Intelligent classification method for Beidou positioning data

Abstract

The invention provides an intelligent classification method for Beidou positioning data, and relates to the technical field of Beidou positioning data classification, and the method comprises the steps: obtaining the Beidou positioning data of a monitoring target; performing standardization processing, abnormal value processing and time sequence alignment processing on the acquired Beidou positioning data to generate a preprocessed time sequence data matrix; constructing a classification model; performing feature extraction, including time feature extraction and spatial feature extraction, on the preprocessed time sequence data matrix by using a spatial-temporal feature extraction module to obtain a time feature vector and a spatial feature vector respectively; carrying out fusion processing on the time feature vector and the space feature vector by adopting a spatial-temporal feature fusion module, and constructing a spatial-temporal joint feature; and inputting the spatio-temporal joint features into a classification module, and obtaining a classification result through feature conversion, attention weight calculation and feature aggregation. According to the invention, accurate identification and classification of the motion state of the monitored target are realized by establishing a data processing and classification framework of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of Beidou positioning data classification, and in particular to an intelligent classification method for Beidou positioning data. Background Art

[0002] In addition to providing basic positioning services, the Beidou satellite navigation system can also obtain multi-dimensional information such as satellite observation data and meteorological data. This information can be used to monitor and identify the movement status of targets and has important application value in the fields of traffic monitoring and logistics tracking.

[0003] At present, the following technical solutions are mainly used for target status classification based on satellite navigation positioning information: first, the original observation data is preprocessed, including signal quality assessment, data standardization and other steps; then the statistical characteristics of the time series are extracted, such as mean, standard deviation, maximum and minimum values, etc.; finally, based on the extracted features, rule judgment or traditional classification algorithm is used for status recognition.

[0004] In practical applications, since satellite signals are affected by multipath effects, ionospheric delays and other factors, the observed data often contain noise and outliers. Traditional anomaly detection methods mainly rely on fixed thresholds for judgment, which is difficult to adapt to complex and changing observation environments.

[0005] In terms of data processing, commonly used methods include median filtering and Kalman filtering. These methods can suppress noise to a certain extent, but they often require pre-setting of noise model parameters, and the processing effect is greatly affected by the selection of initial parameters. In terms of feature extraction, traditional methods mainly focus on statistical features of a single dimension and fail to fully utilize the multi-dimensional observation information provided by the Beidou system. The main technical problems in this field are: satellite navigation positioning information is easily affected by the external environment and generates noise and outliers, and existing technologies are difficult to guarantee the reliability of data processing and the accuracy of classification results.

[0006] Therefore, a classification method for Beidou positioning data is urgently needed to solve the above technical problems. Summary of the invention

[0007] In view of this, the present invention proposes an intelligent classification method for Beidou positioning data, which realizes accurate identification and classification of the motion state of the monitored target by establishing a systematic data processing and classification framework.

[0008] The technical solution of the present invention is achieved in this way:

[0009] The present invention provides an intelligent classification method for Beidou positioning data, comprising:

[0010] S1. Monitor the monitoring target in real time through monitoring equipment and obtain Beidou positioning data of the monitoring target at regular intervals. Beidou positioning data includes Beidou satellite original observation data, meteorological data, ionospheric data and multipath effect data;

[0011] S2, performing standardization, outlier processing and time series alignment processing on the acquired Beidou positioning data to generate a preprocessed time series data matrix;

[0012] S3, constructing a classification model, including a spatiotemporal feature extraction module, a spatiotemporal feature fusion module and a classification module;

[0013] S4, using the spatiotemporal feature extraction module to perform feature extraction on the preprocessed time series data matrix, including time feature extraction and space feature extraction, to obtain a time feature vector and a space feature vector respectively;

[0014] S5, using the spatiotemporal feature fusion module to fuse the temporal feature vector and the spatial feature vector to construct a spatiotemporal joint feature;

[0015] S6. Input the spatiotemporal joint features into the classification module, and obtain the classification results through feature conversion, attention weight calculation and feature aggregation.

[0016] Based on the above technical solution, preferably, step S2 includes:

[0017] S21. Perform zero-mean normalization processing on the Beidou satellite original observation data, meteorological data, ionospheric data and multipath effect data, respectively, so that data of different dimensions are converted to the same scale space;

[0018] S22. Detect and process outliers on the standardized Beidou positioning data, calculate the statistical characteristics of the data based on the sliding time window, identify outliers using the improved 3σ criterion, and correct the identified outliers using the local mean value replacement method;

[0019] S23, aligning the processed data in time series to detect the consistency of data sampling time. For data points with inconsistent sampling time, use the cubic spline interpolation algorithm to resample the time series to ensure that all features are aligned in the time dimension.

[0020] S24. Reorganize the aligned data into a time series data matrix, where the rows of the matrix represent time steps and the columns represent different feature dimensions.

[0021] Based on the above technical solution, preferably, step S22 includes:

[0022] S221, construct a multi-scale time window sequence {W1,W2,...,W k}, where Wi =base window ×2 i-1 ;

[0023] S222, calculating local statistical characteristics at each time window scale, including local mean, local standard deviation, local skewness and local kurtosis;

[0024] S223, calculating the dynamic threshold:

[0025] th i (t) = σ i (t)×[α+β·|s i (t)|+γ·log(1+|k i (t)|)]

[0026] Among them, α is the basic threshold coefficient, β is the skewness influence factor, γ is the kurtosis influence factor, σ i (t) is the local standard deviation of the data when the time window scale is t, s i (t) is the local skewness of the data when the time window scale is t, k i (t) is the local kurtosis of the data when the time window scale is t;

[0027] S224, merging the dynamic thresholds of each scale:

[0028] FU th (t) = ∑(w i ×th i (t))

[0029] Among them, FU th (t) is the final threshold, w i is the weight, which is calculated by the attention mechanism;

[0030] S225, set the fuzzy judgment interval to [FU th (t)×(1-ε),FU th (t)×(1+ε)], where ε is the judgment adjustment factor;

[0031] S226. If the data point falls within the fuzzy determination interval, the data point is determined to be an outlier, and the identified outlier is corrected using a local mean value replacement method.

[0032] On the basis of the above technical solution, preferably, the cubic spline interpolation algorithm includes:

[0033] A comprehensive evaluation index based on signal quality and motion characteristics is constructed to adaptively select the resampling time point. The calculation formula of the comprehensive evaluation index Q(t) is:

[0034]

[0035] In the formula, SNR(t) represents the Beidou signal noise ratio at time t, δ pos (t) represents the rate of change of position, δ vet (t) represents the speed change rate, pos max and vet max are the normalized parameters for position and velocity changes, respectively;

[0036] The distribution of the resampling time period is determined based on the comprehensive evaluation index Q(t), and the time interval Δt is used for division:

[0037]

[0038] In the formula, base interval is the reference sampling interval, Q max is the maximum value of the comprehensive evaluation index, and λ is the adjustment coefficient;

[0039] Construct hybrid spline basis functions:

[0040] B i (t) = ω(t)·N i (t)+(1-ω(t))·H i (t)

[0041] Where N i (t) is the standard cubic B-spline basis function, H i (t) is the Hermite spline basis function considering the physical properties of Beidou positioning data, and ω(t) is the adaptive weight function: ω(t) = sigmoid(η·Q(t)-Q th ), η is the weight adjustment parameter, Q th is the quality assessment threshold;

[0042] Construct the interpolation function:

[0043]

[0044] In the formula, c i is the interpolation coefficient, is the regularization term, which is calculated based on the data smoothness constraint and the data jitter constraint;

[0045] The data points with inconsistent sampling times are resampled according to the interpolation function to complete the time alignment.

[0046] Based on the above technical solution, preferably, the classification model includes:

[0047] Temporal feature extraction module: It consists of a bidirectional LSTM layer, a temporal attention layer, and a feature fusion layer. The bidirectional LSTM layer contains two LSTM units, the forward and backward LSTM units. Each unit contains an input gate, a forget gate, and an output gate. The temporal attention layer uses a scaled dot-product attention mechanism, and the feature fusion layer uses a residual connection structure.

[0048] Spatial feature extraction module: It includes multiple layers of spatial convolutional layers, batch normalization layers, and spatial attention layers. The spatial convolutional layers use 3×3 convolution kernels and ReLU activation functions, the batch normalization layers are used for feature standardization, and the spatial attention layers use channel-wise attention mechanisms.

[0049] Spatiotemporal feature fusion module: It includes feature interaction layer, correlation calculation layer and feature enhancement layer. The feature interaction layer adopts multi-head attention mechanism, the correlation calculation layer uses similarity measurement calculation, and the feature enhancement layer adopts gated linear unit structure.

[0050] Classification module: includes feature conversion layer, global attention layer and classification layer. The feature conversion layer uses a fully connected layer with LayerNorm, the global attention layer adopts a self-attention mechanism, and the classification layer uses a softmax activation function.

[0051] Based on the above technical solution, preferably, the extraction process of the time feature vector is:

[0052] Based on the preprocessed time series data matrix, a bidirectional LSTM network is used to extract features from the time series data in the forward and reverse directions. The forward LSTM captures the time series dependency from the past to the present, and the reverse LSTM captures the time series dependency from the present to the past.

[0053] For the feature sequence extracted by bidirectional LSTM, a temporal attention mechanism is constructed, and the temporal attention weight is calculated using the temporal attention mechanism;

[0054] The feature sequence extracted by the bidirectional LSTM is weighted and fused with the corresponding attention weights to generate the final time feature vector: F temp =∑(a(t)·[hf(t);hb(t)]), where hf(t) is the feature extracted by the forward LSTM, hb(t) is the feature extracted by the backward LSTM, and a(t) is the corresponding temporal attention weight.

[0055] Based on the above technical solution, preferably, the process of extracting the spatial feature vector is:

[0056] The preprocessed time series data matrix is ​​subjected to feature extraction through a multi-layer spatial convolutional network. Each layer of the network includes convolution operations, batch normalization, and ReLU activation functions.

[0057] In the process of feature extraction, the channel attention mechanism is introduced to perform global information statistics on each feature channel, and the channel attention weight of each channel is calculated through nonlinear transformation;

[0058] Adaptively fuse the multi-layer features with attention weights to generate the final spatial feature vector: F space =∑(b(t)·F(t)), where F(t) is the multi-layer feature extracted by the multi-layer spatial convolutional network, and b(t) is the corresponding channel attention weight.

[0059] Based on the above technical solution, preferably, the construction process of the spatiotemporal joint feature is:

[0060] Use multi-head attention mechanism to realize the interaction between temporal feature vector and spatial feature vector;

[0061] Calculate the feature correlation between the temporal feature vector and the spatial feature vector:

[0062]

[0063] In the formula, <·,·> represents the inner product operation, ‖·‖ represents the L2 norm;

[0064] Construct a gated linear unit g, design a gating mechanism, dynamically adjust the fusion ratio of temporal features and spatial features according to the feature correlation, and introduce a residual connection structure to generate spatiotemporal joint features:

[0065] F joint =g⊙tanh(W f [F temp ‖F space ])+(1-g)⊙(W t F temp +W s F space

[0066] Among them, F joint is the spatiotemporal joint feature, g = sigmoid (W g [F temp ‖F space ]), represents the gated linear unit calculation, ∥ represents the feature concatenation operation, ⊙ represents the element-by-element multiplication, W f , W t , W s and W g is the learnable parameter matrix, sigmoid and tanh are the corresponding function operations.

[0067] Based on the above technical solution, preferably, step S6 includes:

[0068] S61, using a fully connected layer to perform an initial transformation on the spatiotemporal joint features, and then performing feature normalization processing through layer normalization, where the layer normalization includes calculating the feature mean and variance, and performing normalization transformation using a learnable scaling factor and bias parameter;

[0069] S62, using the self-attention mechanism to perform global context modeling on the standardized features, specifically generating a query vector, a key vector, and a value vector through linear transformation, calculating the attention weight score, and performing a residual connection between the attention-weighted features and the input features;

[0070] S63. Use a two-layer feedforward neural network to perform nonlinear transformation on the features. The first layer uses the ReLU activation function for feature mapping, the second layer performs dimension conversion, and introduces residual connections after each layer of conversion to preserve the original information.

[0071] S64. Map the features to the predefined category space through the fully connected layer, use the softmax function to calculate the probability distribution of each category, and output the classification result.

[0072] On the basis of the above technical solution, preferably, the spatial convolution layer in the spatial feature extraction module adopts a dilated convolution structure, and the dilation rate increases layer by layer with the depth of the network. The dilation rate d l Set it up as follows:

[0073] d l =2 l-1 ,l=1,2,...,L

[0074] Where L is the total number of spatial convolutional layers.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] (1) The present invention realizes the systematic processing of Beidou positioning data by constructing a classification model including a spatiotemporal feature extraction module, a spatiotemporal feature fusion module and a classification module. The method adopts a multi-level data processing strategy and a deep feature learning mechanism, effectively solving the technical problems of data anomalies, inconsistent sampling and insufficient feature extraction in the target state classification process of Beidou positioning data, and improving the accuracy and reliability of classification;

[0077] (2) The present invention adopts an outlier detection method that combines a multi-scale time window and a dynamic threshold. By constructing a dynamic threshold based on local statistical features and introducing a fuzzy decision interval mechanism, the problem of poor adaptability of the traditional fixed threshold method to dynamic changes in data is overcome, thereby improving the accuracy and robustness of outlier detection.

[0078] (3) The present invention designs an adaptive resampling method based on signal quality and motion characteristics. By constructing a comprehensive evaluation index to dynamically adjust the sampling interval and using a hybrid spline basis function for interpolation, the present invention solves the technical problem that the traditional fixed-interval resampling method is difficult to adapt to changes in signal quality and improves the accuracy of time series alignment.

[0079] (4) The present invention extracts temporal features and spatial features respectively through a bidirectional LSTM network and a spatial convolutional network, and introduces an attention mechanism for feature enhancement, thereby overcoming the problem that traditional feature extraction methods are difficult to capture deep features of data, and improving the richness and distinctiveness of feature expression;

[0080] (5) The present invention designs a feature fusion method based on a multi-head attention mechanism and a gated linear unit. By calculating the feature correlation, the fusion ratio is adaptively adjusted, and a residual connection is introduced to maintain the original information. This solves the technical problem that traditional feature fusion methods are difficult to effectively integrate spatiotemporal features, and enhances the adaptability and effectiveness of feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0082] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0083] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] like Figure 1 As shown, the present invention provides an intelligent classification method for Beidou positioning data, comprising:

[0085] S1. Monitor the monitoring target in real time through monitoring equipment and obtain Beidou positioning data of the monitoring target at regular intervals. Beidou positioning data includes Beidou satellite original observation data, meteorological data, ionospheric data and multipath effect data;

[0086] S2, performing standardization, outlier processing and time series alignment processing on the acquired Beidou positioning data to generate a preprocessed time series data matrix;

[0087] S3, constructing a classification model, including a spatiotemporal feature extraction module, a spatiotemporal feature fusion module and a classification module;

[0088] S4, using the spatiotemporal feature extraction module to perform feature extraction on the preprocessed time series data matrix, including time feature extraction and space feature extraction, to obtain a time feature vector and a space feature vector respectively;

[0089] S5, using the spatiotemporal feature fusion module to fuse the temporal feature vector and the spatial feature vector to construct a spatiotemporal joint feature;

[0090] S6. Input the spatiotemporal joint features into the classification module, and obtain the classification results through feature conversion, attention weight calculation and feature aggregation.

[0091] Specifically, in one embodiment of the present invention, the monitoring device refers to a Beidou navigation and positioning terminal device, which receives Beidou satellite signals and has data storage and transmission functions. The monitoring target can be a vehicle, a ship, or other moving objects. After determining the monitoring target to be analyzed, the system obtains the Beidou positioning data stored in the monitoring device corresponding to the monitoring target, which specifically includes Beidou satellite original observation data, meteorological data, ionospheric data, and multipath effect data. Specifically, the original observation data of Beidou satellites include pseudorange measurement values, carrier phase measurement values, Doppler frequency shift, and signal strength indicators, which directly reflect the position and motion information of the monitored target; meteorological data include meteorological parameters such as atmospheric pressure, temperature, humidity, wind speed and direction, which reflect the atmospheric conditions of the signal propagation environment. These factors will affect the signal propagation path and intensity, and thus affect the positioning accuracy; ionospheric data include ionospheric delay, electron density, TEC (total electron content) value, and ionospheric scintillation index, which characterize the impact of the high-altitude ionosphere on signal propagation. These data can help compensate for the signal delay and distortion caused by the ionosphere; multipath effect data include multipath signal strength, multipath delay, multipath phase error, and multipath attenuation factor, which reflect the reflection and diffraction of the signal during the propagation process. This information helps to identify and eliminate multipath interference.

[0092] Specifically, in one embodiment of the present invention, step S2 includes:

[0093] S21. Perform zero-mean normalization processing on the Beidou satellite original observation data, meteorological data, ionospheric data and multipath effect data respectively, so that data of different dimensions can be converted to the same scale space.

[0094] Specifically, the mean μ and standard deviation σ of the four types of data, namely, the original observation data of Beidou satellites, meteorological data, ionospheric data, and multipath effect data, are calculated, and the standardized transformation x is performed according to the mean and standard deviation. norm =(x-μ) / σ. In this embodiment, different processing strategies are used for different types of data: Beidou satellite raw observation data: pseudorange, carrier phase, etc. are grouped and standardized; meteorological data: temperature, humidity and other parameters are standardized separately; ionospheric data: delay, electron density and other parameters are standardized; multipath effect data: signal strength, delay and other parameters are standardized.

[0095] S22. Perform outlier detection and processing on the standardized Beidou positioning data, calculate the statistical characteristics of the data based on the sliding time window, use the improved 3σ criterion to identify the outliers, and correct the identified outliers using the local mean value replacement method.

[0096] In this embodiment, step S22 includes:

[0097] S221, construct a multi-scale time window sequence {W1,W2,...,W k}, where W i =base window ×2 i-1 ;

[0098] S222, calculating local statistical characteristics at each time window scale, including local mean, local standard deviation, local skewness and local kurtosis;

[0099] S223, calculating the dynamic threshold:

[0100] th i (t) = σ i (t)×[α+β·|s i (t)|+γ·log(1+|k i (t)|)]

[0101] Among them, α is the basic threshold coefficient, β is the skewness influence factor, γ is the kurtosis influence factor, σ i (t) is the local standard deviation of the data when the time window scale is t, s i (t) is the local skewness of the data when the time window scale is t, k i (t) is the local kurtosis of the data when the time window scale is t;

[0102] S224, merging the dynamic thresholds of each scale:

[0103] FU th (t) = ∑(w i ×th i(t))

[0104] Among them, FU th (t) is the final threshold, w i is the weight, which is calculated by the attention mechanism;

[0105] S225, set the fuzzy judgment interval to [FU th (t)×(1-ε),FU th (t)×(1+ε)], where ε is the judgment adjustment factor;

[0106] S226. If the data point falls within the fuzzy determination interval, the data point is determined to be an outlier, and the identified outlier is corrected using a local mean value replacement method.

[0107] The improved 3σ criterion of the present invention effectively solves the limitation problem of the traditional 3σ criterion in processing non-stationary time series data by introducing multi-scale time windows and dynamic threshold mechanism. Specifically, by constructing {W1,W2,...,W k}Multi-scale window sequence, combined with local statistical features (mean, standard deviation, skewness, kurtosis), established a dynamic threshold calculation method based on data distribution characteristics. This improvement realizes the adaptive fusion of multi-scale thresholds through the attention mechanism, and introduces fuzzy judgment intervals, which significantly improves the accuracy and robustness of outlier detection. This method is suitable for processing complex abnormal patterns caused by signal quality fluctuations, multipath effects and other factors in Beidou positioning data. It can accurately identify and process outliers at different scales, while avoiding the overfitting problem caused by fixed thresholds in traditional methods.

[0108] S23. Perform time series alignment on the processed data to detect the consistency of data sampling time. For data points with inconsistent sampling time, use the cubic spline interpolation algorithm to perform time series resampling to ensure that all features are aligned in the time dimension.

[0109] In this embodiment, the cubic spline interpolation algorithm includes:

[0110] A comprehensive evaluation index based on signal quality and motion characteristics is constructed to adaptively select the resampling time point. The calculation formula of the comprehensive evaluation index Q(t) is:

[0111]

[0112] In the formula, SNR(t) represents the Beidou signal noise ratio at time t, δ pos (t) represents the rate of change of position, δ vet (t) represents the speed change rate, pos max and vet max are the normalized parameters for position and velocity changes, respectively;

[0113] The distribution of the resampling time period is determined based on the comprehensive evaluation index Q(t), and the time interval Δt is used for division:

[0114]

[0115] In the formula, base interval is the reference sampling interval, Q max is the maximum value of the comprehensive evaluation index, and λ is the adjustment coefficient;

[0116] Construct hybrid spline basis functions:

[0117] B i (t) = ω(t)·N i (t)+(1-ω(t))·H i (t)

[0118] Where N i (t) is the standard cubic B-spline basis function, H i (t) is the Hermite spline basis function considering the physical properties of Beidou positioning data, and ω(t) is the adaptive weight function: ω(t) = sigmoid(η·Q(t)-Q th ), η is the weight adjustment parameter, Q th is the quality assessment threshold;

[0119] Construct the interpolation function:

[0120]

[0121] In the formula, c i is the interpolation coefficient, is the regularization term, which is calculated based on the data smoothness constraint and the data jitter constraint;

[0122] The data points with inconsistent sampling times are resampled according to the interpolation function to complete the time alignment.

[0123] S24. Reorganize the aligned data into a time series data matrix, where the rows of the matrix represent time steps and the columns represent different feature dimensions.

[0124] The present invention realizes adaptive control of the interpolation process by introducing a comprehensive evaluation index Q(t) based on signal quality and motion characteristics. i(t), the standard cubic B-spline basis function is dynamically fused with the Hermite spline basis function that considers the physical characteristics of Beidou positioning data, and the adaptive weight function ω(t) is introduced for regulation. This method solves the problem of insufficient accuracy of traditional interpolation algorithms when processing Beidou positioning data with non-uniform sampling and unstable signal quality. By comprehensively considering the signal-to-noise ratio, position change rate and speed change rate, the intelligent selection of resampling time points and the adaptive adjustment of interpolation parameters are realized, which improves the accuracy and reliability of time series alignment. It is suitable for processing Beidou positioning data with complex situations such as signal jumps and data missing.

[0125] Specifically, in one embodiment of the present invention, the classification model includes:

[0126] Temporal feature extraction module: It consists of a bidirectional LSTM layer, a temporal attention layer, and a feature fusion layer. The bidirectional LSTM layer contains two LSTM units, the forward and backward LSTM units. Each unit contains an input gate, a forget gate, and an output gate. The temporal attention layer uses a scaled dot-product attention mechanism, and the feature fusion layer uses a residual connection structure.

[0127] Spatial feature extraction module: It includes multiple layers of spatial convolutional layers, batch normalization layers, and spatial attention layers. The spatial convolutional layers use 3×3 convolution kernels and ReLU activation functions, the batch normalization layers are used for feature standardization, and the spatial attention layers use channel-wise attention mechanisms.

[0128] Spatiotemporal feature fusion module: It includes feature interaction layer, correlation calculation layer and feature enhancement layer. The feature interaction layer adopts multi-head attention mechanism, the correlation calculation layer uses similarity measurement calculation, and the feature enhancement layer adopts gated linear unit structure.

[0129] Classification module: includes feature conversion layer, global attention layer and classification layer. The feature conversion layer uses a fully connected layer with LayerNorm, the global attention layer adopts a self-attention mechanism, and the classification layer uses a softmax activation function.

[0130] Among them, the spatial convolution layer in the spatial feature extraction module adopts a dilated convolution structure, and the dilation rate increases layer by layer with the depth of the network. The dilation rate d l Set it up as follows:

[0131] d l =2 l-1 ,l=1,2,...,L

[0132] Where L is the total number of spatial convolutional layers.

[0133] In one example of the present invention, the temporal feature vector extraction process is:

[0134] Based on the preprocessed time series data matrix, a bidirectional LSTM network is used to extract features of the time series data from the forward and reverse directions respectively. The forward LSTM captures the time series dependency from the past to the present: the sequence is processed from t=1 to T in sequence, and the hidden state hf(t) is output at each time step; the reverse LSTM captures the time series dependency from the present to the past: the sequence is processed from t=T to 1 in sequence, and the hidden state hb(t) is output at each time step.

[0135] For the feature sequence extracted by bidirectional LSTM, a temporal attention mechanism is constructed, and the temporal attention weight is calculated using the temporal attention mechanism. The query vector q(t) is calculated for each time step t, and the similarity is calculated with the key vector k(t) of all time steps, and the temporal attention weight is obtained by softmax normalization.

[0136] The feature sequence extracted by the bidirectional LSTM is weighted and fused with the corresponding attention weights to generate the final time feature vector: F temp =∑(a(t)·[hf(t);hb(t)]), where hf(t) is the feature extracted by the forward LSTM, hb(t) is the feature extracted by the backward LSTM, and a(t) is the corresponding temporal attention weight.

[0137] In one example of the present invention, the process of extracting the spatial feature vector is as follows:

[0138] The preprocessed time series data matrix is ​​subjected to feature extraction through a multi-layer spatial convolutional network. Each layer of the network includes convolution operations, batch normalization, and ReLU activation functions. The input data passes through 4 layers of dilated convolutions: the size of the feature map of each layer remains unchanged, the receptive field expands layer by layer, and the number of feature channels increases layer by layer.

[0139] In the process of feature extraction, the channel attention mechanism is introduced to perform global information statistics on each feature channel, and calculate the channel attention weight of each channel through nonlinear transformation; for each feature channel: calculate the global average pooling value, calculate the global maximum pooling value, and generate the channel attention weight b(t) through the shared fully connected layer.

[0140] Adaptively fuse the multi-layer features with attention weights to generate the final spatial feature vector: F space =∑(b(t)·F(t)), where F(t) is the multi-layer feature extracted by the multi-layer spatial convolutional network, and b(t) is the corresponding channel attention weight.

[0141] In one example of the present invention, the process of constructing the spatiotemporal joint feature is as follows:

[0142] A multi-head attention mechanism is used to realize the interaction between the temporal feature vector and the spatial feature vector. The temporal feature vector and the spatial feature vector are input into the multi-head attention mechanism, each attention head independently calculates the attention weight, and the outputs of multiple heads are concatenated and linearly transformed.

[0143] Calculate the feature correlation between the temporal feature vector and the spatial feature vector:

[0144]

[0145] In the formula, <·,·> represents the inner product operation, and ‖·‖ represents the L2 norm.

[0146] Construct a gated linear unit g, design a gating mechanism, dynamically adjust the fusion ratio of temporal features and spatial features according to the feature correlation, and introduce a residual connection structure to generate spatiotemporal joint features:

[0147] F joint =g⊙tanh(W f [F temp ‖F space ])+(1-g)⊙(W t F temp +W s F space

[0148] Among them, F joint is the spatiotemporal joint feature, g = sigmoid (W g [F temp ‖F space ]), represents the gated linear unit calculation, ∥ represents the feature concatenation operation, ⊙ represents the element-by-element multiplication, W f , W t , W s and W g is the learnable parameter matrix, sigmoid and tanh are the corresponding function operations.

[0149] Specifically, in one embodiment of the present invention, step S6 includes:

[0150] S61. Use the fully connected layer to perform an initial transformation on the spatiotemporal joint features, and then perform feature normalization through layer normalization. Layer normalization includes calculating the feature mean and variance, and using a learnable scaling factor and bias parameter for normalization transformation; the normalization formula is: Among them, μ is the characteristic mean, σ 2 is the variance, γ is a learnable scaling factor, β is a bias parameter, and ε is a small constant to prevent division by zero.

[0151] S62. Use the self-attention mechanism to model the global context of the standardized features. Specifically, generate the query vector, key vector and value vector through linear transformation, and calculate the attention weight score: And the attention-weighted features are weighted feature Residual connection with input features: output = weighted feature +input feature ; to keep the original information intact.

[0152] S63. Use a two-layer feedforward neural network to perform nonlinear transformation on the features. The first layer uses the ReLU activation function for feature mapping, the second layer performs dimensionality conversion, and a residual connection is introduced after each layer of conversion to preserve the original information.

[0153] S64, map the features to the predefined category space through the fully connected layer, use the softmax function to calculate the probability distribution of each category, and output the classification result. Specifically, during classification, the input is the features of the previous step, and the output is a vector of the predefined number of categories. After calculating the probability distribution of each category, select the category with the highest probability as the prediction result.

[0154] In summary, the present invention constructs a complete set of Beidou positioning data intelligent classification methods: first, the Beidou positioning raw data is preprocessed by the improved 3σ criterion and cubic spline interpolation algorithm; then a bidirectional LSTM and a dilated convolutional network are designed to extract time and space features respectively, and the effective fusion of features is achieved through a multi-head attention mechanism; finally, technical means such as layer normalization, self-attention mechanism and residual connection are used for feature optimization and classification. This method can output the classification probability distribution of the monitoring target and obtain the final category determination result. The classification result can effectively identify and distinguish the behavior patterns and motion characteristics of different types of monitoring targets, and provide reliable technical support for the intelligent application of Beidou positioning data, which is particularly suitable for scenarios where real-time classification and behavior recognition of monitoring targets are required.

[0155] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent classification method for Beidou positioning data, characterized in that: include: S1. Monitor the monitoring target in real time through monitoring equipment and obtain Beidou positioning data of the monitoring target at regular intervals. Beidou positioning data includes Beidou satellite original observation data, meteorological data, ionospheric data and multipath effect data; S2, performing standardization, outlier processing and time series alignment processing on the acquired Beidou positioning data to generate a preprocessed time series data matrix; S3, constructing a classification model, including a spatiotemporal feature extraction module, a spatiotemporal feature fusion module and a classification module; S4, using the spatiotemporal feature extraction module to perform feature extraction on the preprocessed time series data matrix, including time feature extraction and space feature extraction, to obtain a time feature vector and a space feature vector respectively; S5, using the spatiotemporal feature fusion module to fuse the temporal feature vector and the spatial feature vector to construct a spatiotemporal joint feature; S6. Input the spatiotemporal joint features into the classification module, and obtain the classification results through feature conversion, attention weight calculation and feature aggregation.

2. The intelligent classification method of Beidou positioning data as claimed in claim 1, characterized in that: Step S2 includes: S21. Perform zero-mean normalization processing on the Beidou satellite original observation data, meteorological data, ionospheric data and multipath effect data, respectively, so that data of different dimensions are converted to the same scale space; S22. Detect and process outliers on the standardized Beidou positioning data, calculate the statistical characteristics of the data based on the sliding time window, identify outliers using the improved 3σ criterion, and correct the identified outliers using the local mean value replacement method; S23, aligning the processed data in time series to detect the consistency of data sampling time. For data points with inconsistent sampling time, use the cubic spline interpolation algorithm to resample the time series to ensure that all features are aligned in the time dimension. S24. Reorganize the aligned data into a time series data matrix, where the rows of the matrix represent time steps and the columns represent different feature dimensions.

3. The intelligent classification method of Beidou positioning data as claimed in claim 2, characterized in that: Step S22 includes: S221, construct a multi-scale time window sequence {W1, W2, ..., W k }, where W i =base window ×2 i-1 ; S222, calculating local statistical characteristics at each time window scale, including local mean, local standard deviation, local skewness and local kurtosis; S223, calculating the dynamic threshold: th i (t)=σ i (t)×[α+β·|s i (t)|+γ·log(1+|k i (t)|)] Among them, α is the basic threshold coefficient, β is the skewness influence factor, γ is the kurtosis influence factor, σ i (t) is the local standard deviation of the data when the time window scale is t, s i (t) is the local skewness of the data when the time window scale is t, k i (t) is the local kurtosis of the data when the time window scale is t; S224, merging the dynamic thresholds of each scale: FU th (t)=∑(w i ×th i (t)) Among them, FU th (t) is the final threshold, w i is the weight, which is calculated by the attention mechanism; S225, set the fuzzy judgment interval to [FU th (t)×(1-ε),FU th (t)×(1+ε)], where ε is the judgment adjustment factor; S226. If the data point falls within the fuzzy determination interval, the data point is determined to be an outlier, and the identified outlier is corrected using a local mean value replacement method.

4. The intelligent classification method of Beidou positioning data as claimed in claim 2, characterized in that: The cubic spline interpolation algorithm includes: A comprehensive evaluation index based on signal quality and motion characteristics is constructed to adaptively select the resampling time point. The calculation formula of the comprehensive evaluation index Q(t) is: In the formula, SNR(t) represents the Beidou signal noise ratio at time t, δ pos (t) represents the rate of change of position, δ vet (t) represents the speed change rate, pos max and vet max are the normalized parameters for position and velocity changes, respectively; The distribution of the resampling time period is determined based on the comprehensive evaluation index Q(t), and the time interval Δt is used for division: In the formula, base interval is the reference sampling interval, Q max is the maximum value of the comprehensive evaluation index, and λ is the adjustment coefficient; Construct hybrid spline basis functions: B i (t)=ω(t)·Ni(t)+(1-ω(t))·H i (t) Where N i (t) is the standard cubic B-spline basis function, H i (t) is the Hermite spline basis function considering the physical properties of Beidou positioning data, and ω(t) is the adaptive weight function: ω(t) = sigmoid(η·Q(t)-Q th ), η is the weight adjustment parameter, Q th is the quality assessment threshold; Construct the interpolation function: In the formula, c i is the interpolation coefficient, is the regularization term, which is calculated based on the data smoothness constraint and the data jitter constraint; The data points with inconsistent sampling times are resampled according to the interpolation function to complete the time alignment.

5. The intelligent classification method of Beidou positioning data as claimed in claim 1, characterized in that: Classification models include: Temporal feature extraction module: It consists of a bidirectional LSTM layer, a temporal attention layer, and a feature fusion layer. The bidirectional LSTM layer contains two LSTM units, the forward and backward LSTM units. Each unit contains an input gate, a forget gate, and an output gate. The temporal attention layer uses a scaled dot-product attention mechanism, and the feature fusion layer uses a residual connection structure. Spatial feature extraction module: It includes multiple layers of spatial convolutional layers, batch normalization layers, and spatial attention layers. The spatial convolutional layers use 3×3 convolution kernels and ReLU activation functions, the batch normalization layers are used for feature standardization, and the spatial attention layers use channel-wise attention mechanisms. Spatiotemporal feature fusion module: It includes feature interaction layer, correlation calculation layer and feature enhancement layer. The feature interaction layer adopts multi-head attention mechanism, the correlation calculation layer uses similarity measurement calculation, and the feature enhancement layer adopts gated linear unit structure. Classification module: includes feature conversion layer, global attention layer and classification layer. The feature conversion layer uses a fully connected layer with LayerNorm, the global attention layer adopts a self-attention mechanism, and the classification layer uses a softmax activation function.

6. The intelligent classification method of Beidou positioning data as claimed in claim 5, characterized in that: The extraction process of time feature vector is: Based on the preprocessed time series data matrix, a bidirectional LSTM network is used to extract features from the time series data in the forward and reverse directions. The forward LSTM captures the time series dependency from the past to the present, and the reverse LSTM captures the time series dependency from the present to the past. For the feature sequence extracted by bidirectional LSTM, a temporal attention mechanism is constructed, and the temporal attention weight is calculated using the temporal attention mechanism; The feature sequence extracted by the bidirectional LSTM is weighted and fused with the corresponding attention weights to generate the final time feature vector: F temp =∑(a(t)·[hf(t);hb(t)]), where hf(t) is the feature extracted by the forward LSTM, hb(t) is the feature extracted by the backward LSTM, and a(t) is the corresponding temporal attention weight.

7. The intelligent classification method of Beidou positioning data as claimed in claim 6, characterized in that the space The feature vector extraction process is: The preprocessed time series data matrix is ​​subjected to feature extraction through a multi-layer spatial convolutional network. Each layer of the network includes convolution operations, batch normalization, and ReLU activation functions. In the process of feature extraction, the channel attention mechanism is introduced to perform global information statistics on each feature channel, and the channel attention weight of each channel is calculated through nonlinear transformation; Adaptively fuse the multi-layer features with attention weights to generate the final spatial feature vector: F space =∑(b(t)·F(t)), where F(t) is the multi-layer feature extracted by the multi-layer spatial convolutional network, and b(t) is the corresponding channel attention weight.

8. The intelligent classification method of Beidou positioning data as claimed in claim 7, characterized in that: The construction process of spatiotemporal joint features is: Use multi-head attention mechanism to realize the interaction between temporal feature vector and spatial feature vector; Calculate the feature correlation between the temporal feature vector and the spatial feature vector: In the formula, <·,·> represents the inner product operation, ‖·‖ represents the L2 norm; Construct a gated linear unit g, design a gating mechanism, dynamically adjust the fusion ratio of temporal features and spatial features according to the feature correlation, and introduce a residual connection structure to generate spatiotemporal joint features: F joint =g⊙tanh(W f [F temp ‖F space ])+(1-g)⊙(W t F temp +W s F space Among them, F joint is the spatiotemporal joint feature, g = sigmoid (W g [F temp ‖F space ]), represents the gated linear unit calculation, ∥ represents the feature concatenation operation, ⊙ represents the element-by-element multiplication, W f , W t , W s and W g is the learnable parameter matrix, sigmoid and tanh are the corresponding function operations.

9. The intelligent classification method of Beidou positioning data as claimed in claim 5, characterized in that: Step S6 includes: S61, using a fully connected layer to perform an initial transformation on the spatiotemporal joint features, and then performing feature normalization processing through layer normalization, where the layer normalization includes calculating the feature mean and variance, and performing normalization transformation using a learnable scaling factor and bias parameter; S62, using the self-attention mechanism to perform global context modeling on the standardized features, specifically generating a query vector, a key vector, and a value vector through linear transformation, calculating the attention weight score, and performing a residual connection between the attention-weighted features and the input features; S63. Use a two-layer feedforward neural network to perform nonlinear transformation on the features. The first layer uses the ReLU activation function for feature mapping, the second layer performs dimension conversion, and introduces residual connections after each layer of conversion to preserve the original information. S64. Map the features to the predefined category space through the fully connected layer, use the softmax function to calculate the probability distribution of each category, and output the classification result.

10. The intelligent classification method of Beidou positioning data according to claim 5, characterized in that: The spatial convolution layer in the spatial feature extraction module adopts a dilated convolution structure. The dilation rate increases layer by layer with the depth of the network. The dilation rate d l Set it up as follows: d l 2 l-1 ,l=1,2,...,L Where L is the total number of spatial convolutional layers.

Images