Vehicle Beidou positioning data abnormity restoration method based on deep learning

Abstract

The invention relates to the technical field of Beidou positioning, in particular to a vehicle Beidou positioning data abnormity repairing method based on deep learning. Comprising the following steps: multi-source pretreatment; generating an abnormal mask; trajectory completion: constructing a deep learning model which is based on an Encoder-Decoder structure, fuses an attention mechanism and introduces assistance of an auto-encoder, and adopts an end-to-end training frame to pass through a multi-objective loss function comprising a mean square error loss item of a position error, a trajectory curvature continuity constraint loss item and a physical rationality constraint loss item; and real-time output and system deployment are fused. According to the method, the space-time association relationship and long-range dependency in the vehicle trajectory data are captured through the deep learning technology, the cooperative processing logic of anomaly detection and trajectory completion is combined, intermittent missing segments and sudden jump points can be accurately identified, the physical rationality of the repaired trajectory is ensured, and the vehicle trajectory repairing efficiency is improved. Therefore, the reliability and the operation precision of a downstream intelligent traffic system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of Beidou positioning technology, in particular to a vehicle Beidou positioning data anomaly repair method based on deep learning. BACKGROUND

[0002] As a global satellite navigation system independently constructed by China, the Beidou satellite navigation system (BDS) has been widely applied in key fields such as transportation, intelligent driving, and unmanned equipment navigation. The positioning data provided by the Beidou satellite navigation system is the core basis for the precise operation of downstream systems (such as intelligent traffic scheduling and unmanned vehicle path planning), and plays an irreplaceable role in ensuring the navigation reliability of various mobile carriers. However, in actual application scenarios, vehicles often face the influence of complex urban environments, such as signal obstruction caused by high-rise buildings, signal interruption during tunnel travel, and signal distortion caused by electromagnetic interference. These factors can cause frequent short-term loss, coordinate jumping, and position drift of Beidou positioning data, directly damaging the continuity and accuracy of the positioning data.

[0003] To solve the above-mentioned Beidou positioning data anomaly problem, traditional filtering methods such as Kalman filtering and particle filtering are often used for repair in the prior art. However, such methods can only perform simple smoothing based on the short-term statistical characteristics of the data, and are difficult to effectively capture the spatio-temporal correlation (such as the motion trend correlation of trajectory points at different time steps) and long-range dependence (such as the time sequence transmission law of speed and heading angle during continuous driving) hidden in vehicle trajectory data, resulting in limited anomaly repair effect - neither can accurately identify intermittent missing segments and sudden jump points in complex scenarios, nor can ensure the physical reasonableness of the repaired trajectory, ultimately causing serious impact on the reliability and operation accuracy of downstream intelligent transportation systems. Therefore, there is an urgent need for a new Beidou positioning data anomaly repair method that combines deep learning technology and takes into account the accuracy of anomaly detection and the reasonableness of trajectory completion. In view of this, we provide a vehicle Beidou positioning data anomaly repair method based on deep learning. SUMMARY

[0004] The purpose of the present application is to provide a vehicle Beidou positioning data anomaly repair method based on deep learning to solve the problem of poor traditional repair effect of Beidou positioning anomaly as proposed in the background.

[0005] To solve the above technical problems, the present application provides a vehicle Beidou positioning data anomaly repair method based on deep learning, comprising the following steps: S1, multi-source preprocessing: real-time collection of vehicle Beidou positioning data and multi-source auxiliary sensor data, extraction of driving speed, acceleration, heading angle and direction change characteristics in the Beidou positioning data through a sliding window, first-order difference analysis of driving speed, acceleration and heading angle, and completion of preprocessing; and fusion of the preprocessed Beidou positioning data and multi-source auxiliary sensor data into a multi-dimensional feature set; S2, abnormality mask generation: based on the Beidou positioning data in the multi-dimensional feature set fused in S1, combining statistical methods and DBSCAN clustering algorithm (statistical methods are used to analyze the time sequence correlation of data, and DBSCAN clustering is used to analyze the spatial distribution relationship of data), identifying abnormal points and missing segments in the Beidou positioning data, and generating an abnormality mask matrix; S3, trajectory completion: constructing a deep learning model based on an Encoder-Decoder structure, fusing an attention mechanism and introducing an autoencoder for assistance, using an end-to-end training framework, simulating training samples with random missing segments on existing real trajectory data through a multi-objective loss function including a mean square error loss term containing position error, a trajectory curvature continuity constraint loss term, and a physical rationality constraint loss term, supervising learning and training the deep learning model to learn the context association and physical constraints of trajectory completion, and then using the trained deep learning model in combination with the multi-dimensional feature set in S1 and the abnormality mask matrix in S2 to complete and repair the trajectory abnormal segment with context awareness, and output a continuous trajectory coordinate sequence; S4, fusion of real-time output and system deployment: deploying a lightweight deep learning model on a vehicle terminal or edge computing node, fusing multi-source auxiliary sensor data in the multi-dimensional feature set in S1, physically correcting the trajectory coordinate sequence completed in S3, real-time inputting abnormal trajectories and outputting repair results through a sliding window method, and finally obtaining a Beidou positioning trajectory that meets the vehicle dynamics characteristics.

[0006] As a further improvement of the technical solution, in S1, the vehicle Beidou positioning data at least includes longitude coordinates, latitude coordinates, driving speed, acceleration and heading angle; and the multi-source auxiliary sensor data at least includes three-axis acceleration data of an inertial measurement unit (IMU) carried by the vehicle, cumulative distance data of an odometer, and gyroscopic heading angular velocity data.

[0007] As a further improvement of the technical solution, in S2, identifying abnormal points and missing segments in the Beidou positioning data includes the following steps: S2.1, extracting Beidou positioning data (including speed, acceleration, direction change characteristics extracted by a sliding window and first-order difference analysis results) from the multi-dimensional feature set fused in S1; S2.2, input the extracted Beidou positioning data into the statistical method to analyze the data time sequence correlation, filter out the suspected abnormal data beyond the normal fluctuation range; then input the suspected abnormal data and the remaining normal data into the DBSCAN clustering algorithm, and further analyze the spatial distribution relationship of the data characteristics through the DBSCAN clustering algorithm; S2.3, according to the joint operation results of the statistical method and the DBSCAN clustering algorithm, the isolated data points not belonging to any cluster are determined as the abnormal points of the Beidou positioning data, and the trajectory segments not forming effective clusters continuously and meeting the set time continuity condition are determined as the missing segments.

[0008] As a further improvement of the technical solution, in the S2, the generation of the abnormal mask matrix comprises the following steps: S2.4, an abnormal mask matrix matching the length of the Beidou positioning data time sequence is constructed to ensure that the matrix dimension corresponds to the data time dimension one by one; S2.5, according to the rule that "the abnormal points and missing segments corresponding to the trajectory position are marked as 0, and the normal trajectory position is marked as 1", the elements in the abnormal mask matrix are marked; S2.6, after completing the marking of the abnormal mask matrix, the time index (start time step and end time step) of the missing segment is recorded, and the context feature (normal trajectory feature of the preset time step before and after the missing segment) of the missing segment is extracted to form a complete labeling result containing abnormal position, time range and context information.

[0009] As a further improvement of the technical solution, in the S3, a deep learning model based on Encoder-Decoder structure, fusion attention mechanism and introduction of autoencoder auxiliary is constructed, comprising the following steps: S3.1, a multi-layer bidirectional long short-term memory network Bi-LSTM or a multi-layer Transformer encoder layer is used to extract time sequence features of the input multi-dimensional feature set (containing the multi-source data of S1 and the context feature of S2), and capture the long-range context association of the trajectory; S3.2, a multi-head self-attention module is introduced between the encoder and the decoder, the weight coefficient between the trajectory points is calculated to strengthen the influence of the key trajectory segment on the completion result; S3.3, an autoencoder module is introduced to reduce and reconstruct the time sequence features output by the encoder, and filter out the core features matching the trajectory motion law; S3.4, a multi-layer LSTM or Transformer decoder layer is used to combine the abnormal mask matrix and the encoder output to generate the completed trajectory coordinate sequence step by step.

[0010] As a further improvement of the technical solution, in the S3, the composition and calculation of the multi-target loss function includes the following steps: S3.5, calculate the predicted trajectory coordinates and the Euclidean distance square mean of the true trajectory coordinates when there is no anomaly to obtain the mean square error loss term of the position error ; wherein represents the time step index of the trajectory coordinate sequence; S3.6, calculate the curvature change rate of adjacent trajectory points in the completed trajectory to obtain the trajectory curvature continuity constraint loss term ; S3.7, combine the IMU three-axis acceleration data and the cumulative travel distance data of the multi-source auxiliary sensor in the S1 multi-dimensional feature set to calculate the deviation of the predicted trajectory acceleration, speed and the corresponding sensor measured value, and introduce a weight coefficient (α is the weight coefficient of the IMU three-axis acceleration data, and β is the weight coefficient of the cumulative travel distance data of the odometer) to balance the influence of the two types of sensor data, and obtain the physical rationality constraint loss term .

[0011] As a further improvement of the technical solution, in the S3, the simulation generates a training sample of a random missing segment, which includes the following steps: S3.8, randomly mark the missing segment using two modes on the existing true trajectory data: when using the time-continuous missing mode, the single missing length is set to 5-20 time steps; when using the random discrete missing mode, the missing point ratio in a single trajectory is controlled at 5%-20%; S3.9, the context trajectory of the marked missing segment is used as the input feature, and the context trajectory is the normal trajectory data (including the positioning data and multi-source auxiliary sensor data in the multi-dimensional feature set described in S1) of 20-50 time steps before the start time step of the missing segment and 20-50 time steps after the end time step; the true trajectory coordinates of the missing segment are used as the label to form a training sample pair corresponding to the input feature and the label.

[0012] As a further improvement of the technical solution, in the S3, the context-aware completion repair of the trajectory abnormal segment using the trained deep learning model includes the following steps: S3.10, locate the start time step and the end time step of the abnormal segment using the abnormal mask matrix described in S2, extract the normal trajectory features (including the speed, heading angle and multi-source auxiliary sensor data in the multi-dimensional feature set described in S1) of 30-100 time steps before the start time step of the abnormal segment and 30-100 time steps after the end time step as context input; S3.11, input the context into the trained deep learning model with the abnormal mask matrix input in S2; S3.12, the deep learning model assigns dynamic weights to the context features through an attention mechanism, preferentially strengthens the influence of trajectory points consistent with the motion trend of the abnormal segment, and generates trajectory coordinates of the abnormal segment step by step; S3.13, output a continuous trajectory coordinate sequence, which maintains spatial and temporal continuity with the normal trajectory before and after the abnormal segment, and meets the physical constraint requirements of the multi-target loss function.

[0013] As a further improvement of the technical solution, in the S4, the deployment of the lightweight deep learning model and the real-time data input include the following steps: S4.1, the trained deep learning model is processed to be lightweight, and the processing mode includes model parameter pruning and quantization compression, so as to adapt to the hardware operation requirements of the vehicle terminal or edge computing node; S4.2, a sliding window is configured for real-time data acquisition, and the time step range covered by the window needs to include the latest generated abnormal trajectory data, so as to ensure that the time sequence correlation information of the trajectory can be captured synchronously; S4.3, real-time input data through the sliding window: after completing the window data acquisition once, the abnormal trajectory data (including the abnormal position information marked by the abnormal mask matrix in S2) in the window and the multi-source auxiliary sensor data in the multi-dimensional feature set in S1 are input synchronously, so as to provide complete input information for the deep learning model inference.

[0014] As a further improvement of the technical solution, in the S4, the physical rationality correction and the repair result output include the following steps: S4.4, fuse the IMU three-axis acceleration data, the odometer cumulative running distance data, and the gyroscopic heading angular velocity data in the multi-dimensional feature set in S1, and construct constraint logic based on vehicle dynamics constraints (such as the time sequence correlation of acceleration and speed, and the matching of heading angle and running direction), for checking the rationality of the trajectory motion parameters; S4.5, based on the constraint logic of S4.4, the continuous trajectory coordinate sequence output by S3 is checked and corrected step by step: if the motion parameters (acceleration, speed, heading angle) of a time step trajectory point do not match the actual motion state of the vehicle reflected by the multi-source auxiliary sensor data, then adjust the coordinates of the trajectory point combined with the sensor data, so that the trajectory conforms to the vehicle motion law; S4.6, output the corrected Beidou positioning trajectory, which needs to include the longitude coordinate, latitude coordinate, running speed and heading angle of each time step, and keep consistent with the core parameter dimensions of the vehicle Beidou positioning data in S1.

[0015] Compared with the prior art, the present application has the following advantages: 1. The present application adopts a deep learning-based technical solution, utilizes the processing capability of a deep learning model for vehicle trajectory data, and can effectively improve the continuity and accuracy of positioning data in traffic transportation, intelligent driving, unmanned equipment navigation, and other application scenarios, in the face of short-term loss, coordinate jumping, and position drift of Beidou positioning data caused by factors such as high building obstruction, tunnel passing signal interruption, and electromagnetic interference in complex urban environments. 2. The present application captures the spatio-temporal correlation and long-range dependency in vehicle trajectory data through deep learning technology, combines the cooperative processing logic of anomaly detection and trajectory completion, and can accurately identify intermittent missing sections and sudden jump points in the face of the problem that traditional filtering methods cannot capture the above data characteristics, resulting in inaccurate identification of intermittent missing sections and sudden jump points in complex scenarios, and the repaired trajectory lacks physical rationality, thereby affecting the reliability and operation accuracy of downstream intelligent traffic scheduling, unmanned vehicle path planning, and other systems. In the above application scenarios, the present application can accurately identify intermittent missing sections and sudden jump points, ensure the physical rationality of the repaired trajectory, and thus improve the reliability and operation accuracy of downstream intelligent traffic systems. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The present application is a method step schematic diagram. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0018] As shown in Figure 1 The present embodiment provides a deep learning-based vehicle Beidou positioning data anomaly repair method, which includes: S1, multi-source preprocessing: real-time collection of vehicle Beidou positioning data and multi-source auxiliary sensor data, extraction of driving speed, acceleration, heading angle, and direction change features in the Beidou positioning data through a sliding window, first-order difference analysis of driving speed, acceleration, and heading angle, and completion of preprocessing; and fusion of the preprocessed Beidou positioning data and multi-source auxiliary sensor data into a multi-dimensional feature set; In this step, in the S1, the vehicle Beidou positioning data at least includes longitude coordinates, latitude coordinates, driving speed, acceleration, and heading angle; the multi-source auxiliary sensor data at least includes three-axis acceleration data of an inertial measurement unit (IMU) carried by the vehicle, cumulative driving distance data of an odometer, and gyroscopic heading angular velocity data.

[0019] Specifically, the vehicle end needs to synchronously deploy the Beidou positioning receiving module and the multi-source auxiliary sensor module, wherein: The Beidou positioning receiving module selects a terminal device supporting the BDS B1I / B2I frequency band to ensure real-time output of longitude coordinates, latitude coordinates, driving speed, acceleration and heading angle data; The multi-source auxiliary sensor module needs to integrate an inertial measurement unit (IMU), an odometer and a gyroscope, and each device needs to realize data interaction through the vehicle CAN bus or Ethernet to ensure the collaboration of the collection process.

[0020] At the same time, combined with the change rule of the trajectory characteristics in the vehicle driving process, the collection frequency of the Beidou positioning data and the multi-source auxiliary sensor data is uniformly set to 1-10Hz (which can be adjusted according to the actual application scene, such as 5-10Hz for urban roads to capture frequent turning and starting and stopping features, and 1-5Hz for highways to balance the data volume and timeliness). At the same time, in order to avoid the time deviation of multi-source data due to collection delay, each device needs to be time-synchronized based on the GPS timestamp or the vehicle unified clock signal to ensure that the Beidou data and the auxiliary sensor data at the same time step correspond to each other; In addition, the collected raw data needs to be temporarily stored in the local cache unit (such as an SD card or a solid-state storage) of the vehicle terminal, and the temporary storage format adopts a structured data format (such as JSON or CSV), and each data record needs to include "collection timestamp + Beidou positioning parameters (longitude, latitude, driving speed, acceleration, heading angle) + auxiliary sensor parameters (IMU three-axis acceleration, odometer cumulative driving distance, gyroscope heading angular velocity)", which facilitates fast calling when extracting features in the subsequent sliding window.

[0021] Further, the driving speed, acceleration, heading angle and direction change characteristics in the Beidou positioning data are extracted through the sliding window, and the first-order difference analysis is performed on the driving speed, acceleration and heading angle to complete the preprocessing, including the following steps: S1.1, based on the Beidou positioning data collection frequency (5-10Hz for urban roads and 1-5Hz for highways), set the sliding window length to 5-20 time steps and the sliding step to 1 time step; traverse the real-time Beidou positioning data stream (including longitude, latitude, driving speed, acceleration and heading angle) in time steps, and cut off continuous data blocks according to the window length to form a plurality of sliding window data blocks containing complete positioning parameters; S1.2, for each window data block, extract the features according to the time step: directly extract the driving speed, acceleration and heading angle values to form three types of basic feature sequences; calculate the value difference of adjacent time steps (the difference of the first time step is set to 0 or consistent with the difference of the second time step), and obtain the direction change characteristic sequence, a total of four types of feature sequences; S1.3, take the driving speed, acceleration, and heading angle sequence of S1.2 as the calculation object, respectively calculate the numerical difference of adjacent time steps (the difference of the first time step is set to 0 or consistent with the second time step), and obtain the corresponding three types of first-order difference sequence, which respectively reflects the change rate of driving speed, the fluctuation of acceleration, and the change trend of heading angle; in this step, the "absolute value characteristics" of driving speed, acceleration, and heading angle are converted into "change rate characteristics" through first-order difference, which can more sensitively capture the mutation trend of vehicle motion state (such as sudden acceleration, sudden braking, rapid turning, etc.), and these trend information can be complementary to multi-source auxiliary sensor data (such as IMU acceleration change, gyroscope angular velocity change), providing more rich time sequence change basis for subsequent anomaly detection; S1.4, align the four types of extracted features (driving speed, acceleration, heading angle, and direction change) and the three types of difference features (speed difference, acceleration difference, and heading angle difference) of each sliding window by time step to form a preprocessed feature subset, which provides a basis for subsequent fusion with multi-source auxiliary sensor data to build a multi-dimensional feature set.

[0022] In addition, when the preprocessed Beidou positioning data and multi-source auxiliary sensor data are fused into a multi-dimensional feature set, the preprocessed Beidou positioning data (including the speed, acceleration, heading angle, and direction change features extracted by the sliding window and the first-order difference results of the three) and the multi-source auxiliary sensor data (IMU three-axis acceleration, odometer cumulative driving distance, and gyroscope heading angular velocity) are first subjected to consistency check, and data beyond the physical motion range of the vehicle is removed; then, the two types of data are aligned one by one to build a multi-dimensional feature matrix (each row corresponds to 1 time step, and each column corresponds to 1 feature) with time steps as the index; finally, the matrix is stored in binary format, with feature names, time step ranges, and other meta information, providing complete input for subsequent links.

[0023] S2, anomaly mask generation: based on the Beidou positioning data in the multi-dimensional feature set fused in S1, combined with statistical methods and DBSCAN clustering algorithm (statistical methods are used to analyze the time sequence correlation of data, and DBSCAN clustering is used to analyze the spatial distribution relationship of data), identify the abnormal points and missing segments in the Beidou positioning data, and generate an anomaly mask matrix; In this step, in S2, identifying abnormal points and missing segments of Beidou positioning data includes the following steps: S2.1, extract the Beidou positioning data (including the speed, acceleration, and direction change features extracted by the sliding window and the first-order difference analysis results) from the multi-dimensional feature set fused in S1; Specifically, from the S1 fused multi-dimensional feature set, the Beidou positioning data is extracted step by step according to the time step. The extracted content needs to include the driving speed, acceleration, direction change feature extracted by the sliding window, and the first-order difference analysis result of the three, and the extracted data must maintain the time sequence order consistent with the original multi-dimensional feature set to prevent time sequence disorder from affecting subsequent anomaly analysis.

[0024] S2.2, the extracted Beidou positioning data is first input into the statistical method to analyze the time sequence correlation of the data, and the suspected abnormal data exceeding the normal fluctuation range is screened out; then the suspected abnormal data and the remaining normal data are jointly input into the DBSCAN clustering algorithm, and the spatial distribution relationship of the data features is further analyzed by the DBSCAN clustering algorithm. Specifically, the statistical method is first used to analyze the time sequence correlation of the data, and the mean and standard deviation of the features in the sliding window are calculated to define the normal fluctuation range. The data exceeding this range is marked as suspected abnormal data. Then the suspected abnormal data and the remaining normal data are jointly input into the DBSCAN clustering algorithm. The neighborhood radius and minimum sample size of the algorithm need to be adjusted flexibly according to the data space density of the actual trajectory to adapt to the data space distribution relationship in different scenarios, so as to realize further filtering of the data.

[0025] Further, for the statistical method, the sliding window length is consistent with the sliding window length in S1, which is 5-20 time steps. For driving speed, acceleration, direction change feature and their first-order difference results, the mean and standard deviation of the features in each sliding window are calculated, and the normal fluctuation range of each feature is defined by the mean plus or minus three times the standard deviation. The corresponding data exceeding this range is marked as suspected abnormal data. For the DBSCAN clustering algorithm, the neighborhood radius and minimum sample size need to be dynamically adjusted according to the spatial density of the trajectory data: if it is a city road scene, the spatial density of the trajectory points is high, the neighborhood radius can be set to 1.5-2 times the distance between adjacent trajectory points when the vehicle is normally driving, and the minimum sample size is set to 5-8; if it is a highway scene, the spatial density of the trajectory points is low, the neighborhood radius is set to 2-3 times the distance between adjacent trajectory points when the vehicle is normally driving, and the minimum sample size is set to 3-5. After clustering, the isolated data points not belonging to any cluster are the final determined abnormal points; if the suspected abnormal data is classified into an effective cluster, it is determined as normal data, so as to realize the joint verification of the two methods and improve the accuracy of abnormal identification.

[0026] S2.3, according to the joint operation results of the statistical method and the DBSCAN clustering algorithm, the isolated data points not belonging to any cluster are determined as abnormal points of the Beidou positioning data, and the trajectory segment not forming an effective cluster continuously and meeting the set time continuity condition is determined as a missing segment.

[0027] Specifically, according to the joint operation result of the statistical method and the DBSCAN clustering algorithm, the isolated data points not belonging to any cluster are directly determined as the abnormal points of the Beidou positioning data; for the trajectory segment that does not form an effective cluster continuously, if it meets the preset time continuity condition (conventionally set as 3 or more time steps continuously), it is determined as a missing segment of the Beidou positioning data.

[0028] In this step, in the S2, generating the abnormal mask matrix comprises the following steps: S2.4, an abnormal mask matrix matching the length of the time series of the Beidou positioning data is constructed, the matrix type is set as a one-dimensional matrix, the number of matrix elements needs to correspond to the number of time steps of the Beidou positioning data, and it is ensured that each element in the matrix can accurately map to the corresponding time step position of the data.

[0029] S2.5, according to the rule that the positions of the abnormal points and the missing segments are marked as 0 and the normal trajectory positions are marked as 1, the elements in the abnormal mask matrix are marked; the marking process needs to cover all elements in the matrix to avoid omission or mislabeling.

[0030] S2.6, after completing the marking of the abnormal mask matrix, the time index of the missing segment is recorded (used to clearly indicate the start time step and the end time step of the missing segment), and the context features of the missing segment are extracted, specifically the normal trajectory features of the 20-50 time steps before the start time step and the 20-50 time steps after the end time step of the missing segment, to form a complete labeling result containing the abnormal position, the time range and the context information.

[0031] S3, trajectory completion: a deep learning model based on an Encoder-Decoder structure, fusing an attention mechanism and introducing an autoencoder auxiliary is constructed, an end-to-end training framework is adopted, a multi-objective loss function including a mean square error loss term containing position error, a trajectory curvature continuity constraint loss term and a physical rationality constraint loss term is used to simulate training samples of random missing segments on existing real trajectory data, the deep learning model is supervised and learned to train to make it learn the context association and physical constraints of trajectory completion, and then the trained deep learning model is used to combine the multi-dimensional feature set of S1 and the abnormal mask matrix of S2 to complete and repair the context-aware trajectory abnormal segment, and output a continuous trajectory coordinate sequence. This step S3 realizes the context-aware completion of the trajectory abnormal segment by constructing a deep learning model with a specific structure, combining a multi-objective loss function and a simulated training sample, and the core is to ensure that the model can learn the spatiotemporal association and physical constraints of the trajectory, and finally output a continuous and reasonable trajectory coordinate sequence.

[0032] In this step, in the S3, a deep learning model based on Encoder-Decoder structure, fusion attention mechanism and introduction of auto-encoder auxiliary is constructed, including the following steps: S3.1, a multi-layer bidirectional long short-term memory network Bi-LSTM or a multi-layer Transformer encoder layer is used to extract time sequence features of the input multi-dimensional feature set (including multi-source data of S1 and context features of S2), and long-range context association of the trajectory is captured; Specifically, if Bi-LSTM is selected, the input is the multi-dimensional feature set generated by S1 , wherein: represents the model training batch size, which is 32-64 (matching the subsequent training sample storage batch); represents the input time sequence length, which is consistent with the number of context feature time steps extracted by S2 (20-50 time steps); is the multi-dimensional feature dimension (including Beidou positioning data, IMU three-axis acceleration data, odometer cumulative travel distance data, etc., typical value 10-15); Bi-LSTM fuses features through forward LSTM (captures future time sequence association) and backward LSTM (captures historical time sequence association) to output time sequence feature tensor , represents the number of hidden units of each layer of Bi-LSTM, which is 64-128 (adjusted according to , when , 64 is taken, , 96 is taken), and the forget gate bias is initialized to to enhance long-time sequence trajectory feature memory; If Transformer encoder is selected, sinusoidal position encoding needs to be supplemented, and the formulas are as follows: ; ; , wherein, is the time step position ( ), is the dimension index ( ), and is 64-128 (consistent with of Bi-LSTM); After the input features are processed by , they are input into 2-3 layers of Transformer encoder, each layer is connected by residual connection ( ) and layer normalization to alleviate gradient disappearance, and finally the time sequence feature tensor is output.

[0033] S3.2. A multi-head self-attention module is introduced between the encoder and decoder to enhance the influence of key trajectory segments on the completion result by calculating the weight coefficients between trajectory points. Specifically, a multi-head self-attention module needs to be introduced between the encoder and decoder to enhance the impact of key trajectory segments on the completion results, as detailed below: The encoder output features are received via a multi-head self-attention module. First, map it to a query. ,key ,value ,Right now: ; in, The number of attention heads (values ​​4-8). It is a single-head attention feature dimension, and ; Then, the attention score is calculated using the following formula: ; in, This represents the attention score, which measures... and Similarity; Indicates matrix transpose; and through The weighting coefficients are obtained; the single-head output is... ;in For attention head index, and ; Represents the normalization function; Indicates the first The output of each attention head; Finally, the outputs of all the heads are concatenated and passed through a linear layer, as shown in the formula: ; in, This indicates the output used to focus all attention heads. Concatenate along the feature dimension; This represents a linear transformation layer, used to map the concatenated multi-head features back to the original feature dimension; Optimization during training via backpropagation The mapping parameters tilt the weights toward trajectory points that align with the movement trend of the abnormal segments.

[0034] S3.3. Introduce an autoencoder module to perform dimensionality reduction and reconstruction on the temporal features output by the encoder, and select the core features that match the trajectory motion law. Specifically, to filter the core features of trajectory motion, an autoencoder module needs to be introduced to process the encoder output. Perform dimensionality reduction and reconstruction. Specifically: The encoding process is implemented using the following formula: ; in, The coding layer weight matrix (dimension) ), For coding layer bias (dimension) ), Activation functions are used to introduce non-linear filtering capabilities; This represents the core feature tensor obtained after encoding by the autoencoder, and is a representation of the encoder output. Key features that match the trajectory motion patterns are retained after dimensionality reduction; The decoding process is implemented using the following formula: ; in, The decoding layer weight matrix (dimensions) ), For the decoding layer bias (dimension) ); This represents the feature tensor reconstructed after decoding by the autoencoder, used in conjunction with the original encoder output. Calculate the reconstruction error and constrain the effectiveness of the core features; During training, the reconstruction error of the autoencoder ( The mean squared error function is used as an auxiliary loss term (weights 0.1-0.3) and jointly trained with the main network to ensure... Only the core features that match the trajectory motion pattern are retained.

[0035] S3.4. Employ a multi-layer LSTM or Transformer decoder layer, combined with the anomaly mask matrix and encoder output, to generate the completed trajectory coordinate sequence step by step over time.

[0036] Specifically, if an LSTM decoder is used, the number of layers is the same as the encoder (2-3 layers), and the number of hidden units is the same as the encoder. First, the output of the autoencoder... The anomaly mask matrix generated by S2 ( Outliers / missing segments are marked as 0, and normal segments are marked as 1) and mapped to dimension-consistent segments through a linear layer. merging into Subsequently, an "autoregressive generation" method was adopted, through... Calculate the hidden state; then... Generate complete coordinates ;in To predict longitude, To predict the latitude; denotes the output of the autoencoder and the fused result of the masked tensor after mapping; denotes the hidden state of the LSTM decoder at the time step, records the historical time step information, is the hidden state of the previous time step; If the Transformer decoder is selected, the number of layers is consistent with the encoder, and each layer includes a masked multi-head self-attention sublayer (to prevent future time step information leakage), an encoder-decoder attention sublayer (to focus on the core features of the ), and a feedforward neural network sublayer. First, the and the time series masked tensor mapped by the embedding layer are concatenated into , which is input into the decoder for processing. After processing, the is generated through a linear layer (output dimension = 2). .

[0037] In this step, in the S3, the composition and calculation of the multi-objective loss function include the following steps: To ensure the position accuracy, smoothness, and physical reasonableness of the completed trajectory, a multi-objective loss function needs to be constructed.

[0038] S3.5, calculate the predicted trajectory coordinates and the Euclidean distance squared mean of the true trajectory coordinates when there is no anomaly, to obtain the mean squared error loss term of the position error ; wherein denotes the time step index of the trajectory coordinate sequence; Specifically, the mean squared error loss term of the position error is used to constrain the deviation between the predicted coordinates and the true coordinates, and the formula is: ; wherein is the abnormal segment time step index (located by the S2 mask matrix); is the total number of abnormal segment time steps, and , is the starting time step of the abnormal segment, is the ending time step; is the predicted coordinate, is the true coordinate when there is no anomaly, and this loss is only calculated for the abnormal segment to avoid interference from the normal segment.

[0039] S3.6, calculate the curvature change rate of adjacent trajectory points in the completed trajectory, to obtain the trajectory curvature continuity constraint loss term ; Specifically, the trajectory curvature continuity constraint loss term To ensure the smoothness of the trajectory, the curvature of the adjacent 3 points of the completed trajectory is calculated as follows : ; Wherein, is the cross product of a two-dimensional vector, is the vector length; The curvature rate is calculated by the following formula: ; The final loss formula is .

[0040] S3.7, combined with S1 multi-dimensional feature set, the IMU three-axis acceleration data of the multi-source auxiliary sensor and the cumulative travel distance data of the odometer, the acceleration, speed and deviation of the corresponding sensor measured value of the predicted trajectory are calculated, and a weight coefficient is introduced (α is the weight coefficient of the IMU three-axis acceleration data, and β is the weight coefficient of the cumulative travel distance data of the odometer) to balance the influence of the two types of sensor data, and a physical rationality constraint loss term is obtained .

[0041] Specifically, the physical rationality constraint loss term is used to ensure that the trajectory conforms to the vehicle dynamics characteristics, and the predicted speed is calculated , and the predicted acceleration is calculated ; wherein, is the data acquisition time interval, which is consistent with the acquisition frequency of S1; is the cumulative travel distance of the odometer in S1; , is the IMU three-axis acceleration data in S1; A weight coefficient (α is the weight coefficient of the IMU three-axis acceleration data, and the value is 0.3-0.7) and (β is the weight coefficient of the cumulative travel distance data of the odometer, and the value is 0.3-0.7, and ; urban road , highway ); ); The loss formula is: ; The multi-objective total loss is: ; Wherein, represents the physical plausibility constraint loss term, which measures the deviation of the completed trajectory's speed and acceleration from the sensor's measured values; represents the total loss of multiple targets, which is the weighted sum of the position error, trajectory curvature continuity, and physical plausibility; is a positive number and the sum is 1 by default (the default value is determined based on the technical logic of "prioritizing position accuracy while considering trajectory curvature continuity and physical plausibility").

[0042] In this step, in S3, simulating the generation of training samples with random missing segments includes the following steps: S3.8, randomly mark missing segments on existing real trajectory data in two modes: when using the time-continuous missing mode, the single segment missing length is set to 5-20 time steps; when using the random discrete missing mode, the missing point ratio in a single trajectory is controlled at 5%-20%; S3.9, the context trajectory of the marked missing segment is used as the input feature, which is the normal trajectory data (including the positioning data and multi-source auxiliary sensor data in the multi-dimensional feature set described in S1) of 20-50 time steps before the start time step of the missing segment and 20-50 time steps after the end time step; the real trajectory coordinates of the missing segment are used as the label to form the training sample pair corresponding to the input feature and the label.

[0043] Specifically, model training needs to simulate the generation of training samples with random missing segments, first mark missing segments on real trajectory data (covering urban, highway, and suburban roads, each trajectory time step ≥300, total data volume ≥3000, preprocessed into a multi-dimensional feature set by S1) in two modes: In the time-continuous missing mode, the single segment missing length is 5-20 time steps, and 1-3 non-overlapping intervals are randomly selected for each trajectory; In the random discrete missing mode, the missing ratio is 5%-20%, and randomly select discontinuous time steps for marking, and the marking is only for trajectory coordinates, and the multi-dimensional feature set retains normal values; where is the missing ratio, is the total time step of the trajectory, is the rounding function; Then build a training sample pair, the input feature is the context trajectory of the missing segment ( , the time range is 20-50 time steps before the start of the missing segment and 20-50 time steps after the end, and the boundary is adjusted if insufficient), and the label is the real trajectory coordinates of the missing segment ​​​The total number of missing time steps is stored as a Tensor file (32-64 samples per file) in the form of "sample ID-input feature tensor-label tensor", and is divided into training set, validation set and test set in the ratio of 7:2:1.

[0044] In this step, in the S3, the context-aware completion repair of the trajectory abnormal segment using the trained deep learning model includes the following steps: S3.10, the start time step and the end time step of the trajectory abnormal segment are located in combination with the abnormal mask matrix S2, and the normal trajectory features (including the speed, heading angle and multi-source auxiliary sensor data in the multi-dimensional feature set S1) of 30-100 time steps before the start time step and 30-100 time steps after the end time step of the abnormal segment are extracted as context input; It should be noted that the deep learning model is trained by supervised learning to learn the context association and physical constraints of trajectory completion, which is performed according to the following procedure: First, construct the training sample set: based on the multi-scenario real trajectory data collected by the real vehicle (covering urban roads, highways, suburban roads, etc.), generate sample pairs according to the missing segment marking method (time-continuous missing, random discrete missing) -input context trajectory features before and after the abnormal segment (including the multi-dimensional feature set generated in S1, such as Beidou positioning coordinates, driving speed, IMU three-axis acceleration, odometer distance, etc., and the time step length is 30-100), and the label is the real trajectory coordinates and physical parameters (including longitude, latitude, and corresponding real speed, acceleration, etc.) of the missing segment; Second, loss function joint optimization: during training, the multi-objective loss function jointly constrains model learning: The attention mechanism promotes the model to capture the temporal association of the context trajectory (such as the speed change trend of adjacent trajectory points and the continuity of the heading angle, to ensure that the completed trajectory seamlessly connects with the normal trajectory before and after it in space and time); Force the model to learn the physical constraints (make the speed and acceleration of the completed trajectory consistent with the measured data of IMU, odometer, etc., and comply with the vehicle dynamics law, such as the acceleration and IMU measured value deviation within a reasonable range during acceleration); Finally, model iterative training: batch training is adopted ( ), and the network parameters (such as the forgetting gate weight of LSTM and the attention weight of Transformer) are iteratively optimized by the back propagation algorithm, and when the multi-objective loss on the validation set converges and the completed trajectory meets the "spatiotemporal continuity (the distance between adjacent coordinates meets the vehicle speed constraint) + physical rationality (the deviation of speed and acceleration from sensor data meets the standard)" indicators, the training is completed; The trained model can output a complete trajectory that satisfies both contextual association and physical constraints based on the contextual features of the anomalous segments.

[0045] S3.11 Input the context input and the anomaly mask matrix described in S2 into the trained deep learning model; S3.12. The deep learning model assigns dynamic weights to contextual features through an attention mechanism, prioritizing the enhancement of trajectory points that are consistent with the movement trend of the abnormal segment, and generating the trajectory coordinates of the abnormal segment step by step. S3.13. Output a continuous trajectory coordinate sequence, wherein the trajectory coordinate sequence maintains spatiotemporal continuity with the normal trajectory before and after the abnormal segment, and satisfies the physical constraints of the multi-objective loss function.

[0046] Specifically, when performing context-aware completion and repair on abnormal trajectory segments, the starting time step of the abnormal segment is first located by combining the S2 anomaly mask matrix. With end time step abnormal segment length and according to Extracting contextual input: At that time, the context length is 30 time steps (range) ); At that time, the context length is 100 time steps (range) The context input includes the velocity, heading angle, and multi-source sensor data in S1, and is consistent with the normalization method of the training samples; Next, organize the context input as follows: ( For the total number of context time steps, Extracting the corresponding part of the S2 mask matrix (Context marker 1, abnormal segment marker 0), synchronously input deep learning model.

[0047] Furthermore, deep learning models calculate dynamic weights through an attention mechanism, the expression of which is: ; in, Indicates the abnormal segment number Time step and context The dynamic weights of time-step features are used to measure the influence of contextual features on the completion of anomalous segments; Represents an exponential function; Cosine similarity; Candidate coordinates for the abnormal segment; For contextual features; Then combine the weight and decoder mapping function Generate coordinates: ; wherein, represents the first time step generated by the completion trajectory coordinates, which is a component of the final output continuous trajectory; represents the label (value 0) of the first time step in the abnormal mask matrix.

[0048] Finally, the abnormal segment coordinates are spliced with the normal trajectories before and after to form a complete sequence , and the adjacent coordinate distance is checked to be no more than ( is the maximum design speed of the vehicle, which is read from the vehicle factory parameters or the vehicle-mounted system) and the time step sequence is consistent, and the output format is consistent with the S1 Beidou positioning data (longitude and latitude in degrees, with 6 decimal places).

[0049] S4, fusion of real-time output and system deployment: deploy a lightweight deep learning model on a vehicle terminal or edge computing node, fuse multi-source auxiliary sensor data in the multi-dimensional feature set of S1, and physically correct the trajectory coordinate sequence completed by S3. Through a sliding window, input the abnormal trajectory in real time and output the repair result, and finally obtain the Beidou positioning trajectory consistent with the vehicle dynamics characteristics.

[0050] In this step, in the S4, the deployment and real-time data input of the lightweight deep learning model include the following steps: S4.1, the trained deep learning model is processed to be lightweight, and the processing method includes model parameter pruning and quantization compression to adapt to the hardware operation requirements of the vehicle terminal or edge computing node; Specifically, the trained deep learning model (Encoder-Decoder structure, including attention mechanism and autoencoder auxiliary module) of S3 is processed to be lightweight to adapt to the hardware resource limitations of the vehicle terminal or edge computing node, which specifically includes: Model parameter pruning: adopt a structured pruning strategy to remove redundant parameters and neurons in the network. For the encoder and decoder hidden layers of Bi-LSTM / Transformer, determine the pruning ratio according to the hardware computing power adaptation requirements, remove neurons with low weight contribution; for the multi-head self-attention module, keep the core attention head (the number is adapted according to the model feature dimension), and remove redundant heads with low attention weight mean. After pruning, the performance of the deep learning model is restored through fine-tuning training to ensure that the trajectory completion effect is not affected; Model quantization compression: Linear quantization is used to compress the precision of deep learning model parameters. High-precision parameters in the training stage are converted to low-precision format suitable for hardware inference, reducing the storage volume and computational complexity of deep learning models. During quantization, the scaling factor and offset are determined based on the statistical feature distribution of the training set data to ensure that the inference accuracy of the quantized deep learning model meets the trajectory repair requirements. Then, the quantized deep learning model is converted to an inference format suitable for vehicle-mounted or edge hardware to improve inference efficiency. Deployment adaptation verification: Inference testing is performed on the target hardware platform (vehicle terminal or edge computing node) to verify the compatibility and real-time performance of the deep learning model. The test focuses on the deep learning model loading speed, single inference time consumption, and memory occupation to ensure that the deep learning model can run stably and meet the real-time processing requirements in the vehicle scene, while verifying the physical reasonableness of the completed trajectory and maintaining consistency with the lightweight version before.

[0051] S4.2, configure a sliding window for real-time data collection, the time step range covered by the window needs to include the latest generated abnormal trajectory data to ensure that the time sequence correlation information of the trajectory can be captured synchronously; Specifically, a sliding window is configured for real-time collection and caching of trajectory data to ensure that the model can capture the time sequence correlation information of the trajectory. The parameter configuration is as follows: Window time step range: The total number of time steps covered by the window is consistent with the time step range of the context features in S3, and needs to include the latest generated abnormal trajectory data and the associated normal trajectory data before and after, to ensure that the model can complete the trajectory based on complete context; Window sliding step: The sliding step is adapted to the collection frequency of the Beidou positioning data in S1. After each sliding, the window automatically updates the data, discards the expired data and includes the latest collected trajectory data, achieving real-time dynamic updating of the data; Window caching mechanism: A dedicated cache area is opened in the hardware platform to store the data in the window in time step order, ensuring the time sequence continuity and integrity of the data, and avoiding data loss or disorder.

[0052] S4.3, real-time input data through sliding window: after each window data collection is completed, the abnormal trajectory data (including abnormal position information marked by the abnormal mask matrix in S2) and the multi-source auxiliary sensor data in the multi-dimensional feature set in S1 are input synchronously to provide complete input information for the deep learning model inference.

[0053] Specifically, after each sliding window data collection is completed, the data is input to the lightweight model according to the following process: Data real-time acquisition and synchronization: Collect multi-source auxiliary sensor data (IMU three-axis acceleration, odometer cumulative travel distance, gyro heading angular velocity) and Beidou positioning data in the multi-dimensional feature set of S1 through the vehicle bus, and realize the time sequence alignment of data from different sources based on the timestamp to ensure the consistency of the input data. At the same time, read the abnormal mask matrix generated by S2 to mark the abnormal position of the data in the window; Input data format processing: The abnormal trajectory data, abnormal mask matrix and multi-source auxiliary sensor data in the window are arranged into the input format consistent with the training phase of the deep learning model. The feature data needs to be preprocessed in the same way as during training (such as normalization processing) to ensure that the input feature distribution is consistent with the training, and to ensure the inference accuracy of the deep learning model. Model inference triggering: After the data format verification is passed, the inference process of the lightweight deep learning model is triggered, and the formatted input data is transmitted into the deep learning model to provide complete context information and abnormal position identification for trajectory completion and correction.

[0054] In this step, in S4, the physical reasonableness correction and repair result output includes the following steps: S4.4, fuse the IMU three-axis acceleration data, odometer cumulative travel distance data, and gyro heading angular velocity data in the multi-dimensional feature set of S1, and construct constraint logic based on vehicle dynamics constraints (such as the time sequence correlation of acceleration and speed, and the matching of heading angle and travel direction) to verify the reasonableness of the trajectory motion parameters; Specifically, the specific constraint content of the constraint logic is as follows: Acceleration constraint: Based on the synthesized value of the three-axis acceleration measured by the IMU, the acceleration calculated by the completed trajectory is constrained to be consistent with the actual motion acceleration of the vehicle, ensuring that the trajectory conforms to the performance limit of the vehicle; Speed constraint: Based on the measured speed of the odometer, the speed calculated by the completed trajectory is constrained to match the actual travel speed of the vehicle, while ensuring that the speed does not exceed the reasonable travel range of the vehicle design; Heading angle constraint: Based on the heading angular velocity measured by the gyroscope, the heading angle change rate of the completed trajectory is constrained to be consistent with the actual steering motion of the vehicle, conforming to the physical characteristics of the vehicle steering system.

[0055] S4.5, based on the constraint logic of S4.4, the continuous trajectory coordinate sequence output by S3 is checked and corrected at each time step: if the motion parameters (acceleration, speed, heading angle) of a certain time step trajectory point do not match the actual motion state reflected by the multi-source auxiliary sensor data, adjust the coordinates of the trajectory point in combination with the sensor data to make the trajectory conform to the vehicle motion law; Specifically, based on the above constraint logic, the continuous trajectory coordinate sequence output by S3 is checked and corrected at each time step: Motion parameter calculation: For each time step of the trajectory sequence, the corresponding motion parameters (speed, acceleration, heading angle) are calculated, and the calculation method is consistent with the parameter calculation logic of the trajectory completion process in S3; Constraint compliance verification: Compare the calculated motion parameters with the multi-source sensor measured data and vehicle dynamics constraints to determine whether each time step of the trajectory point meets the physical rationality requirements; Trajectory coordinate correction: For trajectory points that do not meet the physical rationality constraints, adjust the coordinates in combination with sensor measured data. The correction process is based on the actual motion state of the vehicle reflected by the sensor data, and adjusts the longitude and latitude coordinates of the trajectory points, so that the corrected trajectory motion parameters meet all the constraints and comply with the actual driving rules of the vehicle.

[0056] S4.6, output the corrected Beidou positioning trajectory, and the trajectory data needs to include the longitude coordinate, latitude coordinate, driving speed and heading angle of each time step, which is consistent with the core parameter dimension of the vehicle Beidou positioning data in S1.

[0057] Those of ordinary skill in the art can understand that the processes for implementing all or part of the steps of the above embodiments can be completed by hardware, or by programs instructing relevant hardware to complete.

[0058] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for repairing anomalies in vehicle BeiDou positioning data based on deep learning, characterized in that, Includes the following steps: S1. Multi-source preprocessing: Real-time acquisition of vehicle BeiDou positioning data and multi-source auxiliary sensor data; extraction of driving speed, acceleration, heading angle and direction change features from BeiDou positioning data through a sliding window; first-order difference analysis of driving speed, acceleration and heading angle to complete preprocessing; and fusion of preprocessed BeiDou positioning data and multi-source auxiliary sensor data into a multi-dimensional feature set. S2, Anomaly Mask Generation: Based on the BeiDou positioning data fused from the multi-dimensional feature set in S1, combined with statistical methods and the DBSCAN clustering algorithm, anomaly points and missing segments in the BeiDou positioning data are identified, and an anomaly mask matrix is ​​generated. S3. Trajectory Completion: Construct a deep learning model based on an Encoder-Decoder structure, incorporating an attention mechanism and introducing an autoencoder for assistance. Employ an end-to-end training framework and use a multi-objective loss function that includes a mean squared error loss term for positional error, a trajectory curvature continuity constraint loss term, and a physical rationality constraint loss term. Simulate and generate training samples with randomly missing segments on existing real trajectory data. Supervised learning training is then performed on the deep learning model to enable it to learn the contextual associations and physical constraints for trajectory completion. The trained deep learning model is then used in conjunction with the multi-dimensional feature set from S1 and the anomaly mask matrix from S2 to perform context-aware completion and repair of abnormal trajectory segments, outputting a continuous sequence of trajectory coordinates. S4. Real-time output integration and system deployment: Deploy the lightweight deep learning model on the vehicle terminal or edge computing node, integrate the multi-source auxiliary sensor data in the multi-dimensional feature set described in S1, perform physical rationality correction on the trajectory coordinate sequence completed in S3, input abnormal trajectories in real time through a sliding window method and output the repair results, and finally obtain the BeiDou positioning trajectory that conforms to the vehicle dynamics characteristics.

2. The method for repairing vehicle BeiDou positioning data anomalies based on deep learning according to claim 1, characterized in that, In S1, the vehicle's BeiDou positioning data includes at least longitude coordinates, latitude coordinates, driving speed, acceleration, and heading angle; the multi-source auxiliary sensor data includes at least the vehicle's inertial measurement unit (IMU) three-axis acceleration data, odometer cumulative driving distance data, and gyroscope heading angle velocity data.

3. The method for repairing anomalies in vehicle BeiDou positioning data based on deep learning according to claim 2, characterized in that, In step S2, identifying abnormal points and missing segments in BeiDou positioning data includes the following steps: S2.1 Extract BeiDou positioning data from the multi-dimensional feature set fused from S1; S2.2 First, the extracted BeiDou positioning data is input into statistical methods to analyze the temporal correlation of the data and screen out suspected abnormal data that exceed the normal fluctuation range; then, the suspected abnormal data and the remaining normal data are input into the DBSCAN clustering algorithm to further analyze the spatial distribution relationship of the data features. S2.

3. Based on the combined calculation results of statistical methods and DBSCAN clustering algorithm, isolated data points that are not classified into any cluster are identified as outliers in BeiDou positioning data, and trajectory segments that do not form effective clusters and meet the set time continuity conditions are identified as missing segments.

4. The method for repairing vehicle BeiDou positioning data anomalies based on deep learning according to claim 3, characterized in that, In step S2, generating the anomaly mask matrix includes the following steps: S2.4 Construct an anomaly mask matrix that matches the length of the BeiDou positioning data time series; S2.

5. Mark the elements in the anomaly mask matrix according to the rule that "abnormal points and missing segments are marked as 0, and normal trajectory positions are marked as 1". S2.6 After completing the anomaly mask matrix labeling, record the time index of the missing segment and extract the context features of the missing segment to form a complete annotation result containing the anomaly location, time range and context information.

5. The method for repairing vehicle BeiDou positioning data anomalies based on deep learning according to claim 4, characterized in that, In S3, a deep learning model based on an Encoder-Decoder structure, incorporating an attention mechanism and introducing an autoencoder is constructed, including the following steps: S3.

1. Employ a multi-layer bidirectional long short-term memory network Bi-LSTM or a multi-layer Transformer encoder layer to extract temporal features from the input multi-dimensional feature set and capture the long-range contextual association of the trajectory. S3.

2. A multi-head self-attention module is introduced between the encoder and decoder to enhance the influence of key trajectory segments on the completion result by calculating the weight coefficients between trajectory points. S3.

3. Introduce an autoencoder module to perform dimensionality reduction and reconstruction on the temporal features output by the encoder, and select the core features that match the trajectory motion law. S3.

4. Employ a multi-layer LSTM or Transformer decoder layer, combined with the anomaly mask matrix and encoder output, to generate the completed trajectory coordinate sequence step by step over time.

6. The method for repairing vehicle BeiDou positioning data anomalies based on deep learning according to claim 5, characterized in that, In step S3, the construction and calculation of the multi-objective loss function includes the following steps: S3.5 Calculate the coordinates of the predicted trajectory Compared with the actual trajectory coordinates when there are no anomalies The mean squared Euclidean distance is used to obtain the mean squared error loss term of the position error. ;in The time step index represents the trajectory coordinate sequence; S3.6 Calculate the rate of change of curvature of adjacent trajectory points in the completed trajectory to obtain the trajectory curvature continuity constraint loss term. ; S3.

7. Combining the IMU triaxial acceleration data from the multi-source auxiliary sensors in the S1 multi-dimensional feature set with the odometer cumulative driving distance data, calculate the deviation between the acceleration and velocity of the predicted trajectory and the corresponding sensor measured values, and introduce weighting coefficients. Balancing the impact of the two types of sensor data, we obtain the physical rationality constraint loss term. .

7. The method for repairing anomalies in vehicle BeiDou positioning data based on deep learning according to claim 6, characterized in that, In S3, simulating the generation of training samples with randomly missing fragments includes the following steps: S3.

8. For existing real trajectory data, two modes are used to randomly label missing segments: when using the continuous time missing mode, the length of a single missing segment is set to 5-20 time steps; when using the random discrete missing mode, the proportion of missing points in a single trajectory is controlled at 5%-20%. S3.

9. Use the context trajectory of the marked missing segment as the input feature. The context trajectory is the normal trajectory data of the missing segment in the 20-50 time steps before the start time step and the 20-50 time steps after the end time step. Use the real trajectory coordinates of the missing segment as the label to form a training sample pair corresponding to the input feature and the label.

8. The method for repairing vehicle BeiDou positioning data anomalies based on deep learning according to claim 7, characterized in that, In step S3, the context-aware completion and repair of trajectory anomaly segments using the trained deep learning model includes the following steps: S3.

10. Combining the start and end time steps of the abnormal segment of the trajectory with the abnormal mask matrix described in S2, extract the normal trajectory features of the 30-100 time steps before the start time step and the 30-100 time steps after the end time step of the abnormal segment as context input. S3.11 Input the context input and the anomaly mask matrix described in S2 into the trained deep learning model; S3.

12. The deep learning model assigns dynamic weights to contextual features through an attention mechanism, prioritizing the enhancement of the influence of trajectory points consistent with the movement trend of the abnormal segment, and generating the trajectory coordinates of the abnormal segment step by step. S3.

13. Output a continuous trajectory coordinate sequence, wherein the trajectory coordinate sequence maintains spatiotemporal continuity with the normal trajectory before and after the abnormal segment, and satisfies the physical constraints of the multi-objective loss function.

9. The method for repairing vehicle BeiDou positioning data anomalies based on deep learning according to claim 8, characterized in that, In step S4, the deployment of the lightweight deep learning model and the input of real-time data include the following steps: S4.1 Lightweighting of the trained deep learning model, including model parameter pruning and quantization compression, to adapt to the hardware operation requirements of vehicle terminals or edge computing nodes. S4.2 Configure a sliding window for real-time data acquisition. The time step range covered by the window must include the latest generated abnormal trajectory data to ensure that the temporal correlation information of the trajectory can be captured synchronously. S4.3 Real-time data input via sliding window: After each window data acquisition is completed, the abnormal trajectory data in the window and the multi-source auxiliary sensor data in the multi-dimensional feature set described in S1 are input synchronously to provide complete input information for deep learning model inference.

10. The method for repairing anomalies in vehicle BeiDou positioning data based on deep learning according to claim 9, characterized in that, In step S4, the output of the physical rationality correction and repair results includes the following steps: S4.

4. Integrate the IMU three-axis acceleration data, odometer cumulative driving distance data, and gyroscope heading angular velocity data from the multi-dimensional feature set described in S1, and construct constraint logic based on vehicle dynamics constraints to verify the rationality of trajectory motion parameters; S4.5 Based on the constraint logic of S4.4, the continuous trajectory coordinate sequence output by S3 is verified and corrected step by step: if the motion parameters of a trajectory point at a certain time step do not match the actual motion state of the vehicle reflected by the multi-source auxiliary sensor data, the coordinates of the trajectory point are adjusted in combination with the sensor data so that the trajectory conforms to the vehicle motion law. S4.6 Output the corrected BeiDou positioning trajectory. The trajectory data must include the longitude coordinates, latitude coordinates, driving speed and heading angle for each time step, and be consistent with the core parameter dimensions of the vehicle BeiDou positioning data in S1.

Images