Data-driven bus card getting-on station intelligent inference method based on reinforcement learning
By employing a data-driven approach based on reinforcement learning and utilizing bus card swiping records for adaptive clustering and commuter passenger correction, the accuracy and robustness issues of bus boarding station inference are resolved, achieving high-precision station inference applicable to all flat-fare bus systems.
Patent Information
- Application Number
- CN202511018298.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-23
- Publication Date
- 2025-11-07
AI Technical Summary
Existing technologies have limitations in their application scope and robustness when it comes to inferring bus passenger boarding stops. In particular, they are difficult to achieve high-precision and automated stop inference when there is a lack of external auxiliary data.
Employing a data-driven approach based on reinforcement learning, this study uses adaptive clustering and multi-day behavior correction of stable commuters to intelligently infer bus boarding points using card swiping records on single-fare buses. This process includes environmental data cleaning, initial clustering, reinforcement learning adaptive merging, preliminary station inference, and cross-day data correction.
It improves the accuracy and cross-day consistency of bus boarding point inference, has a wide coverage, low data requirements, low computing costs, and is suitable for all urban buses equipped with a single-fare AFC system.
Smart Images
Figure CN120912410A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent transportation, in particular to a data-driven bus card boarding station intelligent inference method based on reinforcement learning. BACKGROUND
[0002] With the rapid development of urban public transportation and the wide application of big data technology, the travel behavior data of bus passengers has become an important support for traffic planning, management and service optimization. The boarding station information of bus passengers is of great significance for building accurate OD (origin-destination) matrix, revealing travel patterns and optimizing line structure. However, the current urban bus system generally adopts a flat-rate charging method, and the automatic fare collection system (AFC) usually only records information such as card swiping time, vehicle number and line number, without recording the boarding and alighting stations of passengers.
[0003] Because the flat-rate charging method does not record the boarding and alighting station information of bus passengers, it has caused great difficulty in analyzing the travel behavior of passengers. Currently, the calculation of the boarding station of flat-rate bus passengers mainly falls into two categories. One is the time matching method based on card swiping data and bus GPS data, which requires the use of GPS data of buses in the same time period. The other is the clustering analysis method based on card swiping data. However, the boarding station calculation method based on bus GPS data cannot be effectively applied to buses without GPS devices or scenarios where GPS data is missing in the same time period, and the clustering analysis method based on card swiping data has problems such as fuzzy clustering boundary and determination of the corresponding station of the cluster.
[0004] Therefore, the existing methods still have the problems of limited practical application range and insufficient robustness in inferring the boarding station of bus passengers. In the absence of external auxiliary data, how to rely only on AFC card swiping data and achieve accurate inference of large-scale bus boarding stations through more intelligent, automated and high-precision algorithms has become a research difficulty and development direction in the field of intelligent transportation. SUMMARY
[0005] The present application aims to provide a data-driven bus card boarding station intelligent inference method based on reinforcement learning, which first performs adaptive clustering and direction inference on single-day card swiping records to obtain the preliminary boarding station of each cluster, and then performs local correction on the preliminary results based on the multi-day behavior of stable commuting passengers, thereby improving the robustness and cross-day consistency of the inference, to solve the problem of relying on artificial threshold and being easily disturbed by incidental noise in the existing bus card boarding station inference method.
[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a data-driven bus card boarding station intelligent inference method based on reinforcement learning, comprising at least the following steps:
[0007] S1: constructing environment data, cleaning the bus card record data set D of the research line bus, and the bus and subway line data and station vector data L, according to the standard deviation of the card time, the number of card times and the travel proportion of working days in the morning and evening peak hours, screening out the commuting set C as the basis for subsequent correction;
[0008] S2: preliminary clustering division;
[0009] S3: reinforcement learning adaptive merging to obtain the optimal station cluster set;
[0010] S4: preliminary station inference;
[0011] S5: performing S1-S4 on all vehicles of the research line;
[0012] S6: commuting passenger stable boarding station extraction;
[0013] S7: cross-day data correction of clustering results, after the correction of all clusters is completed, the final cross-day stable boarding station inference result is generated.
[0014] Further, the screening of the commuting set C as the basis for subsequent correction according to the standard deviation of the card time, the number of card times and the travel proportion of working days in the morning and evening peak hours at least includes the following steps:
[0015] S1.1: peak travel regularity analysis, for each candidate passenger, the card records in the morning peak (6:30-9:30) and evening peak (16:30-21:30) during the working days (Monday to Friday) are counted respectively, and the effective travel days of the morning and evening peak and the proportion in all working days are calculated;
[0016] S1.2: further statistical analysis of the peak travel time distribution of each passenger, the standard deviation of the morning and evening peak travel time is calculated respectively as the measure of the stability of the travel time, only when the peak travel day proportion and the time distribution stability of the passenger in the working days both meet the set threshold (peak travel day proportion≥0.4, travel time standard deviation does not exceed the corresponding peak window), it is determined that the passenger is a commuting passenger;
[0017] S1.3: for all passengers meeting the multiple determination criteria of S1.1 and S1.2, extract their commuting features, the commuting features include card number, line, travel days, peak travel proportion and time stability information, and import into the commuting passenger set.
[0018] Further, the S2 at least includes the following steps:
[0019] S2.1: For the card swiping time series of a vehicle in a research line in a day, initial segmentation is performed according to a minimum time interval threshold to generate an initial cluster set {Gi};
[0020] S2.2: Calculate the feature vector of each initial cluster:
[0021] Vi={C Earliest , C Latest , C var , C size , Δt l , Δt r}\*MERGEFORMAT(1)
[0022] Wherein, C Earliest represents the earliest card swiping time in the cluster; C Latest represents the latest card swiping time in the cluster; C var represents the variance of the card swiping time in the cluster; C size represents the number of passengers in the cluster; Δt l represents the interval between the card swiping time of the last passenger in the cluster and the card swiping time of the passenger in the left cluster, and is 0 if the cluster is the first cluster; Δt r represents the interval between the card swiping time of the first passenger in the cluster and the card swiping time of the passenger in the right cluster, and is 0 if the cluster is the last cluster.
[0023] Further, the S3 at least includes the following steps:
[0024] S3.1: The feature vector obtained in S2.2 is spliced into a state vector S and input into a DQN network for training;
[0025] S3.2: Under the ε-greedy strategy, the agent selects whether to merge or not to merge adjacent cluster pairs;
[0026] S3.3: According to the reward function:
[0027] R=α×ΔT inter -β×σ intra \*MERGEFORMAT(2)
[0028] Update the DQN network parameters, wherein ΔT inter represents the increment of the time interval between clusters, which is used to measure the clarity of the time interval between the current action and the adjacent cluster, and can promote the effective distinction of the clustering boundary; σ intraThe standard deviation of the card swiping time in the cluster after merging represents the consistency of the timing behavior of the members in the cluster. The smaller the standard deviation, the more concentrated the card swiping time of the members in the cluster and the more consistent the behavior. The parameters α and β are adjustable weights for balancing the two goals of clear cluster boundary and high internal consistency. The reward function can effectively guide the agent to preferentially select the cluster division with clear boundary and high internal consistency in the decision-making process, thereby improving the accuracy and robustness of the pick-up site inference. The optimal site cluster set is obtained until all clusters are optimally merged.
[0029] Further, the S4 at least comprises the following steps:
[0030] S4.1: Extract the pick-up site information of the transfer passengers according to the known transfer mode (orthogonal transfer, non-orthogonal transfer, subway transfer bus), and divide the site cluster into reference cluster, candidate cluster and pending cluster according to the different transfer modes, wherein the reference cluster refers to the site cluster with a determined pick-up site, the candidate cluster refers to the site cluster with several expected pick-up sites, and the pending cluster refers to the site cluster without pick-up site information;
[0031] S4.2: Divide all bus trips according to the pick-up site information of all reference clusters;
[0032] S4.3: Infer the direction of all bus trips according to the information of the reference cluster and the optimal site cluster set obtained in step 3;
[0033] S4.4: According to the trip division data obtained in the last step, match the pick-up site information of the candidate cluster and the pending cluster based on the information of the reference cluster according to the minimum deviation rule.
[0034] Further, the S4.3 inferring the direction of all bus trips at least includes the following cases:
[0035] S4.3.1: When there are at least two reference clusters in the trip, the direction of the bus trip can be inferred according to the number of the pick-up site of the reference cluster;
[0036] S4.3.2: When there is only one reference cluster in the trip, the direction of the trip is inferred according to the reference cluster and its pick-up site information; when the site cluster number and the pick-up site number belong to the first half of the number sequence, it proves that the bus trip direction is from the starting station to the terminal station; when the site cluster number is in the first half and the pick-up site number is in the second half, it proves that the bus trip direction is from the terminal station to the starting station;
[0037] S4.3.3: When there is no reference cluster in the trip or the direction of the trip cannot be inferred by S4.3.2 because the bus travel direction changes between adjacent trips, the direction of the trip is inferred according to the context of the trip.
[0038] Further, the S4.4 at least comprises the following steps:
[0039] S4.4.1: After determining all the pending clusters, for each pending cluster p, calculate the time interval between it and the adjacent left and right reference clusters C L and C R , wherein the left reference cluster corresponds to the reference station S L , and the right reference cluster corresponds to the reference station S R , define T p,q as the difference between the card swiping time of the last passenger in the cluster p and the card swiping time of the first passenger in the cluster q, when the pending cluster is located in the first cluster group or the last cluster group, only calculate the time interval with the single-side reference cluster; when the pending cluster is located in the middle cluster group, calculate the time interval with both sides reference clusters, then, according to the principle of minimum time deviation, assign the pending cluster to the nearest (minimum time interval) station, thereby completing the boarding station association of all pending clusters;
[0040] S4.4.2: Due to the calculation ambiguity of the distance between bus stops and the uncertainty of the travel time between stations, especially in mountainous cities, here a fixed time threshold of bus stop travel time interval t is used;
[0041] S4.4.3: When the pending cluster is located in the first cluster group, and there are n l stations in the cluster group, calculate T k = |T p,r -d r ×t|, k = 1, 2, …, n l , d r takes the station number distance between the station and the right reference station, t is the fixed time threshold of bus stop travel time interval mentioned in S4.4.2, here the station corresponding to the minimum T k is taken as the associated station of the pending cluster;
[0042] S4.4.4: When the pending cluster is located in the middle cluster group, and there are n m stations in the cluster group, calculate T k = |T l,p -d l ×t| + |T p,r -d r ×t|, k = 1, 2, …, n m , d l takes the station number distance between the station and the left reference station; d r takes the station number distance between the station and the right reference station, t is the fixed time threshold of bus stop travel time interval mentioned in S4.4.2, here the station corresponding to the minimum T kThe corresponding station is defined as the associated station of the pending cluster.
[0043] S4.4.5: When the pending cluster is located at the last cluster group, and there are n r stations in the cluster group, calculate T k = |T l,p -d l ×t|, k = 1, 2, …, n r , d l is the distance between the station and the reference station on the left, t is the interval of the bus station driving time of the fixed time threshold mentioned in S4.4.2, and the minimum T k is taken here. The corresponding station is defined as the associated station of the pending cluster.
[0044] Further, the S6 at least includes the following steps:
[0045] S6.1: Extract the multi-day card swiping data of each passenger in the commuter passenger set, and count the cross-day boarding frequency f of each passenger at each station;
[0046] S6.2: If the frequency f of a passenger boarding at a specific station is greater than a preset stability threshold (more than 65% of the total number of trips), the station is defined as the stable station of the passenger;
[0047] S6.3: Establish a commuter passenger stable station mapping table to record the stable boarding station of each commuter passenger and the corresponding card number.
[0048] Further, the S7 at least includes the following steps:
[0049] S7.1: For each cluster preliminarily inferred each day, respectively, perform commuter passenger stability verification to check whether multiple commuter passengers in the cluster share the same stable station and are inconsistent with the preliminarily inferred station;
[0050] S7.2: Define a correction condition, when the proportion of commuter passengers with consistent stable stations in the cluster exceeds a set correction threshold (35%), data correction is triggered;
[0051] S7.3: Correction execution, the preliminary inferred station of the cluster is corrected to the stable station of the commuter passenger in the cluster, and after the correction of all clusters is completed, the final cross-day stable boarding station inference result is generated.
[0052] Compared with the prior art, the beneficial effects of the present application are:
[0053] The application only uses the card swiping record of a one-ticket bus AFC system, and accurately calculates the boarding station of a passenger through the following two-stage method: in the first stage, deep reinforcement learning (DQN) is used to adaptively cluster the single-day card swiping time sequence and infer the travel direction, so as to obtain the preliminary boarding station of each time cluster; in the second stage, the spatio-temporal stability of the multi-day card swiping of the bus commuter is screened, and the preliminary result is corrected according to the occurrence frequency, and the occasional misjudgment is removed. This method does not require GPS or additional sensors, and completely relies on card swiping data, and is suitable for all cities with one-ticket AFC system. It has wide coverage, low data demand, low calculation cost, and can significantly improve the accuracy and cross-day consistency of the boarding station calculation. BRIEF DESCRIPTION OF DRAWINGS
[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.
[0055] Figure 1 The method flowchart of the present application is shown in the figure.
[0056] Figure 2 The schematic diagram of the overall framework of the present application is shown in the figure.
[0057] Figure 3 The schematic diagram of the bus travel direction division based on the reference cluster of the present application is shown in the figure.
[0058] Figure 4 The schematic diagram of the boarding station correction based on the stable boarding data of the commuter passenger of the present application is shown in the figure. DETAILED DESCRIPTION
[0059] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all.
[0060] Please refer to FIG. X, the data-driven bus card boarding station intelligent inference method based on reinforcement learning at least includes the following steps:
[0061] S1: Construct environment data, clean the bus card record data set D of the bus on the research line, and the bus and subway line data and station vector data L. According to the standard deviation of the card swiping time in the morning and evening peak period, the number of card swiping and the working day travel ratio, the commuter set C is selected as the subsequent correction basis;
[0062] According to the standard deviation of card swiping time during morning and evening peak hours, the number of card swiping times and the travel proportion on weekdays, a commuting set C is screened out as the basis for subsequent correction, at least comprising the following steps:
[0063] S1.1: Peak travel rule analysis, for each candidate passenger, the card swiping records during morning peak (6:30-9:30) and evening peak (16:30-21:30) on weekdays (Monday to Friday) are counted respectively, and the effective travel days during morning and evening peak and their proportions in all weekdays are calculated;
[0064] S1.2: Further statistical analysis is made on the peak travel time distribution of each passenger, and the standard deviation of morning peak and evening peak travel time is calculated respectively as a measure of travel time stability. Only when the proportion of peak travel days and the stability of travel time of the passenger in weekdays both meet the set threshold (peak travel day proportion ≥ 0.4, travel time standard deviation not more than the corresponding peak window), the passenger is determined as a commuter;
[0065] S1.3: For all passengers meeting the multiple determination criteria of S1.1 and S1.2, their commuting features are extracted, including card number, line, travel days, peak travel proportion and time stability information, which are merged into the commuter set.
[0066] S2: Preliminary clustering division;
[0067] S2 at least comprises the following steps:
[0068] S2.1: The card swiping time series of a vehicle on a line in a day is initially segmented by a minimum time interval threshold to generate an initial cluster set {Gi};
[0069] S2.2: The feature vector of each initial cluster is calculated:
[0070] Vi={C Earliest 、C Latest 、C var 、C size 、Δt l 、Δt r}\*MERGEFORMAT(1)
[0071] Wherein, C Earliest represents the earliest card swiping time in the cluster; C Latest represents the latest card swiping time in the cluster; C var represents the variance of card swiping time in the cluster; C size represents the number of passengers in the cluster; Δt l represents the interval between the card swiping time of the last passenger in the cluster and the card swiping time of the first passenger in the left cluster, which is 0 if the cluster is the first cluster; Δt rrepresents the interval of the cluster and the first passenger of the right cluster, and is 0 if the cluster is the last cluster.
[0072] S3: Reinforcement learning adaptive merging to obtain the optimal set of station clusters;
[0073] S3 at least includes the following steps:
[0074] S3.1: The feature vectors obtained in S2.2 are spliced into a state vector S, which is input into the DQN network for training;
[0075] S3.2: Under the ε-greedy strategy, the agent selects whether to merge or not to merge adjacent cluster pairs;
[0076] S3.3: According to the reward function:
[0077] R = α × ΔT inter - β × σ intra \*MERGEFORMAT(2)
[0078] Update the DQN network parameters, where ΔT inter represents the increment of the time interval between clusters, which is used to measure the clarity of the time interval between the current action and the adjacent cluster, and can promote the effective differentiation of the cluster boundary; σ intra represents the standard deviation of the card swiping time in the merged cluster, which reflects the consistency of the time sequence behavior of the members in the cluster. The smaller the standard deviation, the more concentrated the card swiping time of the members in the cluster and the more consistent the behavior. The parameters α and β are adjustable weights for balancing the clarity of the cluster boundary and the consistency within the cluster. The reward function can effectively guide the agent to preferentially select the cluster division with clear boundary and high internal consistency in the decision-making process, improve the accuracy and robustness of the boarding station inference, and obtain the optimal set of station clusters until all clusters are optimally merged.
[0079] S4: Preliminary station inference;
[0080] S4 at least includes the following steps:
[0081] S4.1: According to the known transfer mode (orthogonal transfer, non-orthogonal transfer, subway transfer bus), extract the boarding station information of the transfer passengers, and according to the different transfer modes, divide the station clusters into reference clusters, candidate clusters and pending clusters, wherein the reference cluster refers to the station cluster with a determined boarding station, the candidate cluster refers to the station cluster with several expected boarding stations, and the pending cluster refers to the station cluster without boarding station information.
[0082] S4.2: According to the boarding station information of all reference clusters, divide all bus trips;
[0083] S4.3: According to the information of the reference cluster, combine the optimal station cluster set obtained in step 3 to infer all directions of the bus trip;
[0084] S4.4: According to the trip division data obtained in the previous step, match the boarding station information of the candidate cluster and the pending cluster based on the information of the reference cluster according to the minimum deviation rule.
[0085] The inference of all directions of the bus in S4.3 includes at least the following cases:
[0086] S4.3.1: When there are at least two reference clusters in the trip, the direction of the bus trip can be inferred according to the number of the boarding station of the reference cluster;
[0087] S4.3.2: When there is only one reference cluster in the trip, the direction of the trip is inferred according to the reference cluster and its boarding station information; when the station cluster number and the boarding station number belong to the first half of the number sequence, it proves that the bus trip direction is from the starting station to the terminal station; when the station cluster number is in the first half and the boarding station number is in the second half, it proves that the bus trip direction is from the terminal station to the starting station;
[0088] S4.3.3: When there is no reference cluster in the trip or the direction of the trip cannot be inferred by S4.3.2 because the direction of the bus trip will change between adjacent trips, the direction of the trip is inferred according to the context of the trip.
[0089] S4.4 includes at least the following steps:
[0090] S4.4.1: After determining all pending clusters, for each pending cluster p, calculate the time interval between and the adjacent left and right reference clusters C L and C R , wherein the reference station of the left reference cluster is denoted as S L , the reference station of the right reference cluster is denoted as S R , and T p,q is defined as the difference between the card swiping time of the last passenger in cluster p and the card swiping time of the first passenger in cluster q; when the pending cluster is located in the first cluster group or the last cluster group, only the time interval with the single-side reference cluster is calculated; when the pending cluster is located in the middle cluster group, the time intervals with both sides of the reference cluster are calculated, then according to the minimum time deviation principle, the pending cluster is assigned to the nearest station (with the smallest time interval), thereby completing the association of the boarding station of all pending clusters;
[0091] S4.4.2: Due to the calculation ambiguity of the distance between bus stations and the uncertainty of the travel time between stations, especially in mountainous cities, a fixed time threshold of the travel time interval t of the bus station is used here;
[0092] S4.4.3: When the undetermined cluster is located in the first cluster group, and there are n cluster groups. l When there are 1 station, calculate T k =|T p,r -d r ×t|,k=1,2,……,n l ,d r The distance between this station and the reference station on the right is taken as the station number distance, and t is the interval of the bus station travel time for the fixed time threshold mentioned in S4.4.2. Here, the minimum T is taken. k The corresponding site is used as the associated site for the cluster to be determined;
[0093] S4.4.4: When the undetermined cluster is located in an intermediate cluster, and there are n clusters within the cluster. m When there are 1 station, calculate T k =|T l,p -d l ×t|+|T p,r -d r ×t|,k=1,2,……,n m ,d l Take the distance between the station number of this station and the reference station on the left; d r The distance between this station and the reference station on the right is taken as the station number distance, and t is the interval of the bus station travel time for the fixed time threshold mentioned in S4.4.2. Here, the minimum T is taken. k The corresponding site is used as the associated site for the cluster to be determined;
[0094] S4.4.5: When the undetermined cluster is located in the last cluster group, and there are n cluster groups. r When there are 1 station, calculate T k =|T l,p -d l ×t|,k=1,2,……,n r ,d l The distance between this station and the left-hand reference station is taken as the station number distance, and t is the interval of bus station travel time with a fixed time threshold mentioned in S4.4.2. Here, the minimum T is taken. k The corresponding site is used as the associated site for the cluster to be determined.
[0095] S5: Perform S1-S4 operations on all vehicles on the research line;
[0096] S6: Extract stable boarding stations for commuter passengers;
[0097] S6 includes at least the following steps:
[0098] S6.1: Extract the card swipe data of each commuter passenger over multiple days and count the cross-day boarding frequency f of each passenger at each station;
[0099] S6.2: If the frequency f of a passenger boarding at a certain station is greater than a preset stable threshold (more than 65% of the total number of trips), the station is defined as the stable station of the passenger;
[0100] S6.3: A commuter passenger stable station mapping table is established to record the stable boarding station of each commuter passenger and the corresponding card number.
[0101] S7: Cross-day data correction of clustering results, after the correction of all clusters is completed, the final cross-day stable boarding station inference result is generated.
[0102] S7 at least includes the following steps:
[0103] S7.1: For each cluster of the daily preliminary inference, a commuter passenger stability check is performed respectively to check whether multiple commuter passengers in the cluster share the same stable station and are inconsistent with the preliminary inferred station;
[0104] S7.2: Define the correction condition, when the proportion of commuter passengers with consistent stable stations in the cluster exceeds the set correction threshold (35%), data correction is triggered;
[0105] S7.3: Correction execution, the preliminary inferred station of the cluster is corrected to the stable station of the commuter passengers in the cluster, after the correction of all clusters is completed, the final cross-day stable boarding station inference result is generated.
[0106] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.
Claims
1. A data-driven bus boarding station intelligent inference method based on reinforcement learning, characterized in that: At least comprising the following steps: S1: constructing environment data, cleaning the bus card record data set D of the research line bus and the bus and subway line data and station vector data L, according to the standard deviation of the card time, the number of card swiping and the travel proportion of working days in the morning and evening peak hours, the commuting set C is selected as the subsequent correction basis; S2: preliminary clustering division; S3: reinforcement learning adaptive merging, obtaining the optimal station cluster set; S4: preliminary station inference; S5: operating S1-S4 on all vehicles of the research line; S6: commuting passenger stable boarding station extraction; S7: cross-day data correction of clustering results, after the correction of all clusters is completed, the final stable boarding station inference result is generated.
2. The data-driven bus boarding station intelligent inference method based on reinforcement learning according to claim 1, characterized in that: The commuting set C is selected as the subsequent correction basis according to the standard deviation of the card time, the number of card swiping and the travel proportion of working days in the morning and evening peak hours at least comprises the following steps: S1.1: peak travel regularity analysis, for each candidate passenger, the card record in the morning and evening peak hours in the working days is counted respectively, and the effective travel days of the morning and evening peak hours and the proportion in all working days are calculated; S1.2: further statistical analysis is made on the peak travel time distribution of each passenger, and the standard deviation of the morning and evening peak travel time is calculated as the measurement of the stability of the travel time, only when the peak travel day proportion of the passenger in the working days and the time distribution stability meet the set threshold, the passenger is determined as a commuter; S1.3: for all passengers meeting the multiple determination criteria of S1.1 and S1.2, the commuter characteristics are extracted, including card number, line, travel days, peak travel proportion and time stability information, which are merged into the commuter passenger set. 3.The method of claim 1, wherein: The S2 at least comprises the following steps: S2.1: the card time sequence of a vehicle in the research line is segmented by the minimum time interval threshold to generate an initial cluster set {Gi}; S2.2: the feature vector of each initial cluster is calculated: Vi = {C Earliest , C Latest , C var , C size , Δt l , Δt r}\*MERGEFORMAT(1) where C Earliest represents the earliest card swiping time in the cluster; C Latest represents the latest card swiping time in the cluster; C var represents the variance of card swiping times within the cluster; C size represents the number of passengers in the cluster; Δt l represents the interval between the card swiping time of the last passenger in the cluster and the card swiping time of the last passenger in the left cluster, and is 0 if the cluster is the first cluster; Δt r represents the interval between the card swiping time of the first passenger in the cluster and the card swiping time of the first passenger in the right cluster, and is 0 if the cluster is the last cluster.
4. The data-driven bus boarding station intelligent inference method based on reinforcement learning according to claim 3, characterized in that: The S3 at least comprises the following steps: S3.1: the feature vector obtained in S2.2 is spliced into a state vector S, which is input into the DQN network for training; S3.2: under the ε-greedy strategy, the agent selects to merge or not to merge adjacent cluster pairs; S3.3: according to the reward function: R = a x ΔT inter - β x σ intra \*MERGEFORMAT(2) update the DQN network parameters until all clusters are optimally merged to obtain the optimal station cluster set.
5. The data-driven bus boarding station intelligent inference method based on reinforcement learning according to claim 4, characterized in that: The S4 at least comprises the following steps: S4.1: according to the known transfer mode, the boarding station information of the transfer passenger is extracted, according to the different transfer modes, the station cluster is divided into reference cluster, candidate cluster and pending cluster, wherein the reference cluster refers to the station cluster with a determined boarding station, the candidate cluster refers to the station cluster with several expected boarding stations, and the pending cluster refers to the station cluster without boarding station information; S4.2: according to the boarding station information of all reference clusters, all bus trips are divided; S4.3: according to the information of the reference cluster and the optimal station cluster set obtained in step 3, the direction of all bus trips is inferred; S4.4: According to the trip division data obtained in the last step, match the pick-up station information of the candidate cluster and the pending cluster based on the information of the reference cluster according to the minimum deviation rule.
6. The data-driven bus boarding station intelligent inference method based on reinforcement learning according to claim 5, characterized in that: The inferred bus trip direction in S4.3 at least includes the following cases: S4.3.1: When there are at least two reference clusters in the trip, the bus trip direction can be inferred according to the number of the pick-up station of the reference cluster; S4.3.2: When there is only one reference cluster in the trip, the trip direction is inferred according to the reference cluster and its pick-up station information; when the station cluster number and the pick-up station number belong to the first half of the number sequence, it proves that the bus trip direction is from the starting station to the terminal station; when the station cluster number is in the first half and the pick-up station number is in the second half, it proves that the bus trip direction is from the terminal station to the starting station; S4.3.3: When there is no reference cluster in the trip or the trip direction cannot be inferred by S4.3.2 because the bus travel direction will change between adjacent trips, the direction of the trip is inferred according to the context of the trip.
7. The data-driven bus boarding station intelligent inference method based on reinforcement learning according to claim 6, characterized in that: S4.4 at least includes the following steps: S4.4.1: After determining all the pending clusters, for each pending cluster p, calculate its time interval with the adjacent left and right reference clusters C L and C R , where the left reference cluster corresponds to the reference station S L , and the right reference cluster corresponds to the reference station S R , define T p,q as the difference between the card swiping time of the last passenger in cluster p and the card swiping time of the first passenger in cluster q, when the pending cluster is located at the first cluster group or the last cluster group, only calculate its time interval with the single-side reference cluster; when the pending cluster is located at the middle cluster group, simultaneously calculate its time interval with both sides reference clusters, then, according to the minimum time deviation principle, assign the pending cluster to the nearest station, thereby completing the boarding station association of all pending clusters; S4.4.2: Due to the uncertainty of the distance between bus stations and the travel time between stations, especially in mountainous cities, a fixed time threshold is used to calculate the travel time interval t between bus stations; S4.4.3: When the pending cluster is located in the first cluster group, and there are n l sites in the cluster group, calculate T k = |T p,r -d r ×t|, k = 1, 2, …, n l , d r is the distance between the site and the reference site on the right, t is the interval of the bus site travel time of the fixed time threshold mentioned in S4.4.2, here take the smallest T k corresponding site as the associated site of the pending cluster; S4.4.4: When the pending cluster is located in the middle cluster group, and there are n m sites in the cluster group, calculate T k = |T l,p -d l ×t| + |T p,r -d r ×t|, k = 1, 2, …, n m , d l is the site number distance between the site and the left reference site; d r is the site number distance between the site and the right reference site, t is the interval of bus site driving time of the fixed time threshold mentioned in S4.4.2, here take the minimum T k corresponding site as the associated site of the pending cluster; S4.4.5: When the pending cluster is located at the last cluster group, and there are n r sites in the cluster group, calculate T k = |T l,p -d l ×t|, k = 1, 2, …, n r , d l is the distance between the site and the reference site on the left, t is the interval of the bus site travel time of the fixed time threshold mentioned in S4.4.2, here take the minimum T k corresponding site as the associated site of the pending cluster. 8.The method of claim 7, wherein the method further comprises: S6 at least includes the following steps: S6.1: Extract the multi-day card swiping data of each commuter passenger in the commuter passenger set, and count the cross-day pick-up frequency f of each passenger at each station; S6.2: If the frequency f of a passenger picking up at a specific station is greater than a preset stability threshold, the station is defined as the passenger's stable station; S6.3: Establish a commuter passenger stable station mapping table to record the stable pick-up station of each commuter passenger and the corresponding card number. 9.The data-driven bus boarding station intelligent inference method based on reinforcement learning according to claim 8, characterized in that: S7 at least includes the following steps: S7.1: For each cluster inferred daily, perform commuter passenger stability verification respectively to check whether multiple commuter passengers in the cluster share the same stable station and are inconsistent with the preliminary inferred station; S7.2: Define a correction condition. When the proportion of commuter passengers with consistent stable stations in the cluster exceeds a set correction threshold, data correction is triggered; S7.3: Correction execution, modify the preliminary inferred station of the cluster to the stable station of the commuter passenger in the cluster. After the correction of all clusters is completed, the final cross-day stable pick-up station inference result is generated.