Metro scene user identification method and system based on trajectory segment compression

Abstract

The application provides a subway scene user identification method and system based on trajectory segment compression, adopts a bucket sorting strategy to perform user-level aggregation on standardized XDR data, performs two-stage trajectory segment compression processing on a user behavior feature sequence, performs feature construction processing on a stable trajectory sequence, obtains a unified aligned mobile user feature sequence, inputs an E-GLM encoder after mapping by using a numerical feature encoder, extracts a CLS global vector corresponding to a first CLS Token by using a bidirectional attention mechanism, and performs two-class calculation on the CLS global vector by using a linear classifier to obtain a user identification result, so as to determine whether the user is a subway scene user. The application constructs a stable user behavior sequence by using two-stage unstable switching processing and trajectory segment compression, and models long-time sequence features by using an E-GLM encoder, so as to realize identification of a subway scene user.

Description

Technical Field

[0001] This invention relates to the fields of temporal deep learning and intelligent user behavior recognition technology, and in particular to a user recognition method and system for subway scenes based on trajectory segment compression. Background Technology

[0002] With the rapid development of mobile communication networks and the large-scale deployment of 5G networks, the scale of data in operators' existing networks continues to grow, and the demand for user behavior analysis based on network-side data in the field of intelligent operation and maintenance of communication networks is constantly increasing. Extended Data Record (XDR) integrates multi-dimensional data such as user behavior information, wireless measurement indicators, and service session characteristics to obtain network-side user behavior features. Compared with traditional drive testing methods, XDR-based analysis methods have the advantages of wide coverage, strong collection continuity, and large data scale, and have become an important method for user behavior identification in communication networks.

[0003] In subway scenarios, XDR data-based user identification has become a common technique. However, existing technologies still have the following problems: rule-based methods rely on static parameters or manually set thresholds for discrimination, resulting in weak generalization ability and frequent parameter adjustments when network topology changes; traditional supervised learning methods (such as XGBoost and Random Forest) have limited ability to model the temporal features of user movement behavior, making it difficult to capture complex subway movement patterns; deep learning methods (such as RNN, LSTM, and CNN) have insufficient ability to model long-range dependencies when processing long-term time-series data; the subway environment is characterized by high-speed movement, enclosed spaces, dense multi-cell coverage, and frequent cell overlap, leading to frequent cell handovers by user equipment, resulting in a large number of unstable handovers and signaling fluctuations in XDR data, further increasing the difficulty of user behavior identification. Especially at the boundary between subway stations and outdoor areas, there are many boundary samples, making it difficult for traditional methods to effectively distinguish between real subway users and users in the surrounding area, resulting in a decrease in identification accuracy.

[0004] Therefore, how to construct stable user behavior sequences in a high-noise XDR data environment and further improve the modeling capability of long-term behavioral features to achieve user identification in complex subway scenarios has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for user identification in subway scenes based on trajectory segment compression. This invention constructs a stable user behavior sequence through two-stage unstable switching processing and trajectory segment compression, and combines an E-GLM encoder to model long-term temporal features, thereby realizing the identification of users in subway scenes.

[0006] In a first aspect, the technical solution provided by this invention is: a subway scene user identification method based on trajectory segment compression, comprising:

[0007] Obtain live XDR data, and perform time field standardization, numerical range compression, and missing value imputation preprocessing on the live XDR data to obtain standardized XDR data;

[0008] Based on the user identifier SUPI, a bucket sorting strategy is used to perform user-level aggregation on the standardized XDR data, and the data of the same user are sorted in chronological order to obtain a user behavior feature sequence.

[0009] A two-stage trajectory segment compression process is performed on the user behavior feature sequence to obtain a stable trajectory sequence, and the user's movement speed feature is calculated based on the stable trajectory sequence.

[0010] The stable trajectory sequence is subjected to feature construction processing to obtain a uniformly aligned user feature sequence;

[0011] The uniformly aligned user feature sequence is mapped to a high-dimensional latent space using a numerical feature encoder. The mapped high-dimensional feature sequence is then input into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS token. A linear classifier is then used to perform binary classification on the CLS global vector to obtain the user identification result, thereby determining whether the user is a subway scene user.

[0012] Preferably, the existing XDR data includes User Identifier (SUPI), Timestamp, Cell Identifier (NCI), Location Area Identifier (TAC), and Radio Measurement Features.

[0013] Preferably, based on the user identifier SUPI, a bucket sorting strategy is used to perform user-level aggregation on the standardized XDR data, and the data of the same user are sorted in chronological order to obtain a user behavior feature sequence, including:

[0014] The total number of distributed buckets B is set according to the cluster parallelism and daily data volume configuration, and the user identifier SUPI of each standardized XDR data is hashed to obtain the bucket number of the corresponding standardized XDR data.

[0015] The standardized XDR data is repartitioned according to the bucket number, and standardized XDR data with the same bucket number are assigned to the same computing partition.

[0016] Within each computational slice, a local hash dictionary is constructed using the user identifier SUPI as the grouping key. All standardized XDR data within the computational slice are traversed, and all standardized XDR data with the same user identifier SUPI are grouped under the same local hash dictionary entry. Each key-value pair in the local hash dictionary represents all standardized XDR data of a single user.

[0017] Sort all standardized XDR data corresponding to each user identifier SUPI in the local hash dictionary by standardized timestamp, and remove duplicate standardized XDR data with the same timestamp;

[0018] Then, the user-standardized XDR data whose length is less than the minimum effective record threshold is removed after sorting, and the user-standardized XDR data of all users are encapsulated into user behavior feature sequences.

[0019] Preferably, a two-stage trajectory segment compression process is performed on the user behavior feature sequence to obtain a stable trajectory sequence, including:

[0020] Traverse the user behavior feature sequence, identify jitter segments that frequently switch back and forth in a short time scale, and uniformly backfill and correct the jitter segments to the primary cell;

[0021] The corrected user behavior feature sequence is divided into multiple independent trajectory segments based on consecutive identical cell identifiers (NCIs).

[0022] Each independent trajectory segment is eliminated or compressed according to its length based on the elimination threshold and compression threshold; all the retained independent trajectory segments and the compressed single compressed segments are spliced ​​together into a first-level compressed sequence;

[0023] Traverse the first-level compressed sequence and identify back-slicing segments that reappear after a gap of multiple records as implicit jitter segments;

[0024] For the detected implicit jitter fragments, calculate their fragment length. If the fragment length of the implicit jitter fragment is greater than or equal to the compression threshold, then the implicit jitter fragment is compressed.

[0025] If the length of an implicit jittering segment is greater than the removal threshold but less than the compression threshold, the current implicit jittering segment is retained.

[0026] All the corrected trajectory segments are reassembled in timestamp order to obtain a stable trajectory sequence.

[0027] Preferably, each independent trajectory segment is eliminated or compressed based on its length according to a rejection threshold and a compression threshold, including:

[0028] If the length of an independent trajectory segment is less than or equal to the removal threshold, then the current independent trajectory segment is removed.

[0029] If the length of an independent trajectory segment is greater than or equal to the compression threshold, the independent trajectory segment is compressed into a single compressed segment.

[0030] If the length of an independent trajectory segment is greater than the removal threshold but less than the compression threshold, the current independent trajectory segment is directly retained.

[0031] Preferably, if an independent trajectory segment meets the compression conditions, the start and end timestamps of a single compressed segment are generated by averaging and aggregating the standardized XDR data of all users within the independent trajectory segment, i.e.:

[0032] , ;

[0033] In the formula, , Independent trajectory segments The start and end timestamps of the t-th user-standardized XDR data before compression; , These are the start and end timestamps of the compressed segment after compressing the original trajectory segment k, respectively; Independent trajectory segments Length;

[0034] And map the representative cell identifier (NCI) of a single compressed segment to a specific longitude. and latitude And mark the stability state bit. ,Right now:

[0035] ;

[0036] In the formula, This is a latitude and longitude mapping function.

[0037] Preferably, the numerical feature encoder adopts a multilayer perceptron (MLP) structure, which maps the uniformly aligned user feature sequence to a high-dimensional latent space through nonlinear transformation to obtain a high-dimensional feature sequence.

[0038] Preferably, the E-GLM encoder removes the unidirectional causal mask, adopts a bidirectional attention mechanism, and is equipped with a dual Pre-LN structure, an RMSnorm normalization module, and a SwiGLU feedforward network.

[0039] Preferably, high-dimensional feature sequences Input the E-GLM encoder, first process the high-dimensional feature sequence The RMSNorm root mean square normalization process is performed, and then the attention output features are extracted through a bidirectional attention mechanism and compared with the high-dimensional feature sequence. The residuals yield intermediate feature sequences. ,Right now:

[0040] ;

[0041] For intermediate feature sequences After RMSOrm normalization, the signal is fed into a Swiglu-activated FFN feedforward network to perform nonlinear feature enhancement and integrate with the intermediate feature sequence. Residual connections are performed to obtain the output characteristics of a single-layer E-GLM encoder. ,Right now:

[0042] ;

[0043] In the formula, It is a SwiGLU feedforward network; This is the RMSNorm normalization module.

[0044] After processing by a multi-layer E-GLM encoder, the high-dimensional vector corresponding to the first CLSToken of the user's feature sequence, which is truncated and uniformly aligned, is obtained as the CLS global vector of the user trajectory. .

[0045] Preferably, the CLS global vector The classification header is used to map categories, resulting in a category probability distribution. The category label corresponding to the highest probability is then used as the final category label.

[0046] ;

[0047] ;

[0048] In the formula, For users The final category label, if Determine the user For users in subway scenarios; For users Category score; , These are the weight matrix and bias term of the classification head, respectively.

[0049] Secondly, the technical solution provided by this invention is: a subway scene user identification system based on trajectory segment compression, comprising:

[0050] The data acquisition module is used to acquire live XDR data and perform preprocessing on the live XDR data, including time field standardization, numerical range compression, and missing value imputation, to obtain standardized XDR data.

[0051] The user behavior feature extraction module is used to perform user-level aggregation on standardized XDR data based on user identifier SUPI, using a bucket sorting strategy, and sort the data of the same user in chronological order to obtain a user behavior feature sequence.

[0052] A two-stage compression module is used to perform two-stage trajectory segment compression processing on the user behavior feature sequence to obtain a stable trajectory sequence, and to calculate the user's movement speed features based on the stable trajectory sequence.

[0053] The user mobility feature construction module is used to perform feature construction processing on the stable trajectory sequence to obtain a uniformly aligned user mobility feature sequence.

[0054] The user prediction module is used to map the uniformly aligned user feature sequence to a high-dimensional latent space using a numerical feature encoder, and input the mapped high-dimensional feature sequence into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS token, and a linear classifier is used to perform binary classification calculation on the CLS global vector to obtain the user identification result, so as to determine whether the user is a subway scene user.

[0055] Thirdly, the present invention provides an electronic device, comprising:

[0056] At least one processor;

[0057] and a memory communicatively connected to the at least one processor;

[0058] The memory stores a computer program that can be executed by at least one processor. When the computer program is executed by at least one processor, it implements the subway scene user identification method based on trajectory segment compression.

[0059] The beneficial technical effects of this invention are as follows:

[0060] 1. This invention reduces the sequence length through two-stage unstable switching processing and trajectory segment compression, thereby reducing subsequent computational complexity, improving processing efficiency, and effectively reducing the interference of signaling jitter and frequent cell handover on user trajectory expression in subway scenarios, thus improving the stability of trajectory sequences;

[0061] 2. This invention uses an E-GLM encoder to perform bidirectional temporal modeling of stable trajectory sequences, adapts to the non-causal characteristics of subway trajectory sequences to capture long-distance temporal dependencies, improves the ability to express long-range user behavior features in complex subway scenarios, enhances the ability to characterize user movement behavior patterns and the accuracy of boundary sample recognition, and achieves accurate recognition of user behavior in complex subway scenarios.

[0062] 3. This invention detects and corrects explicit unstable switching at short time scales and performs trajectory segment compression processing; then it performs stability correction on implicit unstable switching at medium time scales, and finally obtains a stable trajectory sequence, thereby reducing the impact of signaling noise on trajectory representation and improving the stability and modelability of user behavior sequences. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention;

[0064] Figure 2 This is a structural framework diagram of the E-GLM encoder according to an embodiment of the present invention;

[0065] Figure 3 This is a comparison chart of the accuracy of subway user identification in embodiments of the present invention;

[0066] Figure 4 This is a structural framework diagram of the system according to an embodiment of the present invention;

[0067] In the diagram, 10 is the data acquisition module; 20 is the user behavior feature extraction module; 30 is the two-stage compression module; 40 is the user movement feature construction module; and 50 is the user prediction module. Detailed Implementation

[0068] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0069] like Figure 1 As shown, this embodiment provides a user identification method for subway scenes based on trajectory segment compression, including:

[0070] S1: Obtain the existing XDR data, and perform time field standardization, numerical range compression, and missing value filling preprocessing on the existing XDR data to obtain standardized XDR data;

[0071] S2: Based on the user identifier SUPI, a bucket sorting strategy is used to perform user-level aggregation on the standardized XDR data, and the data of the same user are sorted in chronological order to obtain the user behavior feature sequence.

[0072] S3: Perform two-stage trajectory segment compression processing on the user behavior feature sequence to obtain a stable trajectory sequence, and calculate the user's movement speed feature based on the stable trajectory sequence;

[0073] S4: Perform feature construction processing on the stable trajectory sequence to obtain a uniformly aligned user feature sequence;

[0074] S5: The uniformly aligned user feature sequence is mapped to a high-dimensional latent space using a numerical feature encoder. The mapped high-dimensional feature sequence is then input into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLSToken. A linear classifier is then used to perform binary classification on the CLS global vector to obtain the user identification result, thereby determining whether the user is a subway scene user.

[0075] In this embodiment, in step S1, the existing XDR data includes user identifier SUPI, timestamp, cell identifier NCI, location area identifier TAC, and radio measurement characteristics.

[0076] In this embodiment, in step S2, based on the user identifier SUPI, a bucket sorting strategy is used to perform user-level aggregation on the standardized XDR data, and the data of the same user are sorted in chronological order to obtain a user behavior feature sequence, including:

[0077] S21: Configure the total number of distributed buckets B based on the cluster parallelism and daily data volume, and perform a hash operation on the user identifier SUPI of each standardized XDR data to obtain the corresponding bucket number of the standardized XDR data. ,Right now:

[0078] ;

[0079] In the formula, The bucket number used in the hash operation; This is a hash operation; To perform a modulo operation on the total number of buckets B;

[0080] S22: Repartition the standardized XDR data according to the bucket number, and allocate the standardized XDR data with the same bucket number to the same computing partition;

[0081] S23: Within each computing slice, a local hash dictionary is constructed using the user identifier SUPI as the grouping key. All standardized XDR data within the computing slice are traversed, and all standardized XDR data with the same user identifier SUPI are grouped under the same local hash dictionary entry. Each key-value pair in the local hash dictionary represents all standardized XDR data of a single user.

[0082] S24: Sort all standardized XDR data corresponding to each user identifier SUPI in the local hash dictionary by standardized timestamp, and remove duplicate standardized XDR data with the same timestamp.

[0083] S25: Remove user-standardized XDR data whose length is less than the minimum effective record threshold after sorting, and encapsulate the user-standardized XDR data of all users into a user behavior feature sequence.

[0084] In this embodiment, step S3 involves performing a two-stage trajectory segment compression process on the user behavior feature sequence to obtain a stable trajectory sequence, and calculating the user's movement speed features based on the stable trajectory sequence, including:

[0085] In this embodiment, a two-stage trajectory segment compression process is performed on the user behavior feature sequence to obtain a stable trajectory sequence, including:

[0086] S31: Traverse the user behavior feature sequence to identify jitter segments that frequently switch back and forth on a short time scale, i.e.: ; and uniformly backfill and correct the jitter segments in the main residential area. ;

[0087] S32: Divide the corrected user behavior feature sequence into multiple independent trajectory segments based on consecutive identical cell identifiers (NCIs). Each independent trajectory segment... Represented as:

[0088]

[0089] In the formula, , They represent the first Start and end time indices for each independent trajectory segment; Indicates the first The representative cell identifier (NCI) for each independent trajectory segment; This represents the set of all segment indices within a trajectory segment whose NCI remains unchanged for the k-th independent cell;

[0090] S33: For each independent trajectory segment, based on the elimination threshold and compression threshold. Elimination or compression based on length includes:

[0091] If independent trajectory segments length Less than or equal to the rejection threshold Then remove the current independent trajectory segment. ;

[0092] If independent trajectory segments length Greater than or equal to the compression threshold Then the independent trajectory segments Compressed into a single compressed segment ;

[0093] If independent trajectory segments length Greater than the rejection threshold And less than the compression threshold Then the current independent trajectory segment is directly retained. .

[0094] S34: If independent trajectory segments Satisfying compression conditions, through independent trajectory segments The start and end timestamps of a single compressed segment are generated by averaging and aggregating the standardized XDR data of all users within the segment, i.e.:

[0095] , ;

[0096] In the formula, , Independent trajectory segments The start and end timestamps of the t-th user-standardized XDR data before compression; , These are the start and end timestamps of the compressed segment after compressing the original trajectory segment k, respectively; Independent trajectory segments Length;

[0097] And map the representative cell identifier (NCI) of a single compressed segment to a specific longitude. and latitude And mark the stability state bit. ,Right now:

[0098] ;

[0099] In the formula, It is a latitude and longitude mapping function; Indicates the first The representative cell identifier (NCI) for each independent trajectory segment; The NCI is the representative cell identifier for the compressed segment.

[0100] All the retained independent trajectory segments and the compressed single compressed segments are spliced ​​together into a first-level compressed sequence.

[0101] S35: Traverse the first-level compressed sequence and identify back-slicing segments that reappear after multiple records as implicit jitter segments, i.e.: ;

[0102] For each detected implicit jitter segment, its segment length is calculated. If the segment length of the implicit jitter segment is greater than or equal to the compression threshold... Compress according to step S34;

[0103] If the length of the implicit jittering segment is greater than the rejection threshold And less than the compression threshold Preserve the current implicit jitter segment;

[0104] S36: Reassemble all the corrected trajectory segments in timestamp order to obtain a stable trajectory sequence.

[0105] In this embodiment, calculating the user's movement speed characteristics based on the stable trajectory sequence includes:

[0106] Obtain any two consecutive compressed trajectory segments from the stable trajectory sequence. and Longitude and latitude , The spherical distance formula is used to calculate any two consecutive compressed trajectory segments. and Straight-line geographical distance between base stations ,Right now:

[0107] ;

[0108] ;

[0109] In the formula, The average radius of the Earth As an intermediate variable;

[0110] Start timestamp based on compressed trajectory segments and Calculate the time difference between two consecutive compressed trajectory segments. ;

[0111] Based on any two consecutive compressed trajectory segments and Straight-line geographical distance between base stations and time difference Calculate user movement speed characteristics .

[0112] In this embodiment, step S4 involves performing feature construction processing on the stable trajectory sequence to obtain a uniformly aligned user feature sequence, including:

[0113] S41: Based on the start timestamps of any two adjacent compressed trajectory segments and Calculate the time difference features of the differential and compressed versions. ,Right now:

[0114] ;

[0115] S42: Based on two compressed trajectory segments and Longitude and latitude , The difference calculation determines the longitude and latitude characteristics of the compression; that is:

[0116] ;

[0117] ;

[0118] In the formula, , These are the compressed longitude features and compressed latitude features, respectively; Geographic threshold; This is the scaling function.

[0119] S43: Perform segmented mapping on the user movement speed features to obtain compressed speed features. ,Right now:

[0120] ;

[0121] In the formula, , , The segmented velocity threshold; For the first Velocity characteristics of a compressed trajectory segment.

[0122] S44: Introducing NCI Switching Scoring Number of stays score and length of stay rating ;Right now:

[0123] ;

[0124] ;

[0125] ;

[0126] In the formula, For the first Total number of cell handovers within the original trajectory; For the first The total duration of each original trajectory; For the first The dwell length of each original trajectory; This represents the minimum residency threshold. Indicates the first The total length of dwell time for all segments within a single original trajectory; This represents the total number of all segments within the original trajectory. For the first The number of records within the original trajectory that did not experience NCI switching; Indicates the first The number of original records for each original trajectory.

[0127] S45: Time difference features Compressed longitude characteristics and compressed latitude characteristics , NCI switching score Number of stays score and length of stay rating The lengths of the trajectory segments are concatenated to form the user movement feature vector. ;

[0128] S46: Constructing and User Motion Feature Vectors A zero-dimensional vector of the same dimension is used as the class token (CLS Token), and the CLS Token is placed in the user movement feature vector. The first digit is used to obtain a uniformly aligned sequence of mobile user features. .

[0129] In this embodiment, step S5 involves using a numerical feature encoder to map the uniformly aligned user feature sequence to a high-dimensional latent space, and then inputting the mapped high-dimensional feature sequence into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS token, and a linear classifier is used to perform binary classification calculation on the CLS global vector to obtain the user identification result, thereby determining whether the user is a subway scene user. This includes:

[0130] In this embodiment, the numerical feature encoder adopts a multilayer perceptron (MLP) structure, and maps the uniformly aligned user feature sequence to a high-dimensional latent space through nonlinear transformation to obtain a high-dimensional feature sequence. ,Right now:

[0131] .

[0132] In this embodiment, as Figure 2As shown, the E-GLM encoder removes the unidirectional causal mask, adopts a bidirectional attention mechanism, and is equipped with a dual Pre-LN structure, an RMSnorm normalization module, and a SwiGLU feedforward network.

[0133] High-dimensional feature sequences Input the E-GLM encoder, first process the high-dimensional feature sequence The RMSNorm root mean square normalization process is performed, and then the attention output features are extracted through a bidirectional attention mechanism and compared with the high-dimensional feature sequence. The residuals yield intermediate feature sequences. ,Right now:

[0134] ;

[0135] For intermediate feature sequences After RMSOrm normalization, the signal is fed into a Swiglu-activated FFN feedforward network to perform nonlinear feature enhancement and integrate with the intermediate feature sequence. Residual connections are performed to obtain the output characteristics of a single-layer E-GLM encoder. ,Right now:

[0136] ;

[0137] In the formula, It is a SwiGLU feedforward network; This is the RMSNorm normalization module.

[0138] After processing by a multi-layer E-GLM encoder, the high-dimensional vector corresponding to the first CLSToken of the user's feature sequence, which is truncated and uniformly aligned, is obtained as the CLS global vector of the user trajectory. .

[0139] In this embodiment, the CLS global vector is... The classification header is used to map categories, resulting in a category probability distribution. The category label corresponding to the highest probability is then used as the final category label.

[0140] ;

[0141] ;

[0142] In the formula, For users The final category label, if Determine the user For users in subway scenarios; For users Category score; , These are the weight matrix and bias term of the classification head, respectively.

[0143] like Figure 3 As shown, the accuracy of this embodiment compared with existing recognition methods in subway and non-subway scenarios is illustrated. The results show that the method proposed in this embodiment, based on trajectory segment compression and E-GLM encoder, significantly improves the recognition accuracy compared with traditional rule-based discrimination methods and basic supervised learning algorithms when dealing with complex subway lines, including outdoor elevated stations. Furthermore, while achieving the same or higher accuracy, the training parameters of the model in this invention are lower than those of existing general-purpose large model structures, demonstrating the technical effectiveness of this structural improvement in reducing computational overhead, improving inference efficiency, and enhancing scenario robustness.

[0144] like Figure 4 As shown, embodiments of this application also provide a subway scene user identification system based on trajectory segment compression, including:

[0145] Data acquisition module 10 is used to acquire live XDR data and perform time field standardization, numerical range compression and missing value filling preprocessing on the live XDR data to obtain standardized XDR data;

[0146] The user behavior feature extraction module 20 is used to perform user-level aggregation on standardized XDR data based on user identifier SUPI, using a bucket sorting strategy, and sort the data of the same user in chronological order to obtain a user behavior feature sequence.

[0147] The two-stage compression module 30 is used to perform two-stage trajectory segment compression processing on the user behavior feature sequence to obtain a stable trajectory sequence, and calculate the user's movement speed feature based on the stable trajectory sequence.

[0148] User mobility feature construction module 40 is used to perform feature construction processing on the stable trajectory sequence to obtain a uniformly aligned user mobility feature sequence.

[0149] The user prediction module 50 is used to map the uniformly aligned user feature sequence to a high-dimensional latent space using a numerical feature encoder, and input the mapped high-dimensional feature sequence into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS Token, and a linear classifier is used to perform binary classification calculation on the CLS global vector to obtain the user identification result, so as to determine whether the user is a subway scene user.

[0150] In this embodiment, the XDR data collected by the data acquisition module 10 includes user identifier SUPI, timestamp, cell identifier NCI, location area identifier TAC, and wireless measurement features.

[0151] In this embodiment, the user behavior feature extraction module 20, based on the user identifier SUPI, performs user-level aggregation on the standardized XDR data using a bucket sorting strategy, and sorts the data of the same user in chronological order to obtain a user behavior feature sequence, including:

[0152] The total number of distributed buckets B is configured based on the cluster parallelism and daily data volume. A hash operation is then performed on the user identifier SUPI of each standardized XDR data entry to obtain the corresponding bucket number. ,Right now:

[0153] ;

[0154] In the formula, The bucket number used in the hash operation; This is a hash operation; To perform a modulo operation on the total number of buckets B.

[0155] The standardized XDR data is repartitioned based on the bucket number, and standardized XDR data with the same bucket number are assigned to the same computing partition.

[0156] Within each computational slice, a local hash dictionary is constructed using the user identifier SUPI as the grouping key. All standardized XDR data within the computational slice are traversed, and all standardized XDR data with the same user identifier SUPI are grouped under the same local hash dictionary entry. Each key-value pair in the local hash dictionary represents all standardized XDR data of a single user.

[0157] Sort all standardized XDR data corresponding to each user identifier SUPI in the local hash dictionary by standardized timestamp, and remove duplicate standardized XDR data with the same timestamp;

[0158] Then, the user-standardized XDR data whose length is less than the minimum effective record threshold is removed after sorting, and the user-standardized XDR data of all users are encapsulated into user behavior feature sequences.

[0159] In this embodiment, the two-stage compression module 30 performs two-stage trajectory segment compression processing on the user behavior feature sequence to obtain a stable trajectory sequence, including:

[0160] Traverse the user behavior feature sequence to identify jitter segments that frequently switch back and forth within a short time scale, i.e.: ; and uniformly backfill and correct the jitter segments in the main residential area. ;

[0161] The corrected user behavior feature sequence is divided into multiple independent trajectory segments based on consecutive identical cell identifiers (NCIs). Each independent trajectory segment... Represented as:

[0162]

[0163] In the formula, , They represent the first The start and end indices of each independent trajectory segment; Indicates the first The representative cell identifier (NCI) for each independent trajectory segment; This represents the set of all segment indices within a trajectory segment whose NCI remains unchanged for the k-th independent cell;

[0164] Each independent trajectory segment is evaluated based on a rejection threshold and a compression threshold. The segments are discarded or compressed according to length, and all the retained independent trajectory segments, along with the compressed single segments, are spliced ​​together into a first-level compressed sequence, including:

[0165] If independent trajectory segments length Less than or equal to the rejection threshold Then remove the current independent trajectory segment. ;

[0166] If independent trajectory segments length Greater than or equal to the compression threshold Then the independent trajectory segments Compressed into a single compressed segment ;

[0167] If independent trajectory segments length Greater than the rejection threshold And less than the compression threshold Then the current independent trajectory segment is directly retained. .

[0168] If independent trajectory segments Satisfying compression conditions, through independent trajectory segments The start and end timestamps of a single compressed segment are generated by averaging and aggregating the standardized XDR data of all users within the segment, i.e.:

[0169] , ;

[0170] In the formula, , These are the start and end timestamps of the compressed segment after compressing the original trajectory segment k, respectively; , Independent trajectory segments The start and end timestamps of the t-th user-standardized XDR data before compression; Independent trajectory segments Length;

[0171] And map the representative cell identifier (NCI) of a single compressed segment to a specific longitude. and latitude And mark the stability state bit. ,Right now:

[0172] ;

[0173] In the formula, This is a latitude and longitude mapping function.

[0174] Traverse the first-level compressed sequence and identify back-slicing segments that reappear after a gap of multiple records as implicit jitter segments, i.e.: ;

[0175] For each detected implicit jitter segment, its segment length is calculated. If the segment length of the implicit jitter segment is greater than or equal to the compression threshold... Compress according to step S34;

[0176] If the length of the implicit jittering segment is greater than the rejection threshold And less than the compression threshold Preserve the current implicit jitter segment;

[0177] All the corrected trajectory segments are reassembled in timestamp order to obtain a stable trajectory sequence.

[0178] In this embodiment, the two-stage compression module 30 calculates the user's movement speed characteristics based on the stable trajectory sequence, including:

[0179] Obtain any two consecutive compressed trajectory segments from the stable trajectory sequence. and Longitude and latitude , The spherical distance formula is used to calculate any two consecutive compressed trajectory segments. and Straight-line geographical distance between base stations ,Right now:

[0180] ;

[0181] ;

[0182] In the formula, The average radius of the Earth As an intermediate variable;

[0183] Based on trajectory segments and The start timestamp of the trajectory segment and Calculate two consecutive trajectory segments and Time difference ;

[0184] Based on any two consecutive compressed trajectory segments and Straight-line geographical distance between base stations and time difference Calculate user movement speed characteristics .

[0185] In this embodiment, the user mobility feature construction module 40 performs feature construction processing on the stable trajectory sequence to obtain a uniformly aligned user mobility feature sequence, including:

[0186] Based on the start timestamps of any two adjacent compressed trajectory segments and Calculate the time difference features of the differential and compressed versions. ,Right now:

[0187] ;

[0188] Based on two compressed trajectory segments and Longitude and latitude , The difference calculation determines the longitude and latitude characteristics of the compression; that is:

[0189] ;

[0190] ;

[0191] In the formula, , These are the compressed longitude features and compressed latitude features, respectively; Geographic threshold; This is the scaling function.

[0192] The user movement speed feature is segmented and mapped to obtain the compression speed feature. ,Right now:

[0193] ;

[0194] In the formula, , , The segmented velocity threshold; For the first Velocity characteristics of a compressed trajectory segment.

[0195] Introducing NCI switching score Number of stays score and length of stay rating ;Right now:

[0196] ;

[0197] ;

[0198] ;

[0199] In the formula, For the first Total number of cell handovers within the original trajectory; For the first The total duration of each original trajectory; For the first The dwell length of each original trajectory; This represents the minimum residency threshold. Indicates the first The total length of dwell time for all segments within a single original trajectory; This represents the total number of all segments within the original trajectory. For the first The number of records within the original trajectory that did not experience NCI switching; Indicates the first The number of original records for each original trajectory.

[0200] Time difference features Compressed longitude characteristics and compressed latitude characteristics , NCI switching score Number of stays score and length of stay rating The lengths of the trajectory segments are concatenated to form the user movement feature vector. ;

[0201] Construction and user movement feature vectors A zero-dimensional vector of the same dimension is used as the category token CLSToken, and the category token CLSToken is placed in the user movement feature vector. The first digit is used to obtain a uniformly aligned sequence of mobile user features. .

[0202] In this embodiment, the user prediction module 50 uses a numerical feature encoder to map the uniformly aligned user feature sequence to a high-dimensional latent space, and inputs the mapped high-dimensional feature sequence into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS token, and a linear classifier is used to perform binary classification on the CLS global vector to obtain the user identification result, thereby determining whether the user is a user in the subway scene. This includes:

[0203] In this embodiment, the numerical feature encoder adopts a multilayer perceptron (MLP) structure, and maps the uniformly aligned user feature sequence to a high-dimensional latent space through nonlinear transformation to obtain a high-dimensional feature sequence. ,Right now:

[0204] .

[0205] In this embodiment, the E-GLM encoder removes the unidirectional causal mask, adopts a bidirectional attention mechanism, and is equipped with a dual Pre-LN structure, an RMSnorm normalization module, and a SwiGLU feedforward network.

[0206] High-dimensional feature sequences Input the E-GLM encoder, first process the high-dimensional feature sequence The RMSNorm root mean square normalization process is performed, and then the attention output features are extracted through a bidirectional attention mechanism and compared with the high-dimensional feature sequence. The residuals yield intermediate feature sequences. ,Right now:

[0207] ;

[0208] For intermediate feature sequences After RMSOrm normalization, the signal is fed into a Swiglu-activated FFN feedforward network to perform nonlinear feature enhancement and integrate with the intermediate feature sequence. Residual connections are performed to obtain the output characteristics of a single-layer E-GLM encoder. ,Right now:

[0209] ;

[0210] In the formula, It is a SwiGLU feedforward network; This is the RMSNorm normalization module.

[0211] After processing by a multi-layer E-GLM encoder, the high-dimensional vector corresponding to the first CLSToken of the user's feature sequence, which is truncated and uniformly aligned, is obtained as the CLS global vector of the user trajectory. .

[0212] In this embodiment, the CLS global vector is... The classification header is used to map categories, resulting in a category probability distribution. The category label corresponding to the highest probability is then used as the final category label.

[0213] ;

[0214] ;

[0215] In the formula, For users The final category label, if Determine the user For users in subway scenarios; For users Category score; , These are the weight matrix and bias term of the classification head, respectively.

[0216] Embodiments of this application also provide an electronic device, including:

[0217] At least one processor;

[0218] and a memory communicatively connected to the at least one processor;

[0219] The memory stores a computer program that can be executed by at least one processor. When the computer program is executed by at least one processor, it implements the subway scene user identification method based on trajectory segment compression described in the embodiment.

[0220] An embodiment of this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the subway scene user identification method based on trajectory segment compression described in the embodiment.

[0221] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A user identification method for subway scenes based on trajectory segment compression, characterized in that, include: Obtain live XDR data, and perform time field standardization, numerical range compression, and missing value imputation preprocessing on the live XDR data to obtain standardized XDR data; Based on the user identifier SUPI, a bucket sorting strategy is used to perform user-level aggregation on the standardized XDR data, and the data of the same user are sorted in chronological order to obtain a user behavior feature sequence. A two-stage trajectory segment compression process is performed on the user behavior feature sequence to obtain a stable trajectory sequence, and the user's movement speed feature is calculated based on the stable trajectory sequence. The stable trajectory sequence is subjected to feature construction processing to obtain a uniformly aligned user feature sequence; The uniformly aligned user feature sequence is mapped to a high-dimensional latent space using a numerical feature encoder. The mapped high-dimensional feature sequence is then input into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS token. A linear classifier is then used to perform binary classification on the CLS global vector to obtain the user identification result, thereby determining whether the user is a subway scene user.

2. The subway scene user identification method according to claim 1, characterized in that: The aforementioned existing XDR data includes User Identifier (SUPI), timestamp, Cell Identifier (NCI), Location Area Identifier (TAC), and radio measurement characteristics.

3. The subway scene user identification method according to claim 2, characterized in that: Based on the user identifier SUPI, a bucket sort strategy is used to perform user-level aggregation on the standardized XDR data, and the data of the same user are sorted in chronological order to obtain a user behavior feature sequence, including: The total number of distributed buckets B is set according to the cluster parallelism and daily data volume configuration, and the user identifier SUPI of each standardized XDR data is hashed to obtain the bucket number of the corresponding standardized XDR data. The standardized XDR data is repartitioned according to the bucket number, and standardized XDR data with the same bucket number are assigned to the same computing partition. Within each computational slice, a local hash dictionary is constructed using the user identifier SUPI as the grouping key. All standardized XDR data within the computational slice are traversed, and all standardized XDR data with the same user identifier SUPI are grouped under the same local hash dictionary entry. Each key-value pair in the local hash dictionary represents all standardized XDR data of a single user. Sort all standardized XDR data corresponding to each user identifier SUPI in the local hash dictionary by standardized timestamp, and remove duplicate standardized XDR data with the same timestamp; Then, the user-standardized XDR data whose length is less than the minimum effective record threshold is removed after sorting, and the user-standardized XDR data of all users are encapsulated into user behavior feature sequences.

4. The subway scene user identification method according to claim 3, characterized in that: A two-stage trajectory segment compression process is performed on the user behavior feature sequence to obtain a stable trajectory sequence, including: Traverse the user behavior feature sequence, identify jitter segments that frequently switch back and forth in a short time scale, and uniformly backfill and correct the jitter segments to the primary cell; The corrected user behavior feature sequence is divided into multiple independent trajectory segments based on consecutive identical cell identifiers (NCIs). Each independent trajectory segment is eliminated or compressed according to its length based on the elimination threshold and compression threshold; all the retained independent trajectory segments and the compressed single compressed segments are spliced ​​together into a first-level compressed sequence; Traverse the first-level compressed sequence and identify back-slicing segments that reappear after a gap of multiple records as implicit jitter segments; For the detected implicit jitter fragments, calculate their fragment length. If the fragment length of the implicit jitter fragment is greater than or equal to the compression threshold, then the implicit jitter fragment is compressed. If the length of an implicit jittering segment is greater than the removal threshold but less than the compression threshold, the current implicit jittering segment is retained. All the corrected trajectory segments are reassembled in timestamp order to obtain a stable trajectory sequence.

5. The subway scene user identification method according to claim 4, characterized in that: If an independent trajectory segment meets the compression conditions, the start and end timestamps of a single compressed segment are generated by averaging and aggregating the standardized XDR data of all users within the independent trajectory segment, i.e.: ， ； In the formula, , Independent trajectory segments The start and end timestamps of the t-th user-standardized XDR data before compression; , These are the start and end timestamps of the compressed segment after compressing the original trajectory segment k, respectively; , Independent trajectory segments The start and end timestamps of the t-th user-standardized XDR data before compression; Independent trajectory segments Length; , They represent the first Start and end time indices for each independent trajectory segment; And map the representative cell identifier (NCI) of a single compressed segment to a specific longitude. and latitude And mark the stability state bit. ,Right now: ； In the formula, It is a latitude and longitude mapping function; Indicates the first The representative cell identifier (NCI) for each independent trajectory segment; The NCI is the representative cell identifier for the compressed segment.

6. The subway scene user identification method according to claim 1, characterized in that: The stable trajectory sequence is subjected to feature construction processing to obtain a uniformly aligned user feature sequence, including: Based on the start timestamps of any two adjacent compressed trajectory segments, calculate the difference and the time difference features of the compressed time. The longitude and latitude characteristics of the compression are calculated based on the longitude and latitude differences between the two compressed trajectory segments. The user movement speed features are segmented and mapped to obtain compressed speed features; Introduce NCI switching score, number of stays score, and stay duration score; The time difference features, compressed longitude features and compressed latitude features, NCI handover score, dwell time score and dwell duration score, and the length of the trajectory segment are concatenated into a user movement feature vector. Construct a zero vector with the same dimension as the user mobility feature vector as the category token (CLS Token), and place the CLS Token at the beginning and end of the user mobility feature vector to obtain a uniformly aligned user feature sequence.

7. The subway scene user identification method according to claim 1, characterized in that: The E-GLM encoder removes the unidirectional causal mask, adopts a bidirectional attention mechanism, and is equipped with a dual Pre-LN structure, an RMSnorm normalization module, and a SwiGLU feedforward network.

8. The subway scene user identification method according to claim 7, characterized in that: High-dimensional feature sequences Input the E-GLM encoder, first process the high-dimensional feature sequence The RMSNorm root mean square normalization process is performed, and then the attention output features are extracted through a bidirectional attention mechanism and compared with the high-dimensional feature sequence. The residuals yield intermediate feature sequences. ,Right now: ； For intermediate feature sequences After RMSOrm normalization, the signal is fed into a Swiglu-activated FFN feedforward network to perform nonlinear feature enhancement and integrate with the intermediate feature sequence. Residual connections are performed to obtain the output characteristics of a single-layer E-GLM encoder. ,Right now: ； In the formula, It is a SwiGLU feedforward network; For RMSNorm normalization module; After processing by a multi-layer E-GLM encoder, the high-dimensional vector corresponding to the first CLS Token of the user's feature sequence, which is then uniformly aligned, is obtained as the CLS global vector of the user trajectory. ; CLS global vector The classification header is used to map categories, resulting in a category probability distribution. The category label corresponding to the highest probability is then used as the final category label. ； ； In the formula, For users The final category label, if Determine the user For users in subway scenarios; For users Category score; , These are the weight matrix and bias term of the classification head, respectively.

9. A subway scene user recognition system based on trajectory segment compression, characterized in that, include: The data acquisition module is used to acquire live XDR data and perform preprocessing on the live XDR data, including time field standardization, numerical range compression, and missing value imputation, to obtain standardized XDR data. The user behavior feature extraction module is used to perform user-level aggregation on standardized XDR data based on user identifier SUPI, using a bucket sorting strategy, and sort the data of the same user in chronological order to obtain a user behavior feature sequence. A two-stage compression module is used to perform two-stage trajectory segment compression processing on the user behavior feature sequence to obtain a stable trajectory sequence, and to calculate the user's movement speed features based on the stable trajectory sequence. The user mobility feature construction module is used to perform feature construction processing on the stable trajectory sequence to obtain a uniformly aligned user mobility feature sequence. The user prediction module is used to map the uniformly aligned user feature sequence to a high-dimensional latent space using a numerical feature encoder, and input the mapped high-dimensional feature sequence into an E-GLM encoder. A bidirectional attention mechanism is used to extract the CLS global vector corresponding to the first CLS token, and a linear classifier is used to perform binary classification calculation on the CLS global vector to obtain the user identification result, so as to determine whether the user is a subway scene user.

10. An electronic device, comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by at least one processor, characterized in that, when the computer program is executed by at least one processor, it implements the subway scene user identification method based on trajectory segment compression as described in any one of claims 1-8.

Images