Lightweight intelligent multi-base-station sensing fusion positioning method

Abstract

The invention belongs to the field of wireless communication, and particularly relates to a lightweight intelligent multi-base-station sensing fusion positioning method. In order to improve the stability and accuracy of a multi-base-station system for positioning an unmanned aerial vehicle, the invention provides a two-stage screening neural network architecture on the basis that a plurality of base stations utilize a least square subspace rotation invariant technology to respectively estimate and obtain coordinates of the unmanned aerial vehicle. According to the architecture, a dynamic threshold grouping mechanism is used in a first stage, and differentiable base station pre-screening is realized through a straight-through estimator; and in the second stage, a learnable index correction term combining distance and signal-to-noise ratio physical characteristics is designed, a multi-objective loss function is constructed, and the positioning precision and the physical consistency are synchronously optimized. According to the scheme, the realization of millisecond-level real-time positioning in embedded equipment is supported while the lightweight characteristic is maintained, and the navigation reliability of the unmanned aerial vehicle in a complex electromagnetic environment is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of Integrated Sensing And Communications (ISAC) in wireless communications, and specifically relates to a lightweight intelligent multi-base station interawareness fusion positioning method. Background Art

[0002] As the most important carrier in the low-altitude economy, Unmanned Aerial Vehicle (UAV) has the advantages of flexibility, low cost, high precision and strong environmental adaptability. Its unmanned and intelligent technical characteristics have significantly improved operational efficiency and safety. It is reshaping the traditional industrial form and opening up new application scenarios.

[0003] The integrated communication and perception technology can simultaneously utilize communication signals and environmental perception information to extract target feature information, more accurately identify the type, status and location of drones, and achieve high-precision positioning, tracking and identification of drones. Through the integration of multi-dimensional perception data, drones can reduce their own weight, increase their range, and complete their tasks more autonomously.

[0004] In the application of synaesthesia in low-altitude environments, rapid response and accurate positioning are required to assist drone target tracking and real-time obstacle avoidance. Among the current mainstream positioning algorithms, the single-base station system uses algorithms such as Direction of Arrival (DOA) estimation to achieve positioning. Its system structure is relatively simple and has low requirements for time synchronization, but the positioning error is large. The traditional multi-base station positioning system can achieve higher-precision positioning by using methods such as Time Difference of Arrival (TDOA) estimation algorithm and multi-level fusion of multiple base station estimation results. However, these methods lack adaptability and stability in practical applications, and it is difficult to identify and eliminate erroneous positioning information, which may lead to a surge in positioning errors and fail to meet actual needs.

[0005] Compared with traditional methods, multi-base station collaborative sensing assisted by artificial intelligence (AI) can be more flexibly applied to complex practical scenarios. The reason is that through technologies such as deep learning, the system can extract richer features from synaesthesia signals, thereby improving the accuracy and reliability of target detection. In addition, AI algorithms can more accurately identify and distinguish real targets, interference targets or error signals, thereby reducing the probability of false alarms and missed detections. Since the input dimension of multi-base station symbol-level fusion is limited, lightweight neural network models can be better applied to this scenario. As a basic neural network model, multilayer perceptron (MLP) has a small number of parameters and low computational complexity, and can quickly complete data reasoning. When the input data dimension is small, it can effectively learn the features in the input data and make predictions. Compared with complex deep learning models, its reasoning time is shorter and can meet the millisecond-level real-time processing requirements. However, the performance of a single MLP depends largely on the selection of initial weights, and different initial values ​​may lead to different results. In addition, the training process may also encounter problems such as gradient disappearance or gradient explosion, affecting the convergence of the model. Summary of the invention

[0006] The purpose of this invention is to propose a lightweight intelligent multi-base station interaceptive fusion positioning method to improve the stability and accuracy of multi-base station system positioning of drones in response to the above limitations. While maintaining the lightweight characteristics, this solution reduces the positioning error in urban canyon scenes, improves the accuracy of abnormal base station identification, and improves the reliability of drone positioning and tracking in complex electromagnetic environments.

[0007] The technical solution of the present invention is to use a two-stage screening method of a lightweight perceptron combined with an attention mechanism to fuse the estimation results of multiple base stations to complete positioning. Figure 1 As shown, the synaesthesia integrated positioning system includes 1 drone and N communication and perception integrated base stations. Each base station independently transmits a signal based on an orthogonal frequency division multiplexing waveform as a synaesthesia signal, and receives a corresponding echo signal from the drone. The position of the base station is known; it is characterized in that the positioning method includes:

[0008] S1. Use the drone echo signal received by the base station to estimate the pitch angle and azimuth angle of the drone, and then obtain the coordinates of the drone, specifically:

[0009] The μth symbol expression on the mth subcarrier of the drone echo signal received by the nth base station is defined as:

[0010]

[0011] Among them, Rn is the distance of the UAV in the direction corresponding to the nth base station, U n is the amplitude attenuation, △f is the subcarrier spacing, c is the speed of light, η is the noise, the noise has a mean of 0 and a variance of σ 2 Gaussian distribution, (m,μ) represents the μth symbol on the mth subcarrier;

[0012] Assume that the base station is equipped with an XY-axis L-shaped antenna array, the antenna spacing is half the wavelength, the number of antennas on each axis is M, and define θ n and represents the angles between the incident direction of the drone signal and the X-axis and Z-axis of the nth base station, respectively. The antenna on the X-axis of the nth base station is divided into two parallel sub-arrays. The first sub-array X 1 Contains antennas from the 1st to the M-1th, the second subarray X 2 Including from the 2nd to the Mth antenna; the steering vector of the X axis of the nth base station is:

[0013]

[0014] X 1 and X 2 The received echo signals are:

[0015] R x1,r,n =a x,n s r,n +N x1,r,n

[0016] R x2,r,n =a x,n Φ x,n s r,n +N x2,r,n

[0017] in, s r,n is the received signal C n The rth row of x1,r,n and N x2,r,n Represents the subarray X 1 and X 2 Additive Gaussian white noise on ;

[0018] Using the least squares space rotation invariant algorithm, we can get R x1,r,n , R x2,r,n Find Φ x,n ;

[0019] Similarly, the signal received by the Y-axis antenna is extracted to obtain

[0020] Joint Φ x,n With Φ y,n, get the estimated azimuth and the pitch angle estimate

[0021] The steering vector of the nth base station in the distance dimension is:

[0022]

[0023] Among them, d represents the distance dimension, N c Indicates the number of subcarriers of the signal;

[0024] X 1 and X 2 The echo signals in the distance dimension are expressed as:

[0025] R x1,d,r,n =a d,n s r,n +N x1,d,r,n

[0026] R x2,d,r,n =a d,n Φ d,n s r,n +N x2,d,r,n

[0027] Among them, Φ d,n =exp(-j2πΔf2R n / c), N x1,d,r,n and N x2,d,r,n Represents the subarray X 1 and X 2 Additive Gaussian white noise in the distance dimension;

[0028] R x1,d,r,n , R x2,d,r,n Substitute the least squares space rotation invariant algorithm to obtain Φ x,n , by Φ x,n Further calculate the estimated distance of the drone relative to each base station

[0029] By parameter Estimate the relative coordinates of the drone. Since the absolute position of the corresponding base station is known, each base station calculates a set of absolute coordinates of the drone, recorded as At the same time, the signal-to-noise ratio (SNR) of the channel between each base station and the drone is estimated. n ;

[0030] S2. Generate training data using the method of S1. Specifically, randomly generate intermediate obstacles between the drone and the base station, and form a vector with the parameters generated by each base station according to the method of S1. Then, the vectors of all base stations are combined into an N×5-dimensional vector as the input of the neural network, and the real drone coordinates corresponding to each set of vectors are recorded as labels; multiple sets of neural network input vectors are generated to form training data;

[0031] S3, build a fusion positioning network, the network includes a data input module, a base station group screening module, a group fine weighting module and a coordinate output module; the network processes the data as follows: the data input module processes the coordinate components of the input N×5-dimensional vector Normal distribution is used for standardization, and the distance After logarithmic transformation and normalization, the parameter vector of the nth base station is converted to

[0032] The normalized data is input into the base station group screening module. In the base station group screening module, the first fully connected layer uses the ReLU activation function to obtain the feature projection h n =ReLU(W 1 v n +b 1 ), L 1 represents the dimension of the fully connected layer, W 1 It is an L 1 ×5-dimensional weight matrix, b 1 It is an L 1 The second connection layer performs quality scoring and obtains the score corresponding to the nth base station. Where W 2 1×L 1 The weight matrix of dimension b 2 is a one-dimensional bias vector; then the threshold is adaptively selected and the straight-through estimator is used to compare s n The gradient and threshold of s dynamically generate a 0 / 1 binary mask. n Base stations with values ​​less than the threshold are assigned 0, and obviously abnormal base stations are eliminated, and a base station group consisting of N′ normal base stations is initially selected;

[0033] The group-wide fine weighting module performs fine fusion of physical guidance for the screened base station group. Since the positioning error of each base station is positively correlated with the estimated distance and negatively correlated with the signal-to-noise ratio, a coding layer that combines coordinate coding and physical coding is used. The coordinate coding layer extracts the coordinates. Spatial characteristics Where W 3 YesL 3 ×3-dimensional weight matrix, W 4 YesL 4 ×L 3 The weight matrix of dimension b 3 YesL 3 dimensional bias vector; the physical coding layer extracts physical quantity features Concatenate the features of the two encoding layers to obtain the joint feature f n =[c n ;p n ], f n Input the attention generation layer to get the basic weight α n , by introducing a learnable exponential correction term Strengthen the physical law constraints, and then obtain the final weight α of the nth base station after normalization of the mask constraints. n ′; use weight α n ′ and the estimated values ​​of each base station Get the fused positioning coordinates

[0034] S4, using the training data obtained in S2 to train the fusion positioning network constructed in S3, to obtain a trained fusion positioning network;

[0035] S5. The newly acquired set of data Input into the trained fusion positioning network to obtain the positioning result.

[0036] The beneficial effect of the present invention is that while maintaining the lightweight characteristics, it supports millisecond-level real-time positioning in embedded devices and significantly improves the navigation reliability of drones in complex electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the synaesthesia integration system model of the present invention;

[0038] Figure 2 Schematic diagram of the two-stage cascade neural network architecture of the present invention. DETAILED DESCRIPTION

[0039] The present invention uses a two-stage screening method of a lightweight perceptron combined with an attention mechanism to fuse the estimation results of multiple base stations to complete positioning. Figure 1 As shown in the figure, the synaesthesia integrated positioning system includes 1 UAV and N communication and perception integrated base stations. Each base station independently transmits a signal based on an orthogonal frequency division multiplexing (OFDM) waveform as a synaesthesia signal, and receives the corresponding echo signal from the UAV to estimate the coordinates of the UAV. The estimated coordinates, distance, and signal-to-noise ratio are input into a lightweight neural network for training.

[0040] The present invention comprises the following steps:

[0041] S1. Divide the transmit signal matrix by the receive signal matrix to construct a channel waveform model. The expression of the response of the μth symbol on the mth subcarrier of the nth carrier of the OFDM echo is as follows:

[0042]

[0043] Among them, R n is the distance of the UAV in the direction corresponding to the nth base station, U n is the amplitude attenuation, △f is the subcarrier spacing, c is the speed of light, η is the noise, the noise has a mean of 0 and a variance of σ 2 Gaussian distribution, (m,μ) represents the μth symbol on the mth subcarrier;.

[0044] Elevation and azimuth angles are estimated. Each base station is equipped with an XY-axis L-shaped antenna array, with an antenna spacing of half a wavelength and the number of antennas on each axis being M, θ n and Respectively represent the angles between the incident direction and the X-axis and Z-axis of the nth base station. If the drone is above the base station, the pitch angle The value range is (0,π / 2), the azimuth angle θ n The value range is (-π,π). The antenna on the X-axis of the nth base station is divided into two parallel sub-arrays: sub-array X 1 Contains antennas from 1st to M-1th, and subarray X 2 Contains antennas from the 2nd to the Mth day. Therefore, the steering vector of the nth base station on the X axis is:

[0045]

[0046] Subarray X 1 and the subarray X 2 The received echo signals are

[0047] R x1,r,n =a x,n s r,n +N x1,r,n , (Formula 3)

[0048] R x2,r,n =a x,n Φ x,n s r,n +N x2,r,n , (Formula 4)

[0049] in s r,n is the received signal C n The rth row of x1,r,n and N x2,r,n Represents the subarray X 1and X 2 The additive Gaussian white noise on the Y axis can be obtained in the same way. 1 , Y 2 The received echo signal. Similarly, the steering vector of the nth base station in the distance dimension is:

[0050]

[0051] The two subarrays on the X-axis can obtain echo signals in the distance dimension. Using the Total-Least-Squares Estimating Signal Parameters via RotationalInvariance Techniques (TLS-ESPRIT), the distance of the drone relative to each base station can be obtained. Azimuth and pitch angle The relative coordinates of the UAV can be obtained from these parameters. By knowing the absolute position of the corresponding base station, each base station can calculate a set of absolute coordinates of the UAV, which can be recorded as At the same time, the signal-to-noise ratio (SNR) of the channel between each base station and the drone is estimated. n .

[0052] S2, randomly generate 0-2 intermediate obstacles, so that the coordinates of the drone estimated by each base station are randomly wrong. Each base station generates a vector The estimated vectors of all base stations are synthesized into an N×5-dimensional vector as the input of the neural network, and the real drone coordinates corresponding to each set of vectors are recorded to provide labels. Repeat S2 to generate a sufficient amount of training data.

[0053] S3, construct a two-stage cascade neural network architecture. Figure 2 As shown, the network structure of the present invention can be divided into four modules: data input, base station group screening, fine weighting within the group, and coordinate output. The data input module performs normalization preprocessing on the input N×5 dimensional vector, and the coordinate component Normal distribution is used for standardization, and the distance Logarithmic transformation is used for normalization. As the first stage, the base station group screening module uses a lightweight perceptron network and dynamic thresholds to quickly filter out obviously abnormal base stations: after inputting the coordinates, distance and signal-to-noise ratio features of each base station, the lightweight fully connected layer generates a preliminary quality score and adaptively selects the threshold, and uses the straight-through estimator (STE) to dynamically generate a 0 / 1 binary mask. The differentiable hard screening mechanism is used to quickly pre-select the base station group and eliminate obviously abnormal base stations (such as nodes with too low SNR or extremely poor geometric consistency). From practical experience, it can be seen that in general, the positioning error of each base station is positively correlated with the estimated distance and negatively correlated with the signal-to-noise ratio. Therefore, as the second stage, the intra-group fine weighting module needs to perform physically guided fine fusion for the effective base station group: design a hybrid coding layer to extract coordinate space features (MLP encoding geometric relationship) and physical quantity features (joint features of distance and signal-to-noise ratio) respectively. After splicing the features of the two coding layers, the basic weights are calculated through the attention generation layer, and a learnable exponential correction term is introduced. Strengthen the physical law constraints, and finally normalize the mask constraints to force the weights of abnormal base stations to zero. The coordinate output module outputs the final weights.

[0054] S4. Input the generated training data into the neural network. According to the size of the mean square error of the verification coordinates, adjust the parameters of each layer of the network to obtain the optimal model.

[0055] S5. The newly acquired set of data Input into the trained fusion positioning network to obtain the positioning result.

[0056] Use Figure 1 The simulation parameters of the synaesthesia integrated system model shown are as follows: the OFDM carrier frequency is 24 GHz, the frequency interval is 240 KHz, the number of carriers is 128, the number of symbols is 256, the planar antenna array is 4 × 4, the number of synaesthesia integrated base stations is N = 4, the signal-to-noise ratio range of the channel is set to (-10, 20) dB, and the drone is located above all base stations. Assuming that 0-2 sets of estimated coordinates fail randomly in each trial, the simulation generates 10,000 sets of N × 5-dimensional input vectors.

[0057] Reasonable use of practical experience and physical laws can achieve better training effects. Therefore, the present invention constructs the following Figure 2 The two-stage cascade neural network architecture is shown. 8000 sets of data generated by simulation are put into training. During the training process, some data are randomly discarded to prevent overfitting, and 2000 sets of data are used for verification.

[0058] The root mean square error of the verification positioning To judge the training effect, according to the size of the mean square error of the verification coordinates, adjust the parameters of each layer of the network to obtain the optimal model.

[0059] This architecture innovatively decouples hard screening from soft weighting, can achieve millisecond-level real-time reasoning on general devices, and supports expansion of different base station numbers. The actual deployment model of the network is small, which meets the lightweight requirements of practical applications and can be used on more embedded edge devices. And because of the simple architecture, the hierarchy can be flexibly adjusted in different environments, with good adaptability.

Claims

1. A lightweight intelligent multi-base station synaesthesia fusion positioning method, defining a synaesthesia integrated positioning system including 1 drone and N communication perception integrated base stations, each base station independently transmits a signal based on an orthogonal frequency division multiplexing waveform as a synaesthesia signal, and receives a corresponding echo signal from the drone, and the position of the base station is known; it is characterized in that The positioning method comprises: S1. Use the drone echo signal received by the base station to estimate the pitch angle and azimuth angle of the drone, and then obtain the coordinates of the drone, specifically: The μth symbol expression on the mth subcarrier of the drone echo signal received by the nth base station is defined as: Among them, R n is the distance of the UAV in the direction corresponding to the nth base station, U n is the amplitude attenuation, △f is the subcarrier spacing, c is the speed of light, η is the noise, the noise has a mean of 0 and a variance of σ 2 Gaussian distribution, (m,μ) represents the μth symbol on the mth subcarrier; Assume that the base station is equipped with an XY-axis L-shaped antenna array, the antenna spacing is half the wavelength, the number of antennas on each axis is M, and define θ n and They represent the angles between the incident direction of the drone signal and the X-axis and Z-axis of the nth base station, respectively. The antenna on the X-axis of the nth base station is divided into two parallel subarrays. The first subarray X1 contains antennas from the 1st to the M-1th, and the second subarray X2 contains antennas from the 2nd to the Mth. The steering vector of the X-axis of the nth base station is: The echo signals received by X1 and X2 are: R x1,r,n =a x,n s r,n +N x1,r,n R x2,r,n =a x,n F x,n s r,n +N x2,r,n in, s r,n is the received signal C n The rth row of x1,r,n and N x2,r,n denote the additive white Gaussian noise on subarrays X1 and X2 respectively; Using the least squares space rotation invariant algorithm, we can get R x1,r,n , R x2,r,n Find Φ x,n ; Similarly, the signal received by the Y-axis antenna is extracted to obtain Joint Φ x,n With Φ y,n , get the estimated azimuth and the pitch angle estimate The steering vector of the nth base station in the distance dimension is: Among them, d represents the distance dimension, N c Indicates the number of subcarriers of the signal; The echo signals of X1 and X2 in the distance dimension are expressed as: R x1,d,r,n =a d,n s r,n +N x1,d,r,n R x2,d,r,n =a d,n F d,n s r,n +N x2,d,r,n Among them, Φ d,n =exp(-j2πΔf2R n / c), N x1,d,r,n and N x2,d,r,n Respectively represent the additive Gaussian white noise of subarrays X1 and X2 in the distance dimension; R x1,d,r,n , R x2,d,r,n Substitute the least squares space rotation invariant algorithm to obtain Φ x,n , by Φ x,n Further calculate the estimated distance of the drone relative to each base station By parameter Estimate the relative coordinates of the drone. Since the absolute position of the corresponding base station is known, each base station calculates a set of absolute coordinates of the drone, recorded as At the same time, the signal-to-noise ratio (SNR) of the channel between each base station and the drone is estimated. n ; S2. Generate training data using the method of S1. Specifically, randomly generate intermediate obstacles between the drone and the base station, and form a vector with the parameters generated by each base station according to the method of S1. Then, the vectors of all base stations are combined into an N×5-dimensional vector as the input of the neural network, and the real drone coordinates corresponding to each set of vectors are recorded as labels; multiple sets of neural network input vectors are generated to form training data; S3, build a fusion positioning network, the network includes a data input module, a base station group screening module, a group fine weighting module and a coordinate output module; the network processes the data as follows: the data input module processes the coordinate components of the input N×5-dimensional vector Normal distribution is used for standardization, and the distance After logarithmic transformation and normalization, the parameter vector of the nth base station is converted to The normalized data is input into the base station group screening module. In the base station group screening module, the first fully connected layer uses the ReLU activation function to obtain the feature projection h n =ReLU(W1v n +b1), L1 represents the dimension of the fully connected layer, W1 is an L1×5-dimensional weight matrix, and b1 is an L1-dimensional bias vector; the second connection layer performs quality scoring to obtain the score corresponding to the nth base station Where W2 is a 1×L1 dimensional weight matrix and b2 is a one-dimensional bias vector; then the threshold is adaptively selected and the straight-through estimator is used to compare s n The gradient and threshold of s dynamically generate a 0 / 1 binary mask. n Base stations with values ​​less than the threshold are assigned 0, and obviously abnormal base stations are eliminated, and a base station group consisting of N′ normal base stations is initially selected; The group-wide fine weighting module performs fine fusion of physical guidance for the screened base station group. Since the positioning error of each base station is positively correlated with the estimated distance and negatively correlated with the signal-to-noise ratio, a coding layer that combines coordinate coding and physical coding is used. The coordinate coding layer extracts the coordinates. Spatial characteristics Where W3 is the L3×3 dimensional weight matrix, W4 is the L4×L3 dimensional weight matrix, and b3 is the L3 dimensional bias vector; the physical coding layer extracts physical quantity features Concatenate the features of the two encoding layers to obtain the joint feature f n =[c n ;p n ], f n Input the attention generation layer to get the basic weight α n , by introducing a learnable exponential correction term Strengthen the physical law constraints, and then obtain the final weight α of the nth base station after normalization of the mask constraints. n ′; use weight α n ′ and the estimated values ​​of each base station Get the fused positioning coordinates S4, using the training data obtained in S2 to train the fusion positioning network constructed in S3, to obtain a trained fusion positioning network; S5. The newly acquired set of data Input into the trained fusion positioning network to obtain the positioning result.

Images