A method for analyzing rail transit passenger flow based on multi-source data fusion
By constructing a multi-source data fusion method for rail transit passenger flow analysis, using mobile phone signaling and ROM-SDK location records to build anchor points, and combining Siamese neural networks and the Sinkhorn algorithm, the problem of inaccurate user profiles under a single data source is solved, and accurate identification of rail transit passenger flow and analysis of the reasons for passenger loss are realized.
Patent Information
- Application Number
- CN202611413084.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for analyzing passenger flow in rail transit rely on a single data source, making it difficult to accurately obtain user profile information. Furthermore, the identification of potential passenger flow is not accurate enough, and duplicate matching or partial mismatches are prone to occur.
By constructing residential anchor points, office anchor points, and rail station transit anchor points through mobile phone signaling and ROM-SDK location records, cross-source user matching is performed using a Siamese neural network, and combined with the Sinkhorn algorithm and user profiles, potential passenger flow and the reasons for passenger flow loss are identified.
It improves the accuracy and reliability of multi-source rail transit passenger flow analysis, can accurately identify potential passenger flow and the reasons for passenger loss, and enhances the traceability of results.
Smart Images

Figure CN123116162A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rail transit passenger flow analysis technology, and in particular to a rail transit passenger flow analysis method based on multi-source data fusion. Background Technology
[0002] Rail transit passenger flow analysis typically utilizes automated fare collection data, mobile signaling data, or POI data to statistically analyze residents' origin and destination locations, travel times, and the service area of rail stations. It then identifies rail passenger flow and potential passenger flow based on fixed distance thresholds, time characteristics, or trajectory similarity. Some methods also correlate mobile signaling stop points with surrounding POIs to infer travel purposes such as residence, work, commerce, and leisure. However, existing methods often rely on a single data source or simple rules, making it difficult to obtain user profile information such as occupation, consumption preferences, and transportation mode preferences. Furthermore, when there is no direct identification link between mobile signaling users and external profiled users, duplicate matching or partial mismatches are prone to occur. Simultaneously, current potential passenger flow identification often relies solely on station coverage radius to determine "should have taken a ride but didn't," without fully verifying multi-modal connection conditions, train operating hours, and actual travel modes, leading to inaccurate assessments of potential passenger flow numbers and reasons for passenger loss.
[0003] Therefore, how to provide a rail transit passenger flow analysis method based on multi-source data fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] One objective of this invention is to propose a rail transit passenger flow analysis method based on multi-source data fusion. This invention constructs residential anchor points, office anchor points, and rail station transit anchor points using mobile phone signaling and ROM-SDK location records, converting cross-source users lacking direct identification associations into candidate matching objects with spatiotemporal correspondences. By setting a hierarchical matching layer in a Siamese neural network, the local comparison results of various anchor points are recursively derived along the anchor point position order, reducing the masking effect of a single highly similar anchor point on other differing anchor points. Furthermore, the Sinkhorn algorithm is used to implement row and column joint constraints on the cross-source user matching scores, reducing duplicate matching and local mismatches. Based on this, user profiles are attached to the user's travel chain, and potential passenger flow and reasons for passenger loss are identified by combining walking, cycling, and public transport connection conditions, rail transit operating hours, and actual location records, thereby improving the accuracy, reliability, and traceability of multi-source rail transit passenger flow analysis results.
[0005] A method for analyzing rail transit passenger flow based on multi-source data fusion according to an embodiment of the present invention includes the following steps:
[0006] S1. Acquire mobile phone signaling, POI, ROM-SDK, rail transit operation, transportation network and station facility data, and unify the spatiotemporal reference;
[0007] S2. Identify the dwell and movement processes in mobile phone signaling, connect them in sequence according to the occurrence time to form the user's travel chain, and use POI data to label the spatial semantics of the travel start and end locations;
[0008] S3. Extract user spatiotemporal anchor points from mobile phone signaling and ROM-SDK location records respectively, arrange them into anchor point sequences according to the user affiliation in each data source, and form cross-source candidate matching pairs with mobile phone signaling users and ROM-SDK users and anchor point sequences that have spatiotemporal correspondence of the same type of anchor points;
[0009] S4. Input the cross-source candidate matching pairs into the improved Siamese neural network, encode the mobile phone signaling anchor sequence and the ROM-SDK anchor sequence respectively through a shared encoder, and recursively deduce the matching state between the corresponding anchors along the anchor position order through a hierarchical matching layer, and output the cross-source user matching score.
[0010] S5. Arrange the cross-source user matching scores corresponding to each user combination in the mobile signaling user set and the ROM-SDK user set into a user matching matrix, and use the Sinkhorn algorithm to screen out the cross-source matching relationship between mobile signaling users and ROM-SDK users.
[0011] S6. Mount the selected ROM-SDK user profiles to the corresponding user travel chains, use the user profiles to verify the spatial semantics of the travel start and end locations and determine the travel purpose, and obtain a set of profile travel chains.
[0012] S7. Verify the station connection conditions, rail accessibility and actual travel mode of the profile travel chain. Include profile travel chains that meet the station connection conditions, are rail accessible and do not use rail transit into the potential passenger flow set, and generate potential rail transit passenger flow and passenger flow loss diagnosis results.
[0013] Optionally, S1 includes:
[0014] S11. Read the time fields carried by mobile phone signaling, ROM-SDK positioning records and rail transit operation data, convert different time formats into standard timestamps under the same time zone, and divide the operation date according to the start time of the rail transit operation day;
[0015] S12. Read the coordinates and coordinate system identifiers carried by various data records, convert the base station location, POI location, ROM-SDK positioning location, rail station location, transportation network location and station facility location to the same geographic coordinate system, and project them to a unified plane coordinate space.
[0016] Optionally, S2 includes:
[0017] S21. Arrange mobile phone signaling according to anonymous user identifier and signaling time, merge short-time round-trip records between adjacent base stations, remove location drift records, group continuous records that meet the stay duration and stay range into stay process, and classify records between adjacent stay processes into movement process, and connect them in sequence according to the occurrence time to obtain user travel chain;
[0018] S22. Query the POIs in the neighborhood of the start and end locations of the user's travel chain, merge the query results according to the functional category, and count the number of POIs, spatial distance and category ratio for each functional category.
[0019] S23. Calculate the category semantic weight according to the number of POIs, spatial distance and category proportion of each functional category, and write the functional category with the highest weight and its category semantic weight into the corresponding travel origin and destination.
[0020] Optionally, S3 includes:
[0021] S31. Arrange the mobile phone signaling and ROM-SDK location records according to the user identifier and the recording time, respectively, extract the residential anchor point and the office anchor point from the repeated stay locations across days, and extract the track station passing anchor point from the track station locations passed by the trajectory.
[0022] S32. Arrange the spatial location, occurrence time period and cross-day repetition status of the corresponding anchor points according to the user affiliation and anchor point type in each data source to obtain the mobile signaling anchor point sequence and ROM-SDK anchor point sequence.
[0023] S33. Verify the spatial proximity, time period overlap, and cross-day repetition of anchor points of the same type in the two types of anchor point sequences, and form cross-source candidate matching pairs with mobile signaling users and ROM-SDK users and their anchor point sequences that have spatiotemporal correspondence of anchor points of the same type.
[0024] Optionally, the improved Siamese neural network includes a shared encoder, a hierarchical matching layer, and an output layer;
[0025] The shared encoder encodes the mobile signaling anchor sequence and the ROM-SDK anchor sequence respectively, and outputs the hidden state corresponding to each anchor.
[0026] The hierarchical matching layer is set between the shared encoder and the output layer, including multiple matching nodes connected sequentially according to the anchor point position order. Each matching node corresponds to the same anchor point in the two types of anchor point sequences, and a matching state transmission connection is established between adjacent matching nodes.
[0027] Each matching node receives the hidden state of the corresponding anchor point, compares the two hidden states, merges the comparison result with the matching state passed in by the previous matching node, and outputs the current matching state.
[0028] The first matching node outputs the initial matching state based on the comparison result of the hidden states of the corresponding anchor points.
[0029] The output layer converts the matching status output by the last matching node into a cross-source user matching score.
[0030] Optionally, each matching node outputs its current matching status, including:
[0031] Input the hidden states of the anchor points corresponding to the position order in the two types of anchor point sequences into the same matching node, perform element-wise absolute difference calculation and element-wise product calculation on the two anchor point hidden states, and concatenate the calculation results into the current anchor point comparison vector;
[0032] The first matching node performs a state mapping on the current anchor point comparison vector to obtain the initial matching state;
[0033] Each matching node other than the first matching node concatenates the current anchor comparison vector with the matching state output by the previous matching node. Candidate state mapping and update weight mapping are then performed on the concatenation result. Hyperbolic tangent activation is applied to the candidate state mapping result to obtain the candidate state, and logical function activation is applied to the update weight mapping result to obtain the update weight. The update weight is then multiplied element-wise by the candidate state. The difference between the unit value of each dimension and the corresponding update weight is multiplied element-wise by the previous matching state, and the two products are summed to obtain the current matching state.
[0034] Each matching node transmits its current matching status according to the anchor point position order until the last matching node outputs its last matching status.
[0035] Optionally, S5 includes:
[0036] S51. Extract mobile signaling users and ROM-SDK users from cross-source candidate matching pairs. With mobile signaling users as row objects and ROM-SDK users as column objects, write the corresponding cross-source user matching scores into the user matching matrix, and write the preset masking scores into the matrix positions where there is no cross-source candidate matching relationship.
[0037] S52. The Sinkhorn algorithm is used to perform iterative scaling of the user matching matrix under row and column constraints to obtain the cross-source matching assignment matrix.
[0038] S53. From the matrix positions with cross-source candidate matching relationships, filter out the user matching positions where the matching allocation weight reaches the matching threshold and is the highest matching allocation weight in the corresponding row and column, and obtain the cross-source matching relationship between the mobile signaling user and the ROM-SDK user.
[0039] Optionally, S52 includes:
[0040] S521. Perform exponential mapping on the cross-source user matching scores at each matrix position in the user matching matrix, and write the exponential mapping result into the corresponding matrix position to obtain the matching transmission kernel;
[0041] S522. Count the number of rows and columns of the matching transmission core, divide the total unit allocation equally according to the number of rows, and write the division result into each row as the row target allocation amount; divide the total unit allocation equally according to the number of columns, and write the division result into each column as the column target allocation amount; initialize the scaling amount of each row and the scaling amount of each column to the unit value;
[0042] S523. Summarize the weights of each row in the matching transmission core after adjustment by the current column scaling amount, divide the target allocation amount of the corresponding row by the sum of the weights of that row to obtain the new row scaling amount;
[0043] S524. Summarize the weights of each column in the matching transmission core after the new row scaling amount, and divide the target allocation amount of the corresponding column by the sum of the weights of that column to obtain the new column scaling amount;
[0044] S525. Adjust the matching transmission core according to the current row scaling and column scaling, summarize the adjusted weights of each row and each column respectively, calculate the absolute difference between the total adjusted weight of each row and the target allocation of the corresponding row, and the absolute difference between the total adjusted weight of each column and the target allocation of the corresponding column; return to S523 when the maximum value among all absolute differences is greater than the preset convergence threshold, and stop updating when the maximum value is not greater than the preset convergence threshold;
[0045] S526. Multiply the weight of each matrix position in the matching transmission kernel by the row scaling of the row and the column scaling of the column respectively to obtain the cross-source matching assignment matrix.
[0046] Optionally, S6 includes:
[0047] S61. According to the cross-source matching relationship, write the ROM-SDK user profile field into the corresponding mobile signaling user identifier and attach it to the user travel chain of the mobile signaling user;
[0048] S62. Verify the start and end positions of the user's travel chain with the corresponding user's residential anchor point and office anchor point execution positions, and correct the spatial semantics of the start and end positions according to the verification results.
[0049] S63. Combine the corrected origin and destination spatial semantics, travel time period and user profile fields into a purpose determination item, compare each item with the preset travel purpose determination rule, and write the matched travel purpose into the user's travel chain.
[0050] S64. Write the user profile fields, the corrected spatial semantics, and the travel purpose into the corresponding user travel chain to form a profile travel chain set.
[0051] Optionally, S7 includes:
[0052] S71. Using transportation network data, calculate the walking, cycling, and ground public transport connection routes from the start and end points of the profile travel chain to the surrounding rail stations. Verify the connection distance and connection time of each connection route, and screen out the candidate entry and exit stations that meet the corresponding connection conditions.
[0053] S72. Read the track connection relationship, train running time and transfer relationship between the candidate entry station and the candidate exit station, connect the track sections that can be continuously passed during the travel period, and form a track reachable path.
[0054] S73. Perform location verification between the mobile phone signaling movement record and ROM-SDK location record associated with the profile travel chain and the candidate entry station and candidate exit station, respectively, and perform time verification between the movement time period of the profile travel chain and the passable time period of the track reachable path; mark the profile travel chain that matches both the location verification and the time verification as a track travel chain, and mark the rest of the profile travel chains as non-track travel chains.
[0055] S74. Extract profiled travel chains from non-rail travel chains that simultaneously have candidate entry stations, candidate exit stations, and rail reachable paths, and write them into the potential passenger flow set.
[0056] S75. Compare the connection distance, connection time, number of rail transfers, rail running time and station facility status corresponding to the travel chain of each profile in the potential passenger flow set with the corresponding diagnostic conditions, write the passenger flow loss reason label, and merge the potential passenger flow set according to the candidate entry station, candidate exit station, travel time, travel purpose and user profile fields to form the potential passenger flow and passenger flow loss diagnosis results of rail transit.
[0057] The beneficial effects of this invention are:
[0058] 1. This invention constructs residential anchor points, office anchor points, and rail station passing anchor points by using mobile phone signaling and ROM-SDK location records respectively. It also forms cross-source candidate matching pairs according to the spatial location, time of occurrence, and cross-day repetition status of similar anchor points, so that user records scattered in different data sources can be analyzed under a unified spatiotemporal semantics, thereby improving the basic quality of multi-source rail transit passenger flow data fusion.
[0059] 2. This invention replaces the terminal single distance metric structure of the original Siamese neural network with a hierarchical matching layer, which writes the comparison results of residential, office and rail stations along the anchor point position order. This allows the local matching evidence generated by different anchor points to continuously correct the previous matching state, reduce the masking of other different anchor points by a single highly similar anchor point, and improve the accuracy of cross-source user matching.
[0060] 3. This invention utilizes the Sinkhorn algorithm to apply row and column joint constraints to cross-source user matching scores, and combines user profiles, station connection conditions, rail reachable paths, and actual travel modes to screen potential passenger flows. This enables potential passenger flow results to correspond to specific travel chains, candidate entry and exit stations, and reasons for passenger flow loss, thereby improving the reliability and traceability of rail transit passenger flow diagnosis results. Attached Figure Description
[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0062] Figure 1 This is a flowchart of a rail transit passenger flow analysis method based on multi-source data fusion proposed in this invention;
[0063] Figure 2 This is a schematic diagram of an improved twin neural network structure for a rail transit passenger flow analysis method based on multi-source data fusion proposed in this invention. Detailed Implementation
[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0065] refer to Figures 1-2 A method for analyzing rail transit passenger flow based on multi-source data fusion includes the following steps:
[0066] S1. Acquire mobile phone signaling, POI, ROM-SDK, rail transit operation, transportation network and station facility data, and unify the spatiotemporal reference;
[0067] S2. Identify the dwell and movement processes in mobile phone signaling, connect them in sequence according to the occurrence time to form the user's travel chain, and use POI data to label the spatial semantics of the travel start and end locations;
[0068] S3. Extract user spatiotemporal anchor points from mobile phone signaling and ROM-SDK location records respectively, arrange them into anchor point sequences according to the user affiliation in each data source, and form cross-source candidate matching pairs with mobile phone signaling users and ROM-SDK users and anchor point sequences that have spatiotemporal correspondence of the same type of anchor points;
[0069] S4. Input the cross-source candidate matching pairs into the improved Siamese neural network, encode the mobile phone signaling anchor sequence and the ROM-SDK anchor sequence respectively through a shared encoder, and recursively deduce the matching state between the corresponding anchors along the anchor position order through a hierarchical matching layer, and output the cross-source user matching score.
[0070] S5. Arrange the cross-source user matching scores corresponding to each user combination in the mobile signaling user set and the ROM-SDK user set into a user matching matrix, and use the Sinkhorn algorithm to screen out the cross-source matching relationship between mobile signaling users and ROM-SDK users.
[0071] S6. Mount the selected ROM-SDK user profiles to the corresponding user travel chains, use the user profiles to verify the spatial semantics of the travel start and end locations and determine the travel purpose, and obtain a set of profile travel chains.
[0072] S7. Verify the station connection conditions, rail accessibility and actual travel mode of the profile travel chain. Include profile travel chains that meet the station connection conditions, are rail accessible and do not use rail transit into the potential passenger flow set, and generate potential rail transit passenger flow and passenger flow loss diagnosis results.
[0073] In this embodiment, S1 includes:
[0074] S11. Read the time fields carried by mobile phone signaling, ROM-SDK positioning records and rail transit operation data, convert different time formats into standard timestamps under the same time zone, and divide the operation date according to the start time of the rail transit operation day;
[0075] S12. Read the coordinates and coordinate system identifiers carried by various data records, convert the base station location, POI location, ROM-SDK positioning location, rail station location, transportation network location and station facility location to the same geographic coordinate system, and project them to a unified plane coordinate space.
[0076] In the specific implementation process, a unified second-level timestamp in the East 8 time zone is adopted. The 14-bit date and time field of the mobile phone signaling is parsed according to year, month, day, hour, minute, and second; the ROM-SDK millisecond timestamp is converted to an East 8 time zone second-level timestamp after being divided by 1000; the rail transit operation data combines the operating date with the planned time and the actual time respectively, and the hour value reaching 24 hours is converted to the next calendar day. 04:00 is the starting time of the operating day, and records from 00:00 to 04:00 are included in the previous operating day. When the same train spans calendar days, the operating day to which the originating record belongs is used.
[0077] After time conversion, records are arranged according to data source identifier, anonymous user identifier, and standard timestamp. Duplicate records with the same anonymous user identifier, standard timestamp, and location from the same data source are retained only once. When multiple locations exist for the same user and the same standard timestamp in the ROM-SDK positioning record, the record with the smallest positioning error radius is retained. Records with a positioning error radius greater than 200 meters are marked with a low-precision tag and are not included in the extraction of residential anchor points, office anchor points, and rail station transit anchor points.
[0078] The spatial reference system adopts the 2000 National Geodetic Coordinate System. Coordinate system identifiers and transformation parameters from various data sources are read, and the coordinates of base stations, POIs, ROM-SDK positioning points, rail stations, transportation networks, and station facilities are transformed to this coordinate system. Records lacking coordinate system identifiers or transformation parameters are not included in distance and path calculations. After transformation, metric plane coordinates are generated using the Gauss-Kruger 3-degree zone projection, with the same central meridian used within the same operating area. The transportation network is projected point-by-point, preserving line segment connections, directions of travel, and modes of transport.
[0079] After completing time and space processing, mobile signaling and ROM-SDK records retain data source identifiers, anonymous user identifiers, standard timestamps, operating dates, horizontal coordinates, vertical coordinates, and data quality markers; rail transit operation records retain line identifiers, train number identifiers, station identifiers, operating direction, planned operating timestamps, and actual operating timestamps; and transportation network data retains line segment identifiers, endpoint coordinates, travel modes, and connection relationships.
[0080] In this embodiment, S2 includes:
[0081] S21. Arrange mobile phone signaling according to anonymous user identifier and signaling time, merge short-time round-trip records between adjacent base stations, remove location drift records, group continuous records that meet the stay duration and stay range into stay process, and classify records between adjacent stay processes into movement process, and connect them in sequence according to the occurrence time to obtain user travel chain;
[0082] S22. Query the POIs in the neighborhood of the start and end locations of the user's travel chain, merge the query results according to the functional category, and count the number of POIs, spatial distance and category ratio for each functional category.
[0083] S23. Calculate the category semantic weight according to the number of POIs, spatial distance and category proportion of each functional category, and write the functional category with the highest weight and its category semantic weight into the corresponding travel origin and destination.
[0084] In the specific implementation process, mobile signaling records that have achieved unified spatiotemporal benchmarks are processed based on anonymous user identifiers and operating dates. Each mobile signaling record includes at least an anonymous user identifier, a standard timestamp, a base station identifier, the horizontal coordinate of the base station plane, and the vertical coordinate of the base station plane. When the time interval between adjacent records exceeds 30 minutes, a trajectory breakpoint is set between the two records, and records on both sides of the trajectory breakpoint are not included in the same dwell or movement process.
[0085] Perform short-time round-trip verification on consecutive base station records for the same user. If a record starts from a base station, passes through one or more adjacent base stations, and returns to the original base station, and the time from departure to return does not exceed 150 seconds, delete the intermediate base station records formed during the departure from the original base station, and connect the original base station records before and after the return.
[0086] Position drift records are identified based on both their positional relationship and the recording interval. If the spatial distance between the current record and the previous record is greater than 800 meters, the spatial distance between the current record and the next record is greater than 800 meters, the spatial distance between the previous record and the next record is no greater than 300 meters, and the time interval between the current record and both the previous and next records does not exceed 5 minutes, then the current record is marked as a position drift record and deleted. Positions located at the beginning or end of the recording sequence, where neither the preceding nor following record can be simultaneously obtained, are not subject to this drift determination.
[0087] After completing the record cleanup, establish a location record group starting from the first valid record. If a new record is within 300 meters of the current group's center, add it to the current group and update the group center using the average coordinates within the group; if the distance exceeds 300 meters, end the current group and create a new group. If the time difference between the first and last records within a group is not less than 10 minutes, record the group as a stationary process, and record the group center, start time, and end time; if the time difference is less than 10 minutes, do not create a separate stationary process, and its record is included in the movement record between adjacent stationary processes.
[0088] All valid mobile signaling records between two adjacent stopover processes are grouped into the same movement process. The end time of the previous stopover process is recorded as the start time of the movement, and the start time of the subsequent stopover process is recorded as the end time of the movement. When the center distance between two adjacent stopover processes is greater than 300 meters, the previous stopover process, the corresponding movement process, and the subsequent stopover process are sequentially placed into the same user travel chain; when the center distance is no greater than 300 meters, the two stopover processes are merged, and the merged stopover location and duration are recalculated. User records that form only one stopover process do not generate an independent user travel chain.
[0089] The user travel chain should retain at least the anonymous user identifier, operating date, initial stop location, final stop location, departure time, arrival time, and movement record sequence. The initial stop location and final stop location are respectively represented by the group center coordinates of the corresponding stop process.
[0090] Using the starting and ending points of a user's travel chain as query centers, POIs within 250 meters of the query center are retrieved. The retrieved POIs are then categorized into residential, office, commercial services, education, medical, transportation, and leisure functions. The number of POIs in each function category, the proportion of POIs in that category to the total number of POIs within the query range, and the average distance of each POI in that category to the query center are also calculated.
[0091] For any functional category, divide the number of POIs in that category by the maximum number of POIs in all functional categories to obtain the quantity score; divide the average distance of that category by 250 meters, then subtract the resulting ratio from 1 to obtain the distance score; divide the number of POIs in that category by the total number of POIs in the query range to obtain the category percentage. Multiply the quantity score by 35%, the distance score by 35%, and the category percentage by 30%, and add these three results together to obtain the category semantic weight of that functional category.
[0092] The functional categories are sorted from highest to lowest according to their category semantic weights. The first functional category and its category semantic weight are written as the initial spatial semantics of the starting or ending position, and the second functional category and its category semantic weight are written as the candidate spatial semantics field. When multiple functional categories have the same category semantic weight, the functional category with the smaller average distance is prioritized; if the average distance is still the same, the functional category with more POIs is prioritized. When no POI exists within the query range, both the initial spatial semantics and candidate spatial semantics are written as unrecognized, and the corresponding category semantic weight is written as 0.
[0093] In this embodiment, S3 includes:
[0094] S31. Arrange the mobile phone signaling and ROM-SDK location records according to the user identifier and the recording time, respectively, extract the residential anchor point and the office anchor point from the repeated stay locations across days, and extract the track station passing anchor point from the track station locations passed by the trajectory.
[0095] S32. Arrange the spatial location, occurrence time period and cross-day repetition status of the corresponding anchor points according to the user affiliation and anchor point type in each data source to obtain the mobile signaling anchor point sequence and ROM-SDK anchor point sequence.
[0096] S33. Verify the spatial proximity, time period overlap, and cross-day repetition of anchor points of the same type in the two types of anchor point sequences, and form cross-source candidate matching pairs with mobile signaling users and ROM-SDK users and their anchor point sequences that have spatiotemporal correspondence of anchor points of the same type.
[0097] In the specific implementation process, mobile phone signaling records and ROM-SDK location records from 30 consecutive operating days were selected as anchor point statistics, arranged according to anonymous user identifiers and standard timestamps, respectively. The mobile phone signaling side used both dwell and movement processes; the ROM-SDK side grouped consecutive location records of the same anonymous user with adjacent locations no more than 300 meters apart and adjacent record intervals no more than 30 minutes into the same location record group. The average planar coordinates of each location point within the group were written as the location of the location record group, and the times of the first and last location records were written as the start and end times, respectively, with the difference between the two written as the duration. Location record groups with a duration of not less than 10 minutes participated in the extraction of residential and office anchor points, while location record groups with a duration of less than 10 minutes only participated in the extraction of rail station transit anchor points.
[0098] When extracting residential anchor points, the system reads the dwell time or ROM-SDK location record groups between 22:00 and 06:00 the following day. If multiple nighttime dwell locations exist within an operating day, a weighted average is performed on the horizontal and vertical coordinates of each location based on its duration to obtain the nighttime dwell locations for that operating day. Nighttime dwell locations with a distance of no more than 300 meters between different operating days are grouped into the same nighttime location group. The number of operating days in which each nighttime location group appears is counted, and the nighttime location group with the most operating days is determined as the residential anchor point. The average coordinates of the nighttime dwell locations within the group are then recorded as the residential anchor point location.
[0099] When extracting office anchor points, the system reads the dwell time or ROM-SDK location record groups between 09:00 and 17:00 on weekdays. For users who have already established residential anchor points, daytime dwell locations within a 500-meter horizontal distance of their residential anchor points are deleted. The remaining daytime dwell locations within a 500-meter distance on different weekdays are grouped into the same daytime location group. The number of weekdays and cumulative dwell time for each daytime location group are counted, and the daytime location group with the most weekday occurrences is identified as the office anchor point. If multiple daytime location groups have the same number of weekday occurrences, the group with the longer cumulative dwell time is identified as the office anchor point, and the average coordinates of the daytime dwell locations within that group are recorded as the office anchor point location.
[0100] For residential anchor points, the entry and exit times for each nighttime location group on each operating day are retained. The median of the entry times is recorded as the start time of the residential anchor point's appearance, and the median of the exit times is recorded as the end time. For office anchor points, the entry and exit times for each daytime location group on each workday are retained. The median of the entry times is recorded as the start time of the office anchor point's appearance, and the median of the exit times is recorded as the end time. The original appearance time periods for each operating day and workday are also retained to verify the overlap of appearance time periods for the same type of anchor point in different data sources.
[0101] When extracting track station passing anchor points, query mobile phone signaling movement records and ROM-SDK location records within a 300-meter radius of the track station center. Records of the same user within the same station area with adjacent time intervals not exceeding 15 minutes are grouped into one station passing event, and the station identifier, station center coordinates, event start time, event end time, and intermediate time are recorded. Station passing events within 30 operating days are grouped according to anonymous user identifier and station identifier. The median of the start time, end time, and intermediate time of each event is recorded as the anchor point occurrence start time, occurrence end time, and station passing time, respectively. Only one track station passing anchor point is formed for the same user at the same station.
[0102] For each residential and office anchor point, the number of operating days is counted, and this number is divided by the number of operating days with valid records for the corresponding user to obtain the cross-day repetition ratio. For anchor points where rail stations pass through, the number of operating days at that rail station is divided by the number of operating days with movement records for the user to obtain the cross-day repetition ratio. The cross-day repetition ratio is recorded as the cross-day repetition status of the corresponding anchor point in the anchor point record.
[0103] Anchor points for the same user are arranged in the order of residential anchor points, office anchor points, and anchor points passed through rail stations. When multiple rail stations pass through anchor points, they are arranged from earliest to latest according to the time of passage. Each anchor point record includes anchor point type, horizontal axis, vertical axis, time period, and cross-day repetition ratio; rail station passing through anchor points also includes rail station identifiers, thus forming a mobile signaling anchor point sequence and a ROM-SDK anchor point sequence.
[0104] When verifying anchor points of the same type in the two types of anchor point sequences, read the cross-day repetition ratio of the anchor points on the mobile signaling side and the anchor points on the ROM-SDK side, and calculate the absolute difference between the two cross-day repetition ratios. If the absolute difference is not greater than 0.30, continue to verify the spatial location and occurrence time of the anchor points of the same type; if the absolute difference is greater than 0.30, do not record the two anchor points as corresponding anchor points.
[0105] When verifying the residential anchor points of two users, calculate the horizontal distance between the two residential anchor points and calculate the proportion of the cumulative overlap time of the two anchor points to the shorter cumulative overlap time. If the horizontal distance is no more than 300 meters, the proportion is no less than 60%, and the absolute difference of the cross-day repetition proportion of the two residential anchor points is no greater than 0.30, the two residential anchor points are recorded as corresponding anchor points.
[0106] When verifying the office anchor points of two users, calculate the planar distance between the two office anchor points, and count the number of workdays in which both office anchor points appear on the same workday. Divide the number of workdays in which the two office anchor points appear by the smaller number of workdays to obtain the workday coverage rate. If the planar distance is no more than 500 meters, the workday coverage rate is no less than 50%, and the absolute difference in the cross-day repetition ratio of the two office anchor points is no greater than 0.30, the two office anchor points are recorded as corresponding anchor points.
[0107] When verifying the passage of two users' track stations through anchor points, read the track station identifier and the time of passage. If the track station identifiers of the two track stations passing through anchor points are the same, the difference in the time of passage does not exceed 15 minutes, and the absolute difference in the cross-day repetition ratio is not greater than 0.30, record the two track stations passing through anchor points as corresponding anchor points.
[0108] When there is a corresponding anchor of any type between a mobile signaling user and a ROM-SDK user, the anonymous user identifiers of the two users, the mobile signaling anchor sequence, the ROM-SDK anchor sequence, and the type and bit order of the corresponding anchors are written into the same cross-source candidate matching pair.
[0109] In this embodiment, the improved Siamese neural network includes a shared encoder, a hierarchical matching layer, and an output layer;
[0110] The shared encoder encodes the mobile signaling anchor sequence and the ROM-SDK anchor sequence respectively, and outputs the hidden state corresponding to each anchor.
[0111] The hierarchical matching layer is set between the shared encoder and the output layer, including multiple matching nodes connected sequentially according to the anchor point position order. Each matching node corresponds to the same anchor point in the two types of anchor point sequences, and a matching state transmission connection is established between adjacent matching nodes.
[0112] Each matching node receives the hidden state of the corresponding anchor point, compares the two hidden states, merges the comparison result with the matching state passed in by the previous matching node, and outputs the current matching state.
[0113] The first matching node outputs the initial matching state based on the comparison result of the hidden states of the corresponding anchor points.
[0114] The output layer converts the matching status output by the last matching node into a cross-source user matching score.
[0115] In this implementation, each matching node outputs its current matching status, including:
[0116] Input the hidden states of the anchor points corresponding to the position order in the two types of anchor point sequences into the same matching node, perform element-wise absolute difference calculation and element-wise product calculation on the two anchor point hidden states, and concatenate the calculation results into the current anchor point comparison vector;
[0117] The first matching node performs a state mapping on the current anchor point comparison vector to obtain the initial matching state;
[0118] Each matching node other than the first matching node concatenates the current anchor comparison vector with the matching state output by the previous matching node. Candidate state mapping and update weight mapping are then performed on the concatenation result. Hyperbolic tangent activation is applied to the candidate state mapping result to obtain the candidate state, and logical function activation is applied to the update weight mapping result to obtain the update weight. The update weight is then multiplied element-wise by the candidate state. The difference between the unit value of each dimension and the corresponding update weight is multiplied element-wise by the previous matching state, and the two products are summed to obtain the current matching state.
[0119] Each matching node transmits its current matching status according to the anchor point position order until the last matching node outputs its last matching status.
[0120] In the specific implementation process, the improved Siamese neural network retains the original Siamese neural network's shared encoder and output layer, and replaces the end distance metric structure connected to the output of the shared encoder with a hierarchical matching layer. The hierarchical matching layer sets up matching nodes connected sequentially according to the position order of the anchor points passed through the residential anchor point, office anchor point, and rail station. Each matching node performs a comparison of the hidden states of the same anchor point in the two types of anchor point sequences, and recursively merges the current anchor point comparison result with the previous matching state. The matching state output by the last matching node is converted into a cross-source user matching score by the output layer.
[0121] From each cross-source candidate matching pair, read the mobile signaling user identifier, ROM-SDK user identifier, mobile signaling anchor sequence, ROM-SDK anchor sequence, and corresponding anchor relationship. Configure the anchor position order according to the order of residential anchor, office anchor, and rail station passing anchor. When multiple rail stations pass anchor, first arrange them according to the station's passing time from earliest to latest, and if the station's passing time is the same, arrange them according to the rail station identifier. Only the two anchors marked as corresponding anchors enter the same matching node. Single-sided anchors that do not form a corresponding relationship are retained in the original anchor sequence and participate in shared coding, but no separate matching node is established.
[0122] Each anchor point input includes anchor point type, horizontal coordinate, vertical coordinate, start time, end time, day-to-day repetition ratio, and rail station identifier. Residential anchor points, office anchor points, and rail station-passing anchor points are each represented using a 3-digit one-hot code. The horizontal coordinate is normalized by subtracting the minimum horizontal coordinate of the rail transit operating area from the horizontal coordinate, and then dividing by the difference between the maximum and minimum horizontal coordinates; the horizontal vertical coordinate is normalized using the same process.
[0123] The start and end times are converted to minutes relative to 04:00 on the operating day and divided by 1440 to obtain normalized start and end time values between 0 and 1. The cross-day repetition ratio is directly written into the corresponding field. Track station identifiers are converted into 8-dimensional station vectors using a station embedding table. Each row of the station embedding table corresponds to a track station identifier, and its vector parameters are updated during model training. Residential and office anchors do not have track station identifiers, and their corresponding station vectors are written as 8-dimensional zero vectors. The 3-dimensional anchor type code, 2-dimensional coordinate normalized value, 2-dimensional time normalized value, 1-dimensional cross-day repetition ratio, and 8-dimensional station vector are concatenated sequentially to obtain a 16-dimensional anchor input vector.
[0124] The shared encoder consists of an input mapping layer and a gated recurrent unit. The input mapping layer receives a 16-dimensional anchor input vector, generates a 64-dimensional anchor embedding through linear mapping, and then performs activation on the mapping result through a rectified linear function. The gated recurrent unit has a hidden state dimension of 64 dimensions and receives each anchor embedding sequentially according to the anchor position order, outputting the corresponding 64-dimensional anchor hidden state at each anchor position order.
[0125] The mobile signaling anchor sequence and the ROM-SDK anchor sequence are input into the same input mapping layer and the same gated recurrent unit twice, and the same weight matrix and bias parameters are used in both encoding processes. During model training, only one set of shared encoding parameters is updated, and independent encoding networks are not set up for the mobile signaling and ROM-SDK respectively.
[0126] The hierarchical matching layer starts from the first corresponding anchor point and reads the two anchor point hidden states output by the shared encoder according to the anchor point position order. For the current matching node, it reads the 64-dimensional hidden state of the current anchor point on the mobile signaling side and the 64-dimensional hidden state of the corresponding anchor point on the ROM-SDK side. It subtracts the values of the corresponding dimensions of the two hidden states and takes the absolute value to obtain a 64-dimensional absolute difference vector; then it multiplies the values of the corresponding dimensions of the two hidden states to obtain a 64-dimensional element-wise product vector. The absolute difference vector and the element-wise product vector are concatenated to form a 128-dimensional current anchor point comparison vector.
[0127] The first matching node receives the first current anchor comparison vector, inputs the 128-dimensional current anchor comparison vector into a fully connected layer with 64 output units, and uses a hyperbolic tangent activation function to constrain each output value to between -1 and 1, resulting in a 64-dimensional initial matching state. The first matching node does not read the previous matching state.
[0128] For each matching node other than the first matching node, the 128-dimensional current anchor point comparison vector is concatenated with the 64-dimensional matching state output by the previous matching node to obtain a 192-dimensional node input vector. The node input vector is then input into the candidate state mapping branch and the update weight mapping branch, respectively.
[0129] The candidate state mapping branch uses a fully connected layer with 64 output units to process the node input vectors and generates 64-dimensional candidate states through the hyperbolic tangent activation function.
[0130] The updated weight mapping branch uses another fully connected layer with 64 output units to process the same node input vector, and generates 64-dimensional updated weights through the Sigmoid activation function. The values of each dimension in the updated weights are between 0 and 1, controlling the writing ratio of the candidate state and the previous matching state in the corresponding hidden dimension.
[0131] The update weight is multiplied element-wise by the candidate state to obtain the current state's written portion. The update weight is subtracted from 1 to obtain the previous state's retained weight. This previous state's retained weight is then multiplied element-wise by the previous matching state to obtain the previous state's retained portion. The current state's written portion is then added element-wise to the previous state's retained portion to obtain the current matching state. The current matching state is then passed along the matching state chain to the next matching node, continuing until all corresponding anchor points have been processed.
[0132] When no office anchor point or rail station transit anchor point is established in a cross-source candidate matching pair, no missing anchor points are artificially generated, nor are matching nodes created for the missing anchor points. The hierarchical matching layer only recursively calculates the matching status along the actual corresponding anchor points. This avoids mistakenly using filler values as actual anchor points in cross-source user matching.
[0133] During batch training, the original anchor sequence lengths of different samples are inconsistent. The two types of anchor sequences in the same batch are padded to the maximum sequence length of that batch. The padded positions are written into a 16-dimensional zero vector, and a sequence masking marker is configured. After reading the sequence masking marker, the shared encoder does not update the hidden state corresponding to the padded position, and the hierarchical matching layer does not create matching nodes for the padded positions.
[0134] The last-match node outputs a 64-dimensional last-match state. The output layer uses a fully connected layer with one output unit to process the last-match state, and then uses a Sigmoid activation function to convert the output value into a cross-source user matching score between 0 and 1. The cross-source user matching score represents the relative matching strength between the current mobile signaling user and the current ROM-SDK user.
[0135] The model training data is constructed from data that has already obtained user authorization and completed cross-source user relationship labeling. Positive samples are formed by combining mobile signaling anchor sequences and ROM-SDK anchor sequences belonging to the same actual user, and the sample label is set to 1; negative samples are formed by combining mobile signaling anchor sequences and ROM-SDK anchor sequences belonging to different actual users, and the sample label is set to 0.
[0136] Negative samples include random negative samples and hard negative samples. Random negative samples are randomly combined from mobile signaling users and ROM-SDK users who do not have an actual correspondence. Hard negative samples are extracted from user combinations that have at least one similar anchor point but have been manually identified as not belonging to the same actual user. Hard negative samples include user combinations whose residential anchor points are no more than 500 meters apart but whose office anchor points are different, as well as user combinations whose residential and office anchor points are close but whose rail station passes through different anchor points.
[0137] Each positive sample is configured with 3 negative samples, including 1 random negative sample and 2 hard negative samples. If the number of hard negative samples is insufficient, random negative samples are supplemented from user combinations with no corresponding relationship. The training samples are divided into training set, validation set and test set according to the actual user identifier, with proportions of 70%, 15% and 15% respectively. All positive and negative samples associated with the same actual user are written to only one dataset to avoid the same user appearing in both the training set and the test set.
[0138] The error between the cross-source user matching score output by the model and the sample label is calculated using binary cross-entropy. During training, the Adam optimizer is used with an initial learning rate of 0.001, a batch size of 128 samples, and a maximum training epoch of 50. After each training epoch, the binary cross-entropy of the validation set is calculated. Training stops when the validation set error fails to decrease for five consecutive epochs, and the site embedding table, shared encoder, hierarchical matching layer, and output layer parameters corresponding to the epoch with the lowest validation set error are retained.
[0139] During training, the site embedding table, shared encoder, hierarchical matching layer, and output layer parameters are updated jointly; after training, the model parameters are frozen, and cross-source user matching scores are output for cross-source candidate matching pairs.
[0140] In this embodiment, S5 includes:
[0141] S51. Extract mobile signaling users and ROM-SDK users from cross-source candidate matching pairs. With mobile signaling users as row objects and ROM-SDK users as column objects, write the corresponding cross-source user matching scores into the user matching matrix, and write the preset masking scores into the matrix positions where there is no cross-source candidate matching relationship.
[0142] S52. The Sinkhorn algorithm is used to perform iterative scaling of the user matching matrix under row and column constraints to obtain the cross-source matching assignment matrix.
[0143] S53. From the matrix positions with cross-source candidate matching relationships, filter out the user matching positions where the matching allocation weight reaches the matching threshold and is the highest matching allocation weight in the corresponding row and column, and obtain the cross-source matching relationship between the mobile signaling user and the ROM-SDK user.
[0144] In this embodiment, S52 includes:
[0145] S521. Perform exponential mapping on the cross-source user matching scores at each matrix position in the user matching matrix, and write the exponential mapping result into the corresponding matrix position to obtain the matching transmission kernel;
[0146] S522. Count the number of rows and columns of the matching transmission core, divide the total unit allocation equally according to the number of rows, and write the division result into each row as the row target allocation amount; divide the total unit allocation equally according to the number of columns, and write the division result into each column as the column target allocation amount; initialize the scaling amount of each row and the scaling amount of each column to the unit value;
[0147] S523. Summarize the weights of each row in the matching transmission core after adjustment by the current column scaling amount, divide the target allocation amount of the corresponding row by the sum of the weights of that row to obtain the new row scaling amount;
[0148] S524. Summarize the weights of each column in the matching transmission core after the new row scaling amount, and divide the target allocation amount of the corresponding column by the sum of the weights of that column to obtain the new column scaling amount;
[0149] S525. Adjust the matching transmission core according to the current row scaling and column scaling, summarize the adjusted weights of each row and each column, calculate the absolute difference between the total adjusted weight of each row and the target allocation of the corresponding row, and the absolute difference between the total adjusted weight of each column and the target allocation of the corresponding column; return to S523 when the maximum value among all absolute differences is greater than the preset convergence threshold, and stop updating when the maximum value is not greater than the preset convergence threshold;
[0150] S526. Multiply the weight of each matrix position in the matching transmission kernel by the row scaling of the row and the column scaling of the column respectively to obtain the cross-source matching assignment matrix.
[0151] In the specific implementation process, cross-source candidate matching pairs formed within the same anchor point statistical period are classified into the same matching batch, and the mobile signaling user identifier, ROM-SDK user identifier, and cross-source user matching score output by the improved Siamese neural network are read from each cross-source candidate matching pair. When there are multiple candidate matching records between the same mobile signaling user and the same ROM-SDK user, the record with the highest cross-source user matching score is retained.
[0152] Summarize the mobile signaling user identifiers and ROM-SDK user identifiers appearing in the current matching batch, and sort them according to the user identifiers. Configure the mobile signaling users as row objects of the user matching matrix in sequence, and configure the ROM-SDK users as column objects of the user matching matrix in sequence. Each row and column intersection position corresponds to a user combination consisting of a mobile signaling user and a ROM-SDK user.
[0153] For user combinations that have formed cross-source candidate matching pairs, the cross-source user matching score output by the improved Siamese neural network is written into the corresponding matrix position, with the cross-source user matching score ranging from 0 to 1. For user combinations that have not formed cross-source candidate matching pairs, -2 is written into the corresponding matrix position as a masking score, so that the masking position, after being subjected to an exponential mapping with a temperature parameter of 0.1, forms a kernel weight that is much lower than the candidate position and remains positive.
[0154] After the user matching matrix is constructed, the maximum score within the matrix is read. For any matrix position, the maximum score is subtracted from the current score, the difference is divided by the temperature parameter 0.1, and finally, the natural exponential function is performed on the result. The result of the natural exponential function is then written to the corresponding matrix position to form the matching transmission kernel. Subtracting the maximum score first limits the maximum input of the exponential mapping to 0, reducing the possibility of numerical overflow during exponential calculation.
[0155] The system calculates the number of rows and columns for the matching transmission core. It divides the total unit allocation (1) by the number of rows and writes the result to each row to form the row target allocation. It also divides the total unit allocation (1) by the number of columns and writes the result to each column to form the column target allocation. Therefore, the sum of all row target allocations is 1, and the sum of all column target allocations is also 1.
[0156] Configure a row scaling for each row in the matching transmission core and a column scaling for each column. Before the first iteration, initialize all row scaling and all column scaling to 1.
[0157] When updating the row scaling, the kernel weight at each position of the matching transmission kernel is multiplied by the current column scaling of the corresponding column. The products of each row are summed, and the target allocation for the corresponding row is divided by the total product of that row to obtain the new row scaling. After the row scaling is completed, the kernel weight at each position is multiplied by the latest row scaling of the corresponding row. The products of each column are summed, and the target allocation for the corresponding column is divided by the total product of that column to obtain the new column scaling.
[0158] After completing one round of row and column scaling updates, the kernel weight of each matrix position in the matching transmission kernel is multiplied by the latest row scaling of its row and then by the latest column scaling of its column to obtain the current matching assignment weight for that matrix position. All current matching assignment weights are arranged according to their original row and column positions to form the current matching assignment matrix.
[0159] Summarize the matching assignment weights for each row and column in the current matching assignment matrix, calculate the absolute difference between the total weight of each row and the target assignment for the corresponding row, and calculate the absolute difference between the total weight of each column and the target assignment for the corresponding column. Read the smaller value between the row target assignment and the column target assignment, multiply the smaller value by 0.01, and obtain the convergence threshold.
[0160] Read the maximum value from all row and column absolute differences. If the maximum value is greater than the convergence threshold, continue updating the row and column scaling. If the maximum value is not greater than the convergence threshold, stop iterating and determine the current matching assignment matrix as the cross-source matching assignment matrix. The maximum number of iterations is set to 200. If the convergence condition is not met after 200 iterations, the current matching batch is marked as a calculation anomaly, and the cross-source matching relationship of that batch is not output. The convergence condition is that the absolute difference between the total weight of each row and the target allocation of the corresponding row, and the absolute difference between the total weight of each column and the target allocation of the corresponding column, are both not greater than the convergence threshold.
[0161] After the cross-source matching assignment matrix is formed, only the matrix positions with cross-source candidate matching relationships are read; the positions where masking scores are written are not included in the cross-source matching relationship screening. The smaller value between the row target assignment amount and the column target assignment amount is read, and the smaller value is multiplied by 60% to obtain the matching threshold for the current matching batch. A matching threshold that varies with the number of rows and columns is used to avoid a fixed numerical threshold corresponding to different matching strengths under different matrix sizes.
[0162] For any matrix position with cross-source candidate matching relationships, first compare the matching assignment weight of that position with the matching threshold. If the matching assignment weight is lower than the matching threshold, delete the matrix position; if the matching assignment weight reaches the matching threshold, continue to compare the matching assignment weight of that position with other candidate positions in its row and column.
[0163] When a matrix position has the highest matching weight in both its row and column, that matrix position is added to the initial matching position set. Matrix positions that have the highest matching weight only in their row or only in their column are not added to the initial matching position set.
[0164] When multiple highest-ranking matches with the same value exist in the same row or column, the original cross-source user matching scores of the corresponding user combination in the user matching matrix are read, and the matrix position with the higher original cross-source user matching score is retained. If the original cross-source user matching scores are still the same, the matrix position is written to the matching uncertainty set, and the corresponding cross-source matching relationship is not output.
[0165] Read the matrix positions retained in the initial matching position set, and write the mobile signaling user identifier of the corresponding row, the ROM-SDK user identifier of the corresponding column, the original cross-source user matching score, and the matching allocation weight into the same cross-source matching relationship record.
[0166] In this embodiment, S6 includes:
[0167] S61. According to the cross-source matching relationship, write the ROM-SDK user profile field into the corresponding mobile signaling user identifier and attach it to the user travel chain of the mobile signaling user;
[0168] S62. Verify the start and end positions of the user's travel chain with the corresponding user's residential anchor point and office anchor point execution positions, and correct the spatial semantics of the start and end positions according to the verification results.
[0169] S63. Combine the corrected origin and destination spatial semantics, travel time period and user profile fields into a purpose determination item, compare each item with the preset travel purpose determination rule, and write the matched travel purpose into the user's travel chain.
[0170] S64. Write the user profile fields, the corrected spatial semantics, and the travel purpose into the corresponding user travel chain to form a profile travel chain set.
[0171] In the specific implementation process, a one-to-one mapping between mobile signaling users and ROM-SDK users is established based on cross-source matching relationships. Age range, occupation category, consumption preferences, interest categories, and transportation mode preferences are read and attached to the entire travel chain of the mobile signaling user. Missing profile fields are written as unknown; profile fields are saved according to travel chains and are not repeatedly written to movement location records.
[0172] Calculate the distances between the start and end points of the travel chain and the residential and office anchor points respectively. If the distance to the residential anchor point is no more than 300 meters, a residential location tag is written; if the distance to the office anchor point is no more than 500 meters, an office location tag is written. If only one anchor point is matched, the spatial semantics are corrected to the corresponding type; if neither is matched, the POI semantics are retained. If both anchor points are matched, the actual distance is divided by 300 meters and 500 meters respectively, and the anchor point type with the smaller normalized distance is adopted; if both are the same, the POI semantics are retained and a semantic conflict tag is written. The original spatial semantics, the corrected spatial semantics, and the basis for the correction are saved simultaneously.
[0173] The determination of travel purpose follows the order of priority: anchor point relationship first, destination spatial semantics second, and user profile as a supplementary distinction. If the starting location is residence and the ending location is office, and the departure time is between 05:00 and 11:00, the travel purpose is written as commuting; if the starting location is office and the ending location is residence, and the arrival time is between 16:00 and 02:00 the next day, the travel purpose is written as return trip.
[0174] When the commuting or return travel rule is not matched, the corrected spatial semantics of the termination location are read. If the termination location is for education and the arrival time is between 06:00 and 20:00, the travel purpose is written as education; if the termination location is for medical purposes, the travel purpose is written as medical; if the termination location is for business services and the arrival time is between 09:00 and 23:00, the travel purpose is written as consumption; if the termination location is for leisure and the arrival time is between 08:00 and 24:00, the travel purpose is written as leisure; if the termination location is for office purposes but the starting location is not residential, the travel purpose is written as business.
[0175] When the termination position retains the semantics of both categories, and the difference in semantic weights between the two categories is no greater than 0.05, auxiliary differentiation is performed by reading the user profile. If the occupation category is "student" and the candidate categories include "education," the educational purpose is retained; if the occupation category is "employed" and the candidate categories include "office," the travel purpose is written as "business"; if the consumption preference matches the business service category, the consumption purpose is retained; if the interest category matches the leisure category, the leisure purpose is retained. If the user profile does not correspond to any candidate category, the purpose with the higher semantic weight is selected.
[0176] When a user's travel chain matches multiple destination determination rules simultaneously, the first matching result is retained in the order of commuting, return trip, education, medical care, business, consumption, and leisure. If no rule is matched, the travel destination is written as "other," and the travel destination is not forced to be generated based on the user profile.
[0177] The user profile fields, original spatial semantics, corrected spatial semantics, semantic correction basis, travel purpose, and purpose determination rule identifier are written into the corresponding user travel chain to form a profile travel chain set. User travel chains that do not form cross-source matching relationships are retained in the original set.
[0178] In this embodiment, S7 includes:
[0179] S71. Using transportation network data, calculate the walking, cycling, and ground public transport connection routes from the start and end points of the profile travel chain to the surrounding rail stations. Verify the connection distance and connection time of each connection route, and screen out the candidate entry and exit stations that meet the corresponding connection conditions.
[0180] S72. Read the track connection relationship, train running time and transfer relationship between the candidate entry station and the candidate exit station, connect the track sections that can be continuously passed during the travel period, and form a track reachable path.
[0181] S73. Perform location verification between the mobile phone signaling movement record and ROM-SDK location record associated with the profile travel chain and the candidate entry station and candidate exit station, respectively, and perform time verification between the movement time period of the profile travel chain and the passable time period of the track reachable path; mark the profile travel chain that matches both the location verification and the time verification as a track travel chain, and mark the rest of the profile travel chains as non-track travel chains.
[0182] S74. Extract profiled travel chains from non-rail travel chains that simultaneously have candidate entry stations, candidate exit stations, and rail reachable paths, and write them into the potential passenger flow set.
[0183] S75. Compare the connection distance, connection time, number of rail transfers, rail running time and station facility status corresponding to the travel chain of each profile in the potential passenger flow set with the corresponding diagnostic conditions, write the passenger flow loss reason label, and merge the potential passenger flow set according to the candidate entry station, candidate exit station, travel time, travel purpose and user profile fields to form the potential passenger flow and passenger flow loss diagnosis results of rail transit.
[0184] In the specific implementation process, the transportation network data includes pedestrian roads, non-motorized vehicle roads, bus stops, bus routes, operating times of adjacent bus stops, first and last bus times, departure intervals, and line segment connections. Rail transit operation data includes rail lines, relationships between adjacent stations, first and last train times, interval operating times, and transfer times; station facility data includes the location of rail station entrances and exits, bus connection facilities, non-motorized vehicle parking facilities, and the configuration status of escalators and accessible elevators.
[0185] Using the starting and ending points of the travel chain as centers, search for rail stations within a straight-line distance of no more than 3000 meters to serve as candidate stations for walking and cycling connections. For bus connections, the 3000-meter straight-line distance restriction is not applied; instead, search for bus stations within 800 meters of the starting and ending points, and then search for bus stations within 800 meters of the rail station entrances / exits along the bus routes. Rail stations reachable via the same bus route or with a single transfer are selected as candidate bus connection stations.
[0186] When calculating pedestrian connection routes, the starting and ending points and rail station entrances / exits are integrated into the pedestrian road network, and Dijkstra's shortest path algorithm is used to search for the shortest cumulative road length. Walking time is calculated by dividing the path length by 1.2 meters per second; if the walking path length does not exceed 800 meters and the walking time does not exceed 12 minutes, the corresponding rail station and its pedestrian connection route are retained.
[0187] When calculating cycling connection routes, the starting and ending points and non-motorized vehicle parking areas around the rail stations are connected to the non-motorized vehicle road network, roads where non-motorized vehicles are prohibited are blocked, and Dijkstra's shortest path algorithm is used to search for cycling routes. Cycling time is calculated by dividing the path length by 4 meters per second, plus 2 minutes of parking and walking transfer time; when the cycling route length does not exceed 3000 meters and the cycling time does not exceed 20 minutes, the corresponding rail station and its cycling connection route are retained.
[0188] When calculating ground public transport connection routes, the bus stops at the origin, the bus operating section, and the bus stops at the rail station are connected according to the bus route connection relationship. When real-time arrival data is available, the time between the query time and the estimated arrival time of the next vehicle is recorded as the waiting time; when real-time arrival data is not available, half of the bus route departure interval is recorded as the waiting time. Bus travel time is accumulated segment by segment according to the travel time of adjacent bus stops. Bus connection time is obtained by adding the starting walking time, waiting time, bus travel time, and walking time after getting off the bus; when multiple bus routes exist, the route with the shortest connection time is retained; when the bus connection time does not exceed 30 minutes, the corresponding rail station and its bus connection route are retained.
[0189] When the same rail station meets multiple connection conditions simultaneously, each connection path is retained, and the connection method with the shortest connection time is designated as the preferred connection method. A candidate inbound station is formed when there is at least one rail station that meets the connection conditions at the starting position; a candidate outbound station is formed when there is at least one rail station that meets the connection conditions at the ending position.
[0190] Walking shuttles, cycling shuttles, and ground public transport shuttles each employ independent road networks, travel time calculation methods, and shuttle retention conditions, without sharing distance thresholds or performing distance-to-time conversions among the three shuttle types. This implementation establishes a shuttle parameter record for the operating area, which includes the operating area identifier, walking shuttle retention distance, walking shuttle retention duration, cycling shuttle retention distance, cycling shuttle retention duration, public transport shuttle retention duration, walking shuttle diagnostic distance, cycling shuttle diagnostic distance, and public transport diagnostic duration. The corresponding parameters in this implementation are 800 meters, 12 minutes, 3000 meters, 20 minutes, 30 minutes, 600 meters, 2000 meters, and 20 minutes, respectively. When processing data from other operating areas, the corresponding operating area shuttle parameter record is read according to the operating area identifier.
[0191] A rail transit network is constructed using rail stations as nodes, adjacent station operating sections as travel edges, and transfer relationships between different lines within the same transfer station as transfer edges. For each group of candidate entry and exit stations, the departure time of the profiled travel chain is added to the entry connection time to obtain the expected entry time; station combinations with no trains available after the expected entry time are directly deleted.
[0192] A time-dependent Dijkstra algorithm is used to search for accessible routes, with the earliest arrival time at each rail station as the node state. When expanding adjacent operating sections, the first train that can be boarded after the current arrival time is read, and the waiting time and interval travel time are added to the current arrival time. When passing through transfer stations, the walking time within the station and the transfer waiting time are also added. When multiple estimated arrival times are generated for the same rail station, only the path state with the earliest estimated arrival time is retained. After arriving at a candidate exit station, the exit connection time is added to form the accessible rail path and its total travel time.
[0193] When the same travel profile forms multiple accessible routes, the route with the shorter total travel time is prioritized; when the total travel time is the same, the route with fewer transfers is retained; when the total travel time and the number of transfers are the same, the route with the shorter connection time is retained.
[0194] If a location record exists within 20 minutes before or after the departure time, within 300 meters of the candidate entry / exit station entrance / exit, an entry location consistency marker is written. If a location record exists within 20 minutes before or after the arrival time, within 300 meters of the candidate exit / exit station entrance / exit, an exit location consistency marker is written. The actual travel time is calculated from the latest entry location record to the earliest exit location record. The train waiting time, section travel time, and transfer time along the accessible track path are added together to obtain the track travel time. If the actual travel time falls between the track travel time minus 10 minutes and plus 20 minutes, a track time consistency marker is written. If all three markers exist, it is considered a track trip; otherwise, it is considered a non-track trip. When continuous positioning records are lacking in underground sections, only the entry / exit locations and actual travel time are checked. When both mobile phone signaling movement records and ROM-SDK positioning records have valid locations, the positional relationship between the two types of records and the candidate entry / exit stations is checked separately. If the results of the checks of the two types of records are consistent, the result is adopted; if only one type of record has a valid location, the result of the check of that type of record is adopted. If both types of records are valid but the verification results are contradictory, the travel chain of the profile will be marked as uncertain and will not be included in the rail travel chain or potential passenger flow set.
[0195] Profiled travel chains that simultaneously possess candidate entry stations, candidate exit stations, and rail-accessible routes are read from non-railway travel chains and written into the potential passenger flow set. Profiled travel chains lacking candidate entry stations, candidate exit stations, or rail-accessible routes are not included in the potential passenger flow.
[0196] Generate a churn reason label for each travel chain in the potential passenger flow set. Read the preferred connection methods and corresponding connection results for the starting and ending sides of the travel chain. If the preferred connection method on either side is walking, and the corresponding walking connection path length exceeds the current operating area's walking connection diagnostic distance by 600 meters, write a "walking connection distance too long" label. If the preferred connection method on either side is cycling, and the corresponding cycling connection path length exceeds the current operating area's cycling connection diagnostic distance by 2000 meters, write a "cycling connection distance too long" label. If the preferred connection method on either side is ground public transport, and the public transport connection time (the sum of the starting walking time, waiting time, bus travel time, and walking time after disembarking) exceeds the current operating area's public transport connection diagnostic time by 20 minutes, write a "public transport connection time too long" label.
[0197] If the number of transfers on the rail-accessible path is no less than 2, it is labeled as having too many transfers; if the waiting time between the estimated arrival time and the departure time of the available train exceeds 10 minutes, it is labeled as having too long a train waiting time; if the total time of the rail-accessible path exceeds 120% of the actual travel time of the profiled travel chain, it is labeled as having too long a rail travel time.
[0198] The preferred mode of transportation is cycling, but if the candidate entry or exit station lacks non-motorized vehicle parking facilities, a label indicating insufficient non-motorized vehicle transportation facilities will be added. If the candidate entry or exit station also lacks escalators and accessible elevators, a label indicating insufficient in-station convenience facilities will be added. Multiple churn reason labels can be added to the same travel profile's travel chain simultaneously.
[0199] Potential passenger flows are aggregated based on candidate entry and exit stations and a 1-hour travel timeframe. The number of trips corresponding to each travel purpose, user profile, and churn reason tag is then calculated. If two travel chains for the same user have the same start and end locations and their departure times differ by no more than 5 minutes, the chain with more travel records is retained. If the number of trips is the same, the chain with the longer time span is retained. The output includes the potential number of trips, the number of anonymous users, the main travel purposes, the distribution of user profiles, and the number of churn reasons for each station combination, arranged from highest to lowest based on the number of churn reasons.
[0200] Example 1: To verify the feasibility of this invention in practice, it was applied to a potential passenger flow analysis project for rail transit in a provincial capital city. This city has six rail transit lines with a total of 138 stations. Some outlying residential areas and industrial parks are far from the rail stations. Previously, the rail transit operators mainly relied on mobile signaling OD results and the coverage radius of fixed stations to count potential passenger flow. The original analysis method could not reliably identify the correspondence between mobile signaling users and ROM-SDK users, nor could it effectively distinguish between actual and potential rail passenger flow by combining user profiles, actual travel modes, and rail accessibility paths. This resulted in the same ROM-SDK user being repeatedly matched by multiple mobile signaling users, and trips that were close to the station but could not actually be taken by rail transit during peak hours being incorrectly counted as potential passenger flow.
[0201] This embodiment analyzes data from 30 consecutive operating days, including 168 million mobile signaling records, 48.2 million ROM-SDK location records, 294,000 POI records, 1.126 million rail transit operation records, as well as data on pedestrian roads, non-motorized vehicle roads, bus routes, and station facilities. After deleting duplicate records, marking low-precision location records, and removing invalid coordinates, valid records for 74,260 anonymous mobile signaling users and 15,480 anonymous ROM-SDK users were obtained.
[0202] First, all types of data are converted to second-level timestamps in the East 8 time zone, with 04:00 as the start time of each operating day. All locations are uniformly converted to the 2000 National Geodetic Coordinate System, and then metric plane coordinates are generated using the Gauss-Kruger 3-degree zone projection. Records with a ROM-SDK positioning error radius exceeding 200 meters are not included in the spatiotemporal anchor point extraction.
[0203] Short-term round-trip record merging and location drift record removal are performed on mobile signaling records. If a record leaves its original base station, passes through one or more adjacent base stations, and returns to its original base station within 150 seconds, the intermediate base station record is deleted. If the distance between the current record and the records before and after it exceeds 800 meters, the distance between the records before and after it does not exceed 300 meters, and the time interval between the records does not exceed 5 minutes, the current drift record is deleted. Subsequently, location records are merged within a 300-meter range. Location record groups with a duration of not less than 10 minutes form dwell processes, and adjacent dwell processes form movement processes, resulting in a total of 264,380 user travel chains.
[0204] For each user's travel chain, POIs within a 250-meter radius of the start and end points are queried and categorized into residential, office, commercial services, education, medical, transportation, and leisure categories. Taking the end point of a travel chain as an example, there are 24 POIs in its vicinity: 12 commercial services, 7 office, 3 leisure, and 2 transportation. The number score for commercial service POIs is 1, with an average distance of 105 meters, corresponding to a distance score of 0.58 and a category percentage of 0.50. The calculated category semantic weight is: 1 × 35% + 0.58 × 35% + 0.50 × 30% = 0.703.
[0205] The category semantic weights for office and leisure are 0.476 and 0.284, respectively. Therefore, business services are written as the initial spatial semantics, and office is written as the candidate spatial semantics.
[0206] Residential anchor points, office anchor points, and rail station passing anchor points are extracted from mobile phone signaling and ROM-SDK location records, respectively. Residential anchor points are extracted from locations that repeatedly appear across days between 22:00 and 06:00 the next day. Office anchor points are extracted from locations that repeatedly appear across days between 09:00 and 17:00 on weekdays and are more than 500 meters away from residential anchor points. Movement records entering the 300-meter range of the center of a rail station are grouped into station passing events at 15-minute intervals.
[0207] After processing, the mobile signaling side generated 68,140 residential anchor points, 46,320 office anchor points, and 118,760 transit anchor points for rail stations; the ROM-SDK side generated 14,310 residential anchor points, 10,840 office anchor points, and 32,680 transit anchor points for rail stations. Based on the conditions that residential anchor points are within 300 meters, office anchor points within 500 meters, and transit time for rail stations differs by no more than 15 minutes, 38,620 cross-source candidate matching pairs were generated.
[0208] The residential anchor points of mobile signaling user M32784 and ROM-SDK user R8461 are 184 meters apart, with a 72% overlap in nighttime occurrence time; their office anchor points are 341 meters apart, with a weekday coverage rate of 64%; both records pass through rail station S47, with a 6-minute difference in station passage time. Therefore, the two users and their anchor point sequences are combined into a cross-source candidate matching pair.
[0209] Two types of anchor sequence are input into an improved Siamese neural network. Each anchor is encoded as a 16-dimensional input vector, which is then processed by a shared encoder to form a 64-dimensional hidden state. A hierarchical matching layer processes residential anchors, office anchors, and track station passing anchors in sequence. The element-wise absolute difference and element-wise product of the hidden states of the same anchors are concatenated into a 128-dimensional anchor comparison vector, and the comparison result of the current anchor is recursively merged with the previous 64-dimensional matching state. The cross-source user matching score between user M32784 and user R8461 is 0.924, while the matching score between user M32784 and another ROM-SDK user with a similar residential location but different office location and track passing patterns is 0.371.
[0210] All cross-source user matching scores were written into the user matching matrix, and masked scores were written into non-candidate positions. The Sinkhorn algorithm was used to perform alternating row and column scaling on the user matching matrix, and the convergence condition was met in the 43rd iteration. After joint screening using the matching threshold, the highest row weight, and the highest column weight, a total of 8436 cross-source user matching relationships were formed.
[0211] To verify the effectiveness of cross-source user matching, 12,000 test samples with completed authorization calibration were selected from the candidate matching pairs. Among them, 3,000 positive samples were formed by the same actual user, and 9,000 negative samples were formed by different actual users. The traditional spatiotemporal threshold matching method, the traditional Siamese neural network method, and the method of this invention were compared. The traditional spatiotemporal threshold matching method calculates the matching results for residential distance, office distance, and station occupancy time separately, and selects the user with the highest comprehensive score. The traditional Siamese neural network method uses a shared encoder to form the overall encoding result of two types of anchor sequence and performs an Euclidean distance calculation at the end of the encoder. The comparison results are shown in Table 1.
[0212] Table 1 Comparison of Cross-Source User Matching Results
[0213] Matching methods Accuracy / % Recall rate / % Accuracy / % F1 value / % Traditional spatiotemporal threshold matching 89.02 80.40 76.77 78.54 Traditional twin neural network 95.00 91.00 89.22 90.10 This invention 98.25 97.00 96.04 96.52
[0214] As shown in Table 1, the accuracy of the traditional spatiotemporal threshold matching method is only 76.77%. This is mainly because users living in close proximity are easily matched repeatedly, and using a fixed threshold alone cannot effectively utilize the progressive relationship between office anchors and transit station anchors. Traditional Siamese neural networks compress all anchors into a holistic encoding result, which improves matching performance, but the differences between local anchors may be weakened during the holistic encoding process.
[0215] The method of this invention achieves an accuracy of 98.25%, a recall of 97.00%, and a precision of 96.04%. The hierarchical matching layer can continue to write the differences between office anchor points and rail station transit anchor points when residential anchor points are close to each other. The Sinkhorn algorithm further reduces repeated matching when multiple mobile signaling users jointly select the same ROM-SDK user through row and column constraints, thus achieving higher overall matching performance.
[0216] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for analyzing rail transit passenger flow based on multi-source data fusion, characterized in that, Includes the following steps: S1. Acquire mobile phone signaling, POI, ROM-SDK, rail transit operation, transportation network and station facility data, and unify the spatiotemporal reference; S2. Identify the dwell and movement processes in mobile phone signaling, connect them in sequence according to the occurrence time to form the user's travel chain, and use POI data to label the spatial semantics of the travel start and end locations; S3. Extract user spatiotemporal anchor points from mobile phone signaling and ROM-SDK location records respectively, arrange them into anchor point sequences according to the user affiliation in each data source, and form cross-source candidate matching pairs with mobile phone signaling users and ROM-SDK users and anchor point sequences that have spatiotemporal correspondence of the same type of anchor points; S4. Input the cross-source candidate matching pairs into the improved Siamese neural network, encode the mobile phone signaling anchor sequence and the ROM-SDK anchor sequence respectively through a shared encoder, and recursively deduce the matching state between the corresponding anchors along the anchor position order through a hierarchical matching layer, and output the cross-source user matching score. S5. Arrange the cross-source user matching scores corresponding to each user combination in the mobile signaling user set and the ROM-SDK user set into a user matching matrix, and use the Sinkhorn algorithm to screen out the cross-source matching relationship between mobile signaling users and ROM-SDK users. S6. Mount the selected ROM-SDK user profiles to the corresponding user travel chains, use the user profiles to verify the spatial semantics of the travel start and end locations and determine the travel purpose, and obtain a set of profile travel chains. S7. Verify the station connection conditions, rail accessibility and actual travel mode of the profile travel chain. Include profile travel chains that meet the station connection conditions, are rail accessible and do not use rail transit into the potential passenger flow set, and generate potential rail transit passenger flow and passenger flow loss diagnosis results.
2. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, S1 includes: S11. Read the time fields carried by mobile phone signaling, ROM-SDK positioning records and rail transit operation data, convert different time formats into standard timestamps under the same time zone, and divide the operation date according to the start time of the rail transit operation day; S12. Read the coordinates and coordinate system identifiers carried by various data records, convert the base station location, POI location, ROM-SDK positioning location, rail station location, transportation network location and station facility location to the same geographic coordinate system, and project them to a unified plane coordinate space.
3. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, S2 includes: S21. Arrange mobile phone signaling according to anonymous user identifier and signaling time, merge short-time round-trip records between adjacent base stations, remove location drift records, group continuous records that meet the stay duration and stay range into stay process, and classify records between adjacent stay processes into movement process, and connect them in sequence according to the occurrence time to obtain user travel chain; S22. Query the POIs in the neighborhood of the start and end locations of the user's travel chain, merge the query results according to the functional category, and count the number of POIs, spatial distance and category ratio for each functional category. S23. Calculate the category semantic weight according to the number of POIs, spatial distance and category proportion of each functional category, and write the functional category with the highest weight and its category semantic weight into the corresponding travel origin and destination.
4. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, S3 includes: S31. Arrange the mobile phone signaling and ROM-SDK location records according to the user identifier and the recording time, respectively, extract the residential anchor point and the office anchor point from the repeated stay locations across days, and extract the track station passing anchor point from the track station locations passed by the trajectory. S32. Arrange the spatial location, occurrence time period and cross-day repetition status of the corresponding anchor points according to the user affiliation and anchor point type in each data source to obtain the mobile signaling anchor point sequence and ROM-SDK anchor point sequence. S33. Verify the spatial proximity, time period overlap, and cross-day repetition of anchor points of the same type in the two types of anchor point sequences, and form cross-source candidate matching pairs with mobile signaling users and ROM-SDK users and their anchor point sequences that have spatiotemporal correspondence of anchor points of the same type.
5. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, The improved Siamese neural network includes a shared encoder, a hierarchical matching layer, and an output layer; The shared encoder encodes the mobile signaling anchor sequence and the ROM-SDK anchor sequence respectively, and outputs the hidden state corresponding to each anchor. The hierarchical matching layer is set between the shared encoder and the output layer, including multiple matching nodes connected sequentially according to the anchor point position order. Each matching node corresponds to the same anchor point in the two types of anchor point sequences, and a matching state transmission connection is established between adjacent matching nodes. Each matching node receives the hidden state of the corresponding anchor point, compares the two hidden states, merges the comparison result with the matching state passed in by the previous matching node, and outputs the current matching state. The first matching node outputs the initial matching state based on the comparison result of the hidden states of the corresponding anchor points. The output layer converts the matching status output by the last matching node into a cross-source user matching score.
6. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 5, characterized in that, Each matching node outputs its current matching status, including: Input the hidden states of the anchor points corresponding to the position order in the two types of anchor point sequences into the same matching node, perform element-wise absolute difference calculation and element-wise product calculation on the two anchor point hidden states, and concatenate the calculation results into the current anchor point comparison vector; The first matching node performs a state mapping on the current anchor point comparison vector to obtain the initial matching state; Each matching node other than the first matching node concatenates the current anchor comparison vector with the matching state output by the previous matching node. Candidate state mapping and update weight mapping are then performed on the concatenation result. Hyperbolic tangent activation is applied to the candidate state mapping result to obtain the candidate state, and logical function activation is applied to the update weight mapping result to obtain the update weight. The update weight is then multiplied element-wise by the candidate state. The difference between the unit value of each dimension and the corresponding update weight is multiplied element-wise by the previous matching state, and the two products are summed to obtain the current matching state. Each matching node transmits its current matching status according to the anchor point position order until the last matching node outputs its last matching status.
7. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, S5 includes: S51. Extract mobile signaling users and ROM-SDK users from cross-source candidate matching pairs. With mobile signaling users as row objects and ROM-SDK users as column objects, write the corresponding cross-source user matching scores into the user matching matrix, and write the preset masking scores into the matrix positions where there is no cross-source candidate matching relationship. S52. The Sinkhorn algorithm is used to perform iterative scaling of the user matching matrix under row and column constraints to obtain the cross-source matching assignment matrix. S53. From the matrix positions with cross-source candidate matching relationships, filter out the user matching positions where the matching allocation weight reaches the matching threshold and is the highest matching allocation weight in the corresponding row and column, and obtain the cross-source matching relationship between the mobile signaling user and the ROM-SDK user.
8. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 7, characterized in that, S52 includes: S521. Perform exponential mapping on the cross-source user matching scores at each matrix position in the user matching matrix, and write the exponential mapping result into the corresponding matrix position to obtain the matching transmission kernel; S522. Count the number of rows and columns of the matching transmission core, divide the total unit allocation equally according to the number of rows, and write the division result into each row as the row target allocation amount; divide the total unit allocation equally according to the number of columns, and write the division result into each column as the column target allocation amount; initialize the scaling amount of each row and the scaling amount of each column to the unit value; S523. Summarize the weights of each row in the matching transmission core after adjustment by the current column scaling amount, divide the target allocation amount of the corresponding row by the sum of the weights of that row to obtain the new row scaling amount; S524. Summarize the weights of each column in the matching transmission core after the new row scaling amount, and divide the target allocation amount of the corresponding column by the sum of the weights of that column to obtain the new column scaling amount; S525. Adjust the matching transmission core according to the current row scaling and column scaling, summarize the adjusted weights of each row and each column respectively, calculate the absolute difference between the total adjusted weight of each row and the target allocation of the corresponding row, and the absolute difference between the total adjusted weight of each column and the target allocation of the corresponding column; return to S523 when the maximum value among all absolute differences is greater than the preset convergence threshold, and stop updating when the maximum value is not greater than the preset convergence threshold; S526. Multiply the weight of each matrix position in the matching transmission kernel by the row scaling of the row and the column scaling of the column respectively to obtain the cross-source matching assignment matrix.
9. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, S6 includes: S61. According to the cross-source matching relationship, write the ROM-SDK user profile field into the corresponding mobile signaling user identifier and attach it to the user travel chain of the mobile signaling user; S62. Verify the start and end positions of the user's travel chain with the corresponding user's residential anchor point and office anchor point execution positions, and correct the spatial semantics of the start and end positions according to the verification results. S63. Combine the corrected origin and destination spatial semantics, travel time period and user profile fields into a purpose determination item, compare each item with the preset travel purpose determination rule, and write the matched travel purpose into the user's travel chain. S64. Write the user profile fields, the corrected spatial semantics, and the travel purpose into the corresponding user travel chain to form a profile travel chain set.
10. The rail transit passenger flow analysis method based on multi-source data fusion according to claim 1, characterized in that, S7 includes: S71. Using transportation network data, calculate the walking, cycling, and ground public transport connection routes from the start and end points of the profile travel chain to the surrounding rail stations. Verify the connection distance and connection time of each connection route, and screen out the candidate entry and exit stations that meet the corresponding connection conditions. S72. Read the track connection relationship, train running time and transfer relationship between the candidate entry station and the candidate exit station, connect the track sections that can be continuously passed during the travel period, and form a track reachable path. S73. Perform location verification between the mobile phone signaling movement record and ROM-SDK location record associated with the profile travel chain and the candidate entry station and candidate exit station, respectively, and perform time verification between the movement time period of the profile travel chain and the passable time period of the track reachable path; mark the profile travel chain that matches both the location verification and the time verification as a track travel chain, and mark the rest of the profile travel chains as non-track travel chains. S74. Extract profiled travel chains from non-rail travel chains that simultaneously have candidate entry stations, candidate exit stations, and rail reachable paths, and write them into the potential passenger flow set. S75. Compare the connection distance, connection time, number of rail transfers, rail running time and station facility status corresponding to the travel chain of each profile in the potential passenger flow set with the corresponding diagnostic conditions, write the passenger flow loss reason label, and merge the potential passenger flow set according to the candidate entry station, candidate exit station, travel time, travel purpose and user profile fields to form the potential passenger flow and passenger flow loss diagnosis results of rail transit.
