Urban-level individual travel trajectory generation method and system with privacy protection mechanism

Abstract

The invention discloses a city-level individual travel trajectory generation method and system with a privacy protection mechanism. The method comprises the following steps: 1, data acquisition and fusion; 2, key position identification; 3, life pattern extraction; 4, carrying out privacy disturbance and resampling; 5, generating a travel sequence; and 6, performing online iterative optimization and outputting. The system comprises a data fusion and preprocessing module, a key position and life pattern extraction module, a privacy protection generation module, a travel behavior generation module and an online optimization and visualization module. According to the method, differential privacy disturbance, multi-source GIS data fusion, behavior pattern learning and a dynamic generation algorithm are introduced, high-fidelity generation and cross-regional migration of individual travel behaviors are achieved, and therefore urban traffic planning, behavior prediction and emergency simulation are supported under the condition that individual real information is not exposed.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and urban big data processing technology, and in particular to a method and system for generating city-level individual travel trajectories with privacy protection mechanisms. Background Technology

[0002] With the widespread application of mobile communication and positioning technologies, urban travel data (including mobile phone signaling, GPS trajectories, smart cards, POIs, etc.) is widely used for traffic planning, travel behavior analysis, and emergency management. However, existing travel data suffers from the following prominent problems:

[0003] 1. High privacy sensitivity: The raw data contains sensitive information such as place of residence, place of work, and travel route, which restricts data sharing and scientific research use;

[0004] 2. Severe issues with missing and noisy data: Actual sampled data suffers from loss, duplication, and signal drift, resulting in insufficient data availability;

[0005] 3. Lack of model transferability: Most travel generation models are only applicable to a single city and are difficult to generalize to different cities;

[0006] 4. Lack of a privacy-preserving framework for generating data: Existing research focuses on behavioral simulation and statistical fitting, lacking a systematic method for generating high-fidelity virtual travel data under privacy constraints.

[0007] Therefore, there is an urgent need for a system and method that can generate city-level, long-term, and individual-level travel behavior data while protecting privacy, in order to achieve intelligent support for data sharing, security analysis, and traffic planning. Summary of the Invention

[0008] In view of the above-mentioned deficiencies of the prior art, the present invention provides a city-level individual travel trajectory generation method and system with privacy protection mechanism. The purpose is to achieve high-fidelity generation and cross-regional migration of individual travel behavior by introducing differential privacy perturbation, multi-source GIS data fusion, behavior pattern learning and dynamic generation algorithm, thereby supporting urban traffic planning, behavior prediction and emergency simulation without exposing personal real information.

[0009] To achieve the above objectives, this invention discloses a city-level individual travel trajectory generation method with a privacy protection mechanism, comprising the following steps:

[0010] Step 1: Data collection and fusion, integrating multi-source urban data to establish a unified GIS rasterized spatial base;

[0011] Step 2: Key location identification, based on time thresholds and clustering algorithms to identify residences, workplaces, and other locations;

[0012] Step 3: Extract lifestyle patterns and construct the spatiotemporal transition probability matrix P u The system then performs singular value decomposition (SVD) to reduce dimensionality, extracts potential travel pattern vectors, and generates a new transition probability matrix.

[0013] Step 4: Privacy perturbation and resampling. Add differential privacy perturbations to key locations and the transition matrix, and then perform normalization and random sampling.

[0014] Step 5: Travel sequence generation. The Markov chain model is used to simulate the transition sequence of an individual in the state space. The duration is determined by combining the GMM to form a spatiotemporally continuous travel trajectory.

[0015] Step 6: Online iterative optimization and output. By comparing the similarity between the generated data and the activity period distribution and OD matrix of the real statistical characteristics, the model parameters are adjusted and the results are output.

[0016] Preferably, step 1 is as follows:

[0017] Step 1.1: Collect multi-source urban data, including mobile phone signaling data, GPS trajectory data, POI data, land use data, road network data, and / or population density data;

[0018] Step 1.2: Use a spatiotemporal alignment algorithm to align the original timestamp t. s Discretized into a time slot index, the mathematical expression is:

[0019]

[0020] Where Δt is the time resolution; t is the time slot index;

[0021] Step 1.3: Use a spatial fusion algorithm to unify the GPS coordinates into the WGS84 projected coordinate system;

[0022] Next, road-level constraints are applied to the travel route, and the nearest neighbor projection GPS point is matched to the road network G.

[0023] Finally, based on the POI type, the region semantic label is given;

[0024] Step 1.4: Data cleaning and imputation, including outlier removal, trajectory repair, and null value filling;

[0025] Step 1.5: Output data, output the standardized spatiotemporal sample set D. fused ;

[0026] The standardized spatiotemporal sample set D fused ={(u,t,g,lat,lon,poi type ,landtype ,pop density The time grid is t∈{0,...,T-1}.

[0027] Preferably, in step 1.1, the field of the mobile phone signaling data is: user id ,cell id timestamp;

[0028] The field of the GPS trajectory data is: user id ,lat,lon,speed,timestamp;

[0029] The fields in the POI data are: poi id ,poi type ,lat,lon;

[0030] The land use data is: rasterized LU(g) (g∈G);

[0031] The field of the road network data is: node id ,edge id ,level,length;

[0032] The population density data is: pop(g) (g∈G).

[0033] Preferably, in step 1.4, the speed threshold filtering for anomaly point removal is >160km / h;

[0034] The trajectory repair refers to linear interpolation of records with intervals greater than 1 hour.

[0035] The term "null padding" refers to filling missing poi values. type The record uses the dominant poi of the raster g. type filling.

[0036] Preferably, step 2 is as follows:

[0037] Step 2.1, Dwell Point Detection, specifically: Detect the time series formed by any user u according to the corresponding time slot index t. If there exists a point whose spatial radius is less than or equal to R within a consecutive number of time slots... stay And the cumulative duration is greater than or equal to τ stay If so, the corresponding point is identified as a stop point;

[0038] Step 2.2, the clustering of the stopping points, specifically: the DBSCAN clustering algorithm is used on the set formed by all the stopping points, and the obtained clusters represent potential key locations, with the centroid of each cluster taken as the location coordinates;

[0039] Step 2.3, Activity Type Labeling:

[0040] If the stay period is from 20:00 to 08:00 and the cumulative duration is ≥5 hours, it is marked as "Residence (H)";

[0041] If the stay period is from 08:00 to 20:00 and the cumulative duration is ≥3 hours, it will be marked as "Workplace (W)";

[0042] The remaining stops are marked as "Other Locations (O)";

[0043] If a user has multiple clusters that meet the criteria of being labeled as "residence (H)" or "workplace (W)," then the Top-k clusters are selected based on the cumulative duration.

[0044] Step 2.4: Extract the set of high-frequency activity locations of individuals based on trajectory dwell point detection and clustering algorithms, forming the key location set K. u ={L i u};

[0045] Among them, each key position L i u All are composed of coordinate points and activity function labels, and all automatically distinguish between residential locations, workplaces and other activity locations through semantic recognition algorithms based on time thresholds and spatial clustering rules.

[0046] More preferably, step 3 is as follows:

[0047] Step 3.1: Construct the set of active states S = {s1, s2, ..., s...} M}; This includes residence H, workplace W, and other locations O;

[0048] Define the user's activity state A at time slot index t. u A u Let t ∈ state set S, t ∈ [1, L]. Then, for each user u, the activity sequence scanned in chronological order is [A1]. u A2 u ,...,A L u ];

[0049] Transition counting matrix C u This indicates that user u, throughout all time slices, from state s i Transition to state s j The number of occurrences is counted, and its size is the number of activity states M × the number of activity states M, expressed as:

[0050] in,

[0051] Where F(·) is an indicator function, which takes the value 1 if the condition is true, and 0 otherwise;

[0052] L is the length of the user activity sequence;

[0053] When the user's activity state in time slot index t is s i The next time slice t+1 is s j Then in the counting matrix s i，j Add 1;

[0054] Constructing a lifestyle matrix P u ;

[0055] The lifestyle matrix P u Indicates from state s i Transition to state s j The probability is obtained by normalizing the counting matrix:

[0056] in,

[0057] Where λ is the smoothing term; the denominator is the state s. i The sum of all transition counts issued; ensuring that row vector probabilities are normalized, i.e.:

[0058]

[0059] Step 3.2: Matrix dimensionality reduction, for the lifestyle pattern matrix P u Perform Singular Value Decomposition (SVD):

[0060] P u =U∑V T ;

[0061] Where U is the individual feature matrix; Σ is the singular value matrix; V T The fundamental matrix of the pattern;

[0062] Step 3.3: Generate the pattern matrix P′ u Specifically, based on the individual feature matrix U, distribution parameters are extracted, and a new individual feature matrix is ​​generated through random sampling.

[0063] Next, the individual feature matrix With the singular value matrix Σ and the pattern fundamental matrix V T Multiply and reconstruct to generate a new lifestyle pattern matrix P′ u ; .

[0064] More preferably, step 4 is as follows:

[0065] Step 4.1: Add Laplace noise to the location, for the set of key locations Ku Two-dimensional Laplace noise is added to each key location coordinate (lat, lon): lat′=lat+η lat lon′=lon+η lon We obtain the key location set K″ after considering privacy encryption. u ;

[0066] in:

[0067] The privacy budget is allocated as follows: Total budget ε includes location budget ε pos With matrix budget ε mat That is, ε = ε pos +ε mat ;

[0068] Step 4.2, in the lifestyle pattern matrix P′ u Noise addition affects the lifestyle matrix P′ u Each element P in i,j Adding Laplace noise yields a new transition probability matrix with added noise. :

[0069]

[0070] Where m is a matrix The column dimension.

[0071] More preferably, step 5 is as follows:

[0072] Step 5.1: Define the state set S = the key location set K′ u ;

[0073] Step 5.2: Use a Dynamic Bayesian Network (DBN) to describe the state transition. The mathematical expression is:

[0074] P(X t =s j |X t-1 =s i ) = A i,j , P(O t |X t ) = B i,j ;

[0075] Among them, X t In active state, O t In observation state;

[0076] Step 5.3, make state s j Corresponding to one or more specific coordinates (lat, lon);

[0077] state s jWhen dealing with multiple specific coordinates (lat, lon), sampling is performed according to weights.

[0078] Then the emission matrix is ​​used for position selection, and the mathematical expression is:

[0079]

[0080] Among them, the weight w can be determined based on the POI weight and population density;

[0081] Step 5.4: Time-duration modeling. For each state s, the activity duration is fitted using a Gaussian Mixture Model (GMM). The mathematical expression is:

[0082]

[0083] The average dwell time error is controlled within ±0.25 hours;

[0084] Step 5.5: Trajectory generation, sampling the initial state X0 from the user's historical distribution or sample prior;

[0085] Each step is based on A t Sample the next state, based on the GMM sampling dwell time, sample the specific coordinates in that state according to the position weight, and record (t) start ,t end ,lat,lon,s), update time t=t+t stay until the required number of days is reached;

[0086] Step 5.6: Smooth the trajectory and perform path fitting on the coordinates between adjacent time slots.

[0087] More preferably, step 6 is as follows:

[0088] Step 6.1: Data statistics. Statistical analysis is performed on the time-period distribution and OD matrix raster population density of the generated and actual data to obtain the statistical result P. gen and P real ;

[0089] Step 6.2: Define the overall relative deviation based on the combination of raster and time, and calculate the deviation using the following formula:

[0090]

[0091] Target threshold: If Bias > θ bias If so, an update will be triggered;

[0092] Step 6.3, Parameter Update Rules:

[0093] If the bias exceeds the threshold and the privacy budget ε is fixed, then adjust the parameters of the generative model and retrain the parameters of the GMM.

[0094] If the bias exceeds the threshold and the privacy budget ε can be relaxed, then the privacy budget parameter should be adjusted first.

[0095] After updating, regenerate small samples and evaluate again. If the bias does not improve, perform deeper parameter retraining, re-decompose SVD or adjust the DBN structure.

[0096] Step 6.4: Output the results and generate the trajectory file GeoJSON;

[0097] Each Feature represents a pause;

[0098] Step 6.5: Visualization components, displaying points / heatmap layers through Leaflet / Mapbox on the Web map; time window slider controlling t animation; comparison panel displaying the difference between Pgen and Preal, i.e., time curve and heatmap difference map; traffic simulation system; access to the city operation monitoring platform.

[0099] The present invention also provides a system for running the above-mentioned city-level individual travel trajectory generation method with privacy protection mechanism, including a data fusion and preprocessing module, a key location and lifestyle pattern extraction module, a privacy protection generation module, a travel behavior generation module, and an online optimization and visualization module;

[0100] The data fusion and preprocessing module is used to receive multi-source heterogeneous input data, complete data alignment through spatial matching and time synchronization algorithms, clean, denoise and rasterize the raw data, and generate a basic spatiotemporal feature set.

[0101] The multi-source heterogeneous input data includes mobile signaling data, GPS trajectory data, POI data, land use data, population density and road network information;

[0102] The key location and lifestyle pattern extraction module identifies an individual's key activity locations based on preprocessed travel data, including Home, Work, and Other activity locations.

[0103] A Life Pattern Matrix is ​​constructed, using time period and location type as dimensions, to record the probability of an individual's movement at different time periods.

[0104] An algorithm based on time thresholds and spatial clustering rules is used for identification:

[0105] The place of residence is defined as the nighttime period, from 20:00 to 08:00, during which the stay is ≥5 hours.

[0106] Workplace refers to the time spent during weekdays, from 08:00 to 20:00, and staying there for ≥3 hours.

[0107] The privacy-preserving generation module introduces a differential privacy mechanism to add Laplace perturbations to the key location coordinates and activity transition probability matrix. It controls the perturbation intensity of data at different levels through a privacy budget allocation algorithm and uses resampling and probability normalization processes to ensure the statistical consistency of the perturbed data, so that the generated data satisfies the privacy protection constraint ε-privacy while maintaining the stability of behavioral distribution characteristics.

[0108] The travel behavior generation module generates time-series travel sequences based on the perturbed key location set and lifestyle matrix, using an improved dynamic Bayesian network (DBN) or Markov chain model. It estimates the duration of each activity state using a Gaussian mixture model (GMM) and outputs an individual-level virtual travel trajectory containing time, location, and activity type.

[0109] The online optimization and visualization module monitors the statistical deviation of the generated results in real time, adjusts the model parameters through an iterative feedback mechanism, outputs the generated travel data to the traffic simulation system or urban operation monitoring platform, and provides map visualization and data interface APIs for system integration.

[0110] The beneficial effects of this invention are:

[0111] This invention balances privacy protection and data availability: through differential privacy and resampling mechanisms, it achieves anonymized generation of travel data, significantly reducing the risk of privacy leakage;

[0112] This invention has the ability to generate high-fidelity travel routes: the correlation coefficient R between the generated trajectory and the real data in terms of spatial distribution and temporal patterns is... 2 ≥0.90;

[0113] This invention has strong cross-city mobility: by using GIS and POI feature input, travel data can be generated in cities without actual measurement data, with an RMSE of less than 0.04;

[0114] The system integration and real-time performance of this invention: It supports online generation and visualization output, and can be directly connected to traffic simulation and planning management platforms;

[0115] The invention improves computational efficiency: compared with traditional full trajectory simulation, the generation speed is increased by about 4 times, and the data storage is reduced by more than 80%.

[0116] This invention combines a city-level individual travel trajectory generation system and method with Geographic Information System (GIS), differential privacy, and multi-source spatiotemporal data analysis. It can be applied to fields such as urban traffic planning, smart travel simulation, privacy and security data sharing, and traffic demand forecasting.

[0117] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0118] Figure 1 A flowchart of an embodiment of the present invention is shown.

[0119] Figure 2 A schematic diagram of the system structure in one embodiment of the present invention is shown. Detailed Implementation

[0120] Example: Figure 1 As shown, the method for generating city-level individual travel trajectories with privacy protection mechanisms includes the following steps:

[0121] Step 1: Data collection and fusion, integrating multi-source urban data to establish a unified GIS rasterized spatial base;

[0122] Step 2: Key location identification, based on time thresholds and clustering algorithms to identify residences, workplaces, and other locations;

[0123] Step 3: Extract lifestyle patterns and construct the spatiotemporal transition probability matrix P u The system then performs singular value decomposition (SVD) to reduce dimensionality, extracts potential travel pattern vectors, and generates a new transition probability matrix.

[0124] Step 4: Privacy perturbation and resampling. Add differential privacy perturbations to key locations and the transition matrix, and then perform normalization and random sampling.

[0125] Step 5: Travel sequence generation. The Markov chain model is used to simulate the transition sequence of an individual in the state space. The duration is determined by combining the GMM to form a spatiotemporally continuous travel trajectory.

[0126] Step 6: Online iterative optimization and output. By comparing the similarity between the generated data and the activity period distribution and OD matrix of the real statistical characteristics, the model parameters are adjusted and the results are output.

[0127] In some embodiments, step 1 is specifically as follows:

[0128] Step 1.1: Collect multi-source urban data, including mobile phone signaling data, GPS trajectory data, POI data, land use data, road network data, and / or population density data;

[0129] Step 1.2: Use a spatiotemporal alignment algorithm to align the original timestamp t. s Discretized into a time slot index, the mathematical expression is:

[0130]

[0131] Where Δt is the time resolution; t is the time slot index;

[0132] In practical applications, the time resolution Δt is usually 15 minutes, which can be adjusted as needed. The corresponding total daily time period T = 24 × 60 / Δt.

[0133] Step 1.3: Use a spatial fusion algorithm to unify the GPS coordinates into the WGS84 projected coordinate system;

[0134] Next, road-level constraints are applied to the travel route, and the nearest neighbor projection GPS point is matched to the road network G.

[0135] Finally, based on the POI type, the region semantic label is given;

[0136] In practical applications, area semantic labels include residential areas, office areas, commercial areas, etc.

[0137] Step 1.4: Data cleaning and imputation, including outlier removal, trajectory repair, and null value filling;

[0138] Step 1.5: Output data, output the standardized spatiotemporal sample set D. fused ;

[0139] Standardized spatiotemporal sample set D fused ={(u,t,g,lat,lon,poi type ,land type ,pop density The time grid is t∈{0,...,T-1}.

[0140] In some embodiments, in step 1.1, the field of the mobile signaling data is: user id ,cell id timestamp;

[0141] The GPS trajectory data field is: user id ,lat,lon,speed,timestamp;

[0142] The fields for POI data are: poi id ,poi type ,lat,lon;

[0143] Land use data is: rasterized LU(g) (g∈G);

[0144] The fields for road network data are: node id ,edge id ,level,length;

[0145] Population density data is: pop(g) (g∈G).

[0146] In some embodiments, in step 1.4, the speed threshold filtering for outlier removal is >160km / h;

[0147] Track repair refers to linear interpolation of records with intervals greater than 1 hour;

[0148] Null padding refers to filling missing POIs type The record uses the dominant poi of the raster g. type filling.

[0149] In some embodiments, step 2 is specifically as follows:

[0150] Step 2.1, Dwell Point Detection, specifically: Detect the time series formed by any user u according to the corresponding time slot index t. If there exists a point whose spatial radius is less than or equal to R within a consecutive number of time slots... stay And the cumulative duration is greater than or equal to τ stay If so, the corresponding point is identified as a stop point;

[0151] In practical applications, R stay =200m, τ stay =20min (or τ) stay =2Δt).

[0152] Step 2.2, Clustering of Stop Points: Specifically, the DBSCAN clustering algorithm is used on the set of all stop points. The resulting clusters represent potential key locations, and the centroid of each cluster is taken as the location coordinates.

[0153] In practical applications, the R-value of the DBSCAN clustering algorithm is typically... ε =300m; MinPts=5; The position coordinates are usually (lat,lon).

[0154] Step 2.3, Activity Type Labeling:

[0155] If the stay period is from 20:00 to 08:00 and the cumulative duration is ≥5 hours, it is marked as "Residence (H)";

[0156] If the stay period is from 08:00 to 20:00 and the cumulative duration is ≥3 hours, it will be marked as "Workplace (W)";

[0157] The remaining stops are marked as "Other Locations (O)";

[0158] If a user has multiple clusters that meet the criteria of being labeled as "residence (H)" or "workplace (W)," then the Top-k clusters are selected based on the cumulative duration.

[0159] In practical applications, k=5 or k=10 is often chosen. When sorting by cumulative duration, the Top-k can be selected by keeping the top 5 that meet the condition of being marked as "residence (H)" or "workplace (W)", and the top 10 that meet the condition of being marked as "other place (O)".

[0160] Step 2.4: Extract the set of high-frequency activity locations of individuals based on trajectory dwell point detection and clustering algorithms, forming the key location set K. u ={L i u};

[0161] Among them, each key position L i u All are composed of coordinate points and activity function labels, and all automatically distinguish between residential locations, workplaces and other activity locations through semantic recognition algorithms based on time thresholds and spatial clustering rules.

[0162] In some embodiments, step 3 is specifically as follows:

[0163] Step 3.1: Construct the set of active states S = {s1, s2, ..., s...} M}; This includes residence H, workplace W, and other locations O;

[0164] Define the user's activity state A at time slot index t. u A u Let t ∈ state set S, t ∈ [1, L]. Then, for each user u, the activity sequence scanned in chronological order is [A1]. u A2 u ,...,A L u ];

[0165] Transition counting matrix C u This indicates that user u, throughout all time slices, from state s i Transition to state s j The number of occurrences is counted, and its size is the number of activity states M × the number of activity states M, expressed as:

[0166] in,

[0167] Where F(·) is an indicator function, which takes the value 1 if the condition is true, and 0 otherwise;

[0168] L is the length of the user activity sequence;

[0169] When the user's activity state in time slot index t is s i The next time slice t+1 is s j Then in the counting matrix s i，jAdd 1;

[0170] Constructing a lifestyle matrix P u ;

[0171] Lifestyle Matrix P u Indicates from state s i Transition to state s j The probability is obtained by normalizing the counting matrix:

[0172] in,

[0173] Where λ is the smoothing term, which is usually taken as 10 in practical applications. -6 ~10 -3 The denominator is state s. i The sum of all transition counts issued; ensuring that row vector probabilities are normalized, i.e.:

[0174]

[0175] Step 3.2: Matrix dimensionality reduction, for the lifestyle pattern matrix P u Perform Singular Value Decomposition (SVD):

[0176] P u =U∑V T ;

[0177] Where U is the individual feature matrix; Σ is the singular value matrix; V T The fundamental matrix of the pattern;

[0178] Step 3.3: Generate the pattern matrix P′ u Specifically, based on the individual feature matrix U, distribution parameters are extracted, and a new individual feature matrix is ​​generated through random sampling.

[0179] Next, the individual feature matrix With the singular value matrix Σ and the pattern fundamental matrix V T Multiply and reconstruct to generate a new lifestyle pattern matrix P′ u ; .

[0180] In some embodiments, step 4 is specifically as follows:

[0181] Step 4.1: Add Laplace noise to the location for the key location set K. u Two-dimensional Laplace noise is added to each key location coordinate (lat, lon): lat′=lat+η lat lon′=lon+η lon We obtain the key location set K″ after considering privacy encryption. u ;

[0182] Where: η lat ,η lon ~Lap(0,b),

[0183] The privacy budget is allocated as follows: Total budget ε includes location budget ε pos With matrix budget ε mat That is, ε = ε pos +ε mat In practical applications, the default value is ε. pos =0.2ε, ε mat =0.8ε.

[0184] Step 4.2, in the lifestyle pattern matrix P′ u Noise addition affects the lifestyle matrix P′ u Each element P in i,j Adding Laplace noise yields a new transition probability matrix with added noise.

[0185]

[0186] Where m is a matrix The column dimension.

[0187] In practical applications, through the above steps, the final set of key locations K" after considering privacy encryption is obtained. u and the new transition probability matrix after adding noise.

[0188] In some embodiments, step 5 is specifically as follows:

[0189] Step 5.1: Define the state set S = key location set K′ u ;

[0190] Step 5.2: Use a Dynamic Bayesian Network (DBN) to describe the state transition. The mathematical expression is:

[0191] P(X t =s j |X t-1 =s i ) = A i,j , P(O t |X t ) = B i,j ;

[0192] Among them, X t In active state, O t In observation state;

[0193] Step 5.3, make state s jCorresponding to one or more specific coordinates (lat, lon);

[0194] state s j When dealing with multiple specific coordinates (lat, lon), sampling is performed according to weights.

[0195] Then the emission matrix is ​​used for position selection, and the mathematical expression is:

[0196]

[0197] Among them, the weight w can be determined based on the POI weight and population density;

[0198] Step 5.4: Time-duration modeling. For each state s, the activity duration is fitted using a Gaussian Mixture Model (GMM). The mathematical expression is:

[0199]

[0200] The average dwell time error is controlled within ±0.25 hours;

[0201] Step 5.5: Trajectory generation, sampling the initial state X0 from the user's historical distribution or sample prior;

[0202] Each step is based on A t Sample the next state, based on the GMM sampling dwell time, sample the specific coordinates in that state according to the position weight, and record (t) start ,t end ,lat,lon,s), update time t=t+t stay This continues until the required number of days is reached; in practice, it is usually 7 days.

[0203] Step 5.6: Smooth the trajectory and perform path fitting on the coordinates between adjacent time slots.

[0204] In practical applications, to ensure the continuity of the journey, path fitting is performed on the coordinates between adjacent time slots, such as using the nearest road path interpolation or the shortest path constraint, and interpolating on the road map G using Dijkstra.

[0205] In some embodiments, step 6 is specifically as follows:

[0206] Step 6.1: Data statistics. Statistical analysis is performed on the time-period distribution and OD matrix raster population density of the generated and actual data to obtain the statistical result P. gen and P real ;

[0207] Step 6.2: Define the overall relative deviation based on the combination of raster and time, and calculate the deviation using the following formula:

[0208]

[0209] Target threshold: If Bias > θ bias If θ is reached, an update will be triggered; in practical applications, the default is θ. bias =0.05, which is 5%.

[0210] Step 6.3, Parameter Update Rules:

[0211] If the bias exceeds the threshold and the privacy budget ε is fixed, then adjust the parameters of the generative model and retrain the parameters of the GMM.

[0212] If the bias exceeds the threshold and the privacy budget ε can be relaxed, then the privacy budget parameter should be adjusted first.

[0213] After updating, regenerate small samples and evaluate again. If the bias does not improve, perform deeper parameter retraining, re-decompose SVD or adjust the DBN structure.

[0214] Step 6.4: Output the results and generate the trajectory file GeoJSON;

[0215] Each Feature represents a pause;

[0216] In practical applications, Feature:

[0217] {"type":"Feature","properties":{"user id ","t start ","t end ","activity"},"geometry":{"type":"Point","coordinates":[lat,lon]}}.

[0218] Step 6.5: Visualization components, displaying points / heatmap layers through Leaflet / Mapbox on the Web map; time window slider controlling t animation; comparison panel displaying the difference between Pgen and Preal, i.e., time curve and heatmap difference map; traffic simulation system; access to the city operation monitoring platform.

[0219] The present invention also provides a system for running the above-mentioned city-level individual travel trajectory generation method with privacy protection mechanism, including a data fusion and preprocessing module, a key location and lifestyle pattern extraction module, a privacy protection generation module, a travel behavior generation module, and an online optimization and visualization module;

[0220] The data fusion and preprocessing module is used to receive multi-source heterogeneous input data, complete data alignment through spatial matching and time synchronization algorithms, clean the raw data, denoise and rasterize it, and generate a basic spatiotemporal feature set.

[0221] Multi-source heterogeneous input data includes mobile signaling data, GPS trajectory data, POI data, land use data, population density and road network information;

[0222] The key location and lifestyle pattern extraction module identifies an individual's key activity locations based on preprocessed travel data, including home, workplace, and other activity locations.

[0223] A Life Pattern Matrix is ​​constructed, using time period and location type as dimensions, to record the probability of an individual's movement at different time periods.

[0224] An algorithm based on time thresholds and spatial clustering rules is used for identification:

[0225] The place of residence is defined as the nighttime period, from 20:00 to 08:00, during which the stay is ≥5 hours.

[0226] Workplace refers to the time spent during weekdays, from 08:00 to 20:00, and staying there for ≥3 hours.

[0227] The privacy-preserving generation module introduces a differential privacy mechanism (DP) to add Laplace perturbations to the key location coordinates and activity transition probability matrix. It controls the perturbation intensity of data at different levels through a privacy budget allocation algorithm (ε-Allocation) and uses resampling and probability normalization processes to ensure the statistical consistency of the perturbed data. This ensures that the generated data meets the privacy protection constraint ε-privacy while maintaining the stability of behavioral distribution characteristics.

[0228] The travel behavior generation module generates time-series travel sequences based on the perturbed key location set and lifestyle matrix, using an improved dynamic Bayesian network (DBN) or Markov chain model. It estimates the duration of each activity state through a Gaussian mixture model (GMM) and outputs individual-level virtual travel trajectories containing time, location, and activity type.

[0229] The online optimization and visualization module monitors the statistical deviation of the generated results in real time, adjusts the model parameters through an iterative feedback mechanism, outputs the generated travel data to the traffic simulation system or urban operation monitoring platform, and provides map visualization and data interface APIs for system integration.

[0230] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for generating city-level individual travel trajectories with privacy protection mechanisms; characterized in that, Includes the following steps: Step 1: Data collection and fusion, integrating multi-source urban data to establish a unified GIS rasterized spatial base; Step 2: Key location identification, based on time thresholds and clustering algorithms to identify residences, workplaces, and other locations; Step 3: Extract lifestyle patterns and construct the spatiotemporal transition probability matrix P u The system then performs singular value decomposition (SVD) to reduce dimensionality, extracts potential travel pattern vectors, and generates a new transition probability matrix. Step 4: Privacy perturbation and resampling. Add differential privacy perturbations to key locations and the transition matrix, and then perform normalization and random sampling. Step 5: Travel sequence generation. The Markov chain model is used to simulate the transition sequence of an individual in the state space. The duration is determined by combining the GMM to form a spatiotemporally continuous travel trajectory. Step 6: Online iterative optimization and output. By comparing the similarity between the generated data and the activity period distribution and OD matrix of the real statistical characteristics, the model parameters are adjusted and the results are output.

2. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 1, characterized in that, Step 1 is as follows: Step 1.1: Collect multi-source urban data, including mobile phone signaling data, GPS trajectory data, POI data, land use data, road network data, and / or population density data; Step 1.2: Use a spatiotemporal alignment algorithm to align the original timestamp t. s Discretized into a time slot index, the mathematical expression is: Where Δt is the time resolution; t is the time slot index; Step 1.3: Use a spatial fusion algorithm to unify the GPS coordinates into the WGS84 projected coordinate system; Next, road-level constraints are applied to the travel route, and the nearest neighbor projection GPS point is matched to the road network G. Finally, based on the POI type, the region semantic label is given; Step 1.4: Data cleaning and imputation, including outlier removal, trajectory repair, and null value filling; Step 1.5: Output data, output the standardized spatiotemporal sample set D. fused ; The standardized spatiotemporal sample set D fused ={(u,t,g,lat,lon,poi type ,land type ,pop density The time grid is t∈{0,...,T-1}.

3. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 1, characterized in that, In step 1.1, the field of the mobile phone signaling data is: user id ,cell id timestamp; The field of the GPS trajectory data is: user id ,lat,lon,speed,timestamp; The fields in the POI data are: poi id ,poi type ,lat,lon; The land use data is: rasterized LU(g) (g∈G); The field of the road network data is: node id ,edge id ,level,length; The population density data is: pop(g) (g∈G).

4. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 1, characterized in that, In step 1.4, the speed threshold filtering for outlier removal is >160km / h; The trajectory repair refers to linear interpolation of records with intervals greater than 1 hour. The term "null padding" refers to filling missing poi values. type The record uses the dominant poi of the raster g. type filling.

5. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 1, characterized in that, Step 2 is as follows: Step 2.1, Dwell Point Detection, specifically: Detect the time series formed by any user u according to the corresponding time slot index t. If there exists a point whose spatial radius is less than or equal to R within a consecutive number of time slots... stay And the cumulative duration is greater than or equal to τ stay If so, the corresponding point is identified as a stop point; Step 2.2, the clustering of the stopping points, specifically: the DBSCAN clustering algorithm is used on the set formed by all the stopping points, and the obtained clusters represent potential key locations, with the centroid of each cluster taken as the location coordinates; Step 2.3, Activity Type Labeling: If the stay period is from 20:00 to 08:00 and the cumulative duration is ≥5 hours, it will be marked as "Residence (H)"; If the stay period is from 08:00 to 20:00 and the cumulative duration is ≥3 hours, it will be marked as "Workplace (W)"; The remaining stops are marked as "Other Locations (O)"; If a user has multiple clusters that meet the criteria of being labeled as "residence (H)" or "workplace (W)," then the Top-k clusters are selected based on the cumulative duration. Step 2.4: Extract the set of high-frequency activity locations of individuals based on trajectory dwell point detection and clustering algorithms, forming the key location set K. u ={L i u }; Among them, each key position L i u All are composed of coordinate points and activity function labels, and all automatically distinguish between residence, workplace and other activity locations through semantic recognition algorithms based on time thresholds and spatial clustering rules.

6. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 5, characterized in that, Step 3 is as follows: Step 3.1: Construct the set of active states S = {s1, s2, ..., s...} M }; This includes residence H, workplace W, and other locations O; Define the user's activity state A at time slot index t. u A u Let t ∈ state set S, t ∈ [1, L]. Then, for each user u, the activity sequence scanned in chronological order is [A1]. u A2 u ,...,A L u ]; Transition counting matrix C u This indicates that user u, throughout all time slices, from state s i Transition to state s j The number of occurrences is counted, and its size is the number of activity states M × the number of activity states M, expressed as: in, Where F(·) is an indicator function, which takes the value 1 if the condition is true, and 0 otherwise; L is the length of the user activity sequence; When the user's activity state in time slot index t is s i The next time slice t+1 is s. j Then in the counting matrix s i，j Add 1; Constructing a lifestyle matrix P u ; The lifestyle matrix P u Indicates from state s i Transition to state s j The probability is obtained by normalizing the counting matrix: in, Where λ is the smoothing term; the denominator is the state s. i The sum of all transition counts issued; ensuring that row vector probabilities are normalized, i.e.: Step 3.2: Matrix dimensionality reduction, for the lifestyle pattern matrix P u Perform Singular Value Decomposition (SVD): P u =U∑V T ; Where U is the individual feature matrix; Σ is the singular value matrix; V T The fundamental matrix of the pattern; Step 3.3: Generate the pattern matrix P′ u Specifically, based on the individual feature matrix U, distribution parameters are extracted, and a new individual feature matrix is ​​generated through random sampling. Next, the individual feature matrix With the singular value matrix Σ and the pattern fundamental matrix V T Multiply and reconstruct to generate a new lifestyle pattern matrix P′ u ; .

7. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 6, characterized in that, Step 4 is as follows: Step 4.1: Add Laplace noise to the location, for the set of key locations K u Two-dimensional Laplace noise is added to each key location coordinate (lat, lon): lat′=lat+η lat lon′=lon+η lon We obtain the key location set K″ after considering privacy encryption. u ; Among them: h lat ,or lon ~Lap(0,b), The privacy budget is allocated as follows: Total budget ε includes location budget ε pos With matrix budget ε mat That is, ε = ε pos +ε mat ; Step 4.2, in the lifestyle pattern matrix P′ u Noise addition affects the lifestyle matrix P′ u Each element P in i,j Adding Laplace noise yields a new transition probability matrix with added noise. : Where m is a matrix The column dimension.

8. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 7, characterized in that, Step 5 is as follows: Step 5.1: Define the state set S = the key location set K′ u ; Step 5.2: Use a Dynamic Bayesian Network (DBN) to describe the state transition. The mathematical expression is: P(X t =s j |X t-1 =s i )=A i,j ,P(O t |X t )=B i,j ; Among them, X t In active state, O t In observation state; Step 5.3, make state s j Corresponding to one or more specific coordinates (lat, lon); state s j When dealing with multiple specific coordinates (lat, lon), sampling is performed according to weights. Then the emission matrix is ​​used for position selection, and the mathematical expression is: Among them, the weight w can be determined based on the POI weight and population density; Step 5.4: Time-duration modeling. For each state s, the activity duration is fitted using a Gaussian Mixture Model (GMM). The mathematical expression is: The average dwell time error is controlled within ±0.25 hours; Step 5.5: Trajectory generation, sampling the initial state X0 from the user's historical distribution or sample prior; Each step is based on A t Sample the next state, based on the GMM sampling dwell time, sample the specific coordinates in that state according to the position weight, and record (t) start ,t end ,lat,lon,s), update time t=t+t stay until the required number of days is reached; Step 5.6: Smooth the trajectory and perform path fitting on the coordinates between adjacent time slots.

9. The method for generating city-level individual travel trajectories with a privacy protection mechanism according to claim 8, characterized in that, Step 6 is as follows: Step 6.1: Data statistics. Statistical analysis is performed on the time-period distribution and OD matrix raster population density of the generated and actual data to obtain the statistical result P. gen and P real ; Step 6.2: Define the overall relative deviation based on the combination of raster and time, and calculate the deviation using the following formula: Target threshold: If Bias > θ bias If so, an update will be triggered; Step 6.3, Parameter Update Rules: If the bias exceeds the threshold and the privacy budget ε is fixed, then adjust the parameters of the generative model and retrain the parameters of the GMM. If the bias exceeds the threshold and the privacy budget ε can be relaxed, then the privacy budget parameter should be adjusted first. After updating, regenerate small samples and evaluate again. If the bias does not improve, perform deeper parameter retraining, re-decompose SVD or adjust the DBN structure. Step 6.4: Output the results and generate the trajectory file GeoJSON; Each Feature represents a pause; Step 6.5: Visualization components, displaying points / heatmap layers through Leaflet / Mapbox on the Web map; time window slider controlling t animation; comparison panel displaying the difference between Pgen and Preal, i.e., time curve and heatmap difference map; traffic simulation system; access to the city operation monitoring platform.

10. A system, characterized in that, The method for generating city-level individual travel trajectories with privacy protection mechanisms as described in any one of claims 1 to 9 includes a data fusion and preprocessing module, a key location and lifestyle extraction module, a privacy protection generation module, a travel behavior generation module, and an online optimization and visualization module. The data fusion and preprocessing module is used to receive multi-source heterogeneous input data, complete data alignment through spatial matching and time synchronization algorithms, clean, denoise and rasterize the raw data, and generate a basic spatiotemporal feature set. The multi-source heterogeneous input data includes mobile signaling data, GPS trajectory data, POI data, land use data, population density and road network information; The key location and lifestyle pattern extraction module identifies an individual's key activity locations based on preprocessed travel data, including Home, Work, and Other activity locations. A Life Pattern Matrix is ​​constructed, using time period and location type as dimensions, to record the probability of an individual's movement at different time periods. An algorithm based on time thresholds and spatial clustering rules is used for identification: The place of residence is defined as the nighttime period, from 20:00 to 08:00, during which the stay is ≥5 hours. Workplace refers to the time spent during weekdays, from 08:00 to 20:00, and staying there for ≥3 hours. The privacy-preserving generation module introduces a differential privacy mechanism to add Laplace perturbations to the key location coordinates and activity transition probability matrix. It controls the perturbation intensity of data at different levels through a privacy budget allocation algorithm and uses resampling and probability normalization processes to ensure the statistical consistency of the perturbed data, so that the generated data satisfies the privacy protection constraint ε-privacy while maintaining the stability of behavioral distribution characteristics. The travel behavior generation module generates time-series travel sequences based on the perturbed key location set and lifestyle matrix, using an improved dynamic Bayesian network (DBN) or Markov chain model. It estimates the duration of each activity state using a Gaussian mixture model (GMM) and outputs an individual-level virtual travel trajectory containing time, location, and activity type. The online optimization and visualization module monitors the statistical deviation of the generated results in real time, adjusts the model parameters through an iterative feedback mechanism, outputs the generated travel data to the traffic simulation system or urban operation monitoring platform, and provides map visualization and data interface APIs for system integration.

Images