Vehicle cooperative positioning method based on multi-intelligent model fusion

Abstract

The invention discloses a vehicle cooperative positioning method based on multi-intelligent model fusion. The method comprises the following steps: acquiring a data set containing GPS observation data and UWB ranging values; randomly dividing the data set into a first data set and a second data set; constructing an initial deep learning model by using the first data set; performing supervised learning on the initial deep learning model by using the second data set to obtain a deep learning model, the deep learning model being used for predicting a GPS observation value; and performing GPS / UWB tight combination vehicle cooperative positioning by using the GPS observation value. A deep learning prediction model based on AOO, CNN, BiGRU and Attention mechanism fusion is designed and used for predicting GPS pseudo-range and Doppler frequency shift observation values in the satellite signal lock losing period, and the stability and reliability of vehicle positioning in a complex environment are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle collaborative positioning, and in particular relates to a vehicle collaborative positioning method based on multi-intelligent model fusion. Background Art

[0002] With the rapid development of autonomous driving, connected vehicles, and smart agriculture, relative positioning between vehicles has become an essential parameter for advanced applications in intelligent transportation systems. Due to complex environments such as urban canyons with tall buildings and densely forested roads, the Global Navigation Satellite System (GNSS) often struggles to meet safety standards for vehicle positioning in practical applications, nor can it fully leverage the advantages of vehicular ad-hoc networks (VANETs) for collaborative positioning. This situation not only affects positioning accuracy but, in severe cases, can even lead to positioning service failure, posing a significant challenge to driving safety. In collaborative positioning, GPS pseudorange and Doppler shift, as fundamental GPS observations, directly impact the accuracy and reliability of collaborative positioning. GPS pseudorange is a geometric distance measurement calculated by measuring the time difference between satellite signal transmission and reception. Doppler shift is caused by the relative radial motion of the satellite and receiver. By accurately estimating and predicting GPS pseudorange and Doppler shift during periods of satellite signal outages, users can be provided with continuous, stable, and high-precision collaborative positioning results.

[0003] However, the performance of existing vehicle collaborative positioning technologies is highly dependent on a favorable observation environment, such as an open sky and good satellite signal quality. In complex environments like urban canyons and tunnels, GPS signals are often blocked or interrupted, resulting in reduced or even ineffective positioning accuracy. This makes it impossible to accurately position vehicles in complex urban road environments. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a vehicle collaborative positioning method based on multi-intelligent model fusion, which improves the stability and reliability of vehicle positioning in complex environments.

[0005] To achieve the above objectives, the present invention provides a vehicle collaborative positioning method based on multi-intelligent model fusion, comprising:

[0006] Get a dataset containing GPS observation data and UWB ranging values;

[0007] randomly dividing the data set into a first data set and a second data set;

[0008] Building an initial deep learning model using the first dataset;

[0009] The initial deep learning model is supervisedly learned using the second data set to obtain a deep learning model, wherein the deep learning model is used to predict GPS observation values, and the GPS observation values ​​are used to perform GPS / UWB tightly combined vehicle collaborative positioning.

[0010] Optionally, the deep learning model includes: a CNN module, a BiGRU module, an attention mechanism module and an AOO optimization module;

[0011] The CNN module is used to capture local features between different time steps in GPS observation data and share the local features at different time points;

[0012] The BiGRU module is used to obtain the bidirectional dependency relationship of the time series data between the local features;

[0013] The attention mechanism module is used to obtain key features based on the bidirectional dependency relationship;

[0014] The AOO optimization module is used to optimize the hyperparameters of the initial deep learning model.

[0015] Optionally, the CNN module includes: an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer;

[0016] The input layer is used to input GPS observation data;

[0017] The convolution layer and the pooling layer are used to extract local features from the GPS observation data;

[0018] The fully connected layer is used to integrate the local features and map them to the sample label space;

[0019] The output layer is used to output the processed local features.

[0020] Optionally, the BiGRU module includes: an input layer, a forward hidden layer, a backward hidden layer and an output layer;

[0021] The input layer is used to input processed local features;

[0022] The forward hidden layer is used to process data in chronological order and capture information features from the past to the present;

[0023] The backward hidden layer is used to process data in reverse chronological order, model the future and the past, and supplement information not captured by the forward processing;

[0024] The output layer is used to output the bidirectional dependency relationship of time series data between local features.

[0025] Optionally, the key features for obtaining the key features according to the bidirectional dependency relationship include:

[0026] Calculating the attention distribution and the weight coefficient according to the bidirectional dependency relationship;

[0027] Obtaining the key features according to the calculation results;

[0028] Wherein, the weight coefficient is:

[0029]

[0030] Among them, α z is the weight coefficient, s(x z ,q) is the attention score function, Softmax is the normalized exponential function, x z is the zth input vector, q is the query vector, used to calculate the z similarity, n is the total number of input elements.

[0031] Optionally, the AOO optimization module is used to optimize the hyperparameters of the initial deep learning model, including:

[0032] Step 1: Set the model parameters of the initial deep learning model, wherein the model parameters include: population size, maximum number of iterations, problem dimension, current number of iterations, and a random number between 0 and 1;

[0033] Step 2: determine whether the model parameters meet the first constraint condition and obtain a first determination result;

[0034] Step 3: Perform biological parameter mapping on the first judgment result to obtain a mapping result;

[0035] Step 4: Determine whether the mapping result meets the second constraint condition and obtain a second determination result;

[0036] Step 5: Update the position according to the second judgment result to obtain the best position and suitable value;

[0037] Step 6: Determine whether the optimal position and the suitable value meet the first constraint condition, and repeat steps 2 to 5 until the hyperparameter optimization is completed.

[0038] Optionally, performing a position update according to the second judgment result to obtain an optimal position and a suitable value includes:

[0039] When the second judgment result meets the second constraint condition, the exploration phase position update is performed:

[0040]

[0041] Among them, UB is the upper bound of the target problem, X k+1 (i) is the position of the i-th individual in the g+1th generation; X best is the position of the best individual in the group, W is the dynamic weight factor, k is the current number of iterations, c is the dynamic adjustment factor, r dim is a random matrix, X k (t) is the position of the i-th individual in the g-th generation;

[0042] According to the update result, the optimal position and the suitable value are obtained;

[0043] When the second judgment result does not meet the second constraint condition, it is determined whether the second judgment result meets the third constraint condition, a third judgment result is obtained, and the optimal position and the suitable value are obtained according to the third judgment result.

[0044] Optionally, according to the third judgment result, obtaining the optimal position and the suitable value includes:

[0045] If the second judgment result meets the third constraint condition, updating the position by rolling caused by moisture absorption movement, and obtaining the optimal position and the suitable value according to the update result;

[0046] If the second judgment result does not meet the third constraint condition, a simplified projectile motion model is used to update the position, and the optimal position and fitness value are obtained according to the update result.

[0047] Optionally, performing GPS / UWB tightly integrated vehicle collaborative positioning using the GPS observation value includes:

[0048] Based on the GPS observation value, double difference processing is performed in combination with the vehicle's own pseudorange and Doppler frequency shift;

[0049] The processing results, UWB ranging values ​​and GPS observation values ​​are integrated to obtain the positioning results.

[0050] Optionally, performing double difference processing includes:

[0051]

[0052] in, is the double difference between the pseudorange observed by the receiver and the satellite at time t, u h (t),u p (t) is the unit line-of-sight vector from receiver a or b to satellite h and satellite p, respectively, r ab (t) is the distance vector between receivers a and b, is the residual of the uncorrelated errors.

[0053] Compared with the prior art, the present invention has the following advantages and technical effects:

[0054] (1) The present invention designs a deep learning prediction model based on the fusion of AOO, CNN, BiGRU and Attention mechanism to predict GPS pseudorange and Doppler shift observations during satellite signal loss, thereby improving the stability and reliability of vehicle positioning in complex environments.

[0055] (2) This paper uses CNN to extract spatiotemporal feature information from data sequences, and uses BiGRU's forward and backward propagation mechanisms to fully explore long-term dependencies in the data. It also introduces an attention mechanism to highlight key features in the data sequence. Furthermore, the AOO algorithm is used to optimize the model's hyperparameters to further improve the model's prediction accuracy.

[0056] (3) As traditional GPS positioning technology is difficult to meet the needs of intelligent transportation systems in complex environments, the new model in this invention uses historical GPS observation data to predict the observation values ​​during signal interruption to fill the positioning information gap.

[0057] (4) The present invention utilizes the characteristics of UWB technology that can provide high-precision ranging information and transmit information, and uses the predicted GPS pseudorange and Doppler frequency shift for GPS / UWB tightly combined vehicle collaborative positioning, providing users with continuous and stable collaborative positioning results.

[0058] (5) During the period when the GPS signal is lost, the relative positioning accuracy of the fusion model proposed in the present invention is improved by 50%, which significantly improves the stability and reliability of vehicle positioning in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0060] Figure 1 This is a flow chart of a vehicle collaborative positioning method based on multi-intelligent model fusion according to an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of a GPS / UWB tightly integrated vehicle collaborative positioning method according to an embodiment of the present invention;

[0062] Figure 3 It is a structural diagram based on the CNN module in an embodiment of the present invention;

[0063] Figure 4 Schematic diagram of the structure based on the BiGRU module in an embodiment of the present invention;

[0064] Figure 5 This is a structural diagram based on the Attention module in an embodiment of the present invention;

[0065] Figure 6 It is a structural diagram based on the AOO module in an embodiment of the present invention;

[0066] Figure 7 This is a schematic diagram of the structure of the AOO-CNN-BiGRU-Attention deep learning model in an embodiment of the present invention;

[0067] Figure 8 These are the results of the collaborative positioning relative position errors in the x, y, and z directions in an embodiment of the present invention. (a) is the result of the collaborative positioning relative position error in the x direction, (b) is the result of the collaborative positioning relative position error in the y direction, and (c) is the result of the collaborative positioning relative position error in the z direction. DETAILED DESCRIPTION

[0068] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0069] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0070] This embodiment proposes a vehicle collaborative positioning method based on multi-intelligent model fusion, such as Figure 1 As shown, the specific steps include:

[0071] Get a dataset containing GPS observation data and UWB ranging values;

[0072] Randomly divide the data set into a first data set and a second data set;

[0073] Building an initial deep learning model using the first dataset;

[0074] The second data set is used to perform supervised learning on the initial deep learning model to obtain a deep learning model, wherein the deep learning model is used to predict GPS observation values, and the GPS observation values ​​are used to perform GPS / UWB tightly combined vehicle collaborative positioning.

[0075] Specifically, S1. obtain preprocessed data to construct a data set, perform normalization processing on it, and divide it into a training set and a test set in proportion.

[0076] S2. The preprocessed data is input into the constructed CNN-BiGRU-Attention model. This model combines the advantages of CNN and BiGRU and introduces the attention mechanism to better capture feature information and improve model performance.

[0077] S3. Use the AOO algorithm for hyperparameter optimization. Use the AOO algorithm to optimize and update model hyperparameters such as convolution kernels, weights, convolution kernel channels, and fully connected layers.

[0078] S4. Taking advantage of the ultra-wideband technology's ability to provide high-precision ranging information and transmit information, the predicted GPS pseudorange and Doppler shift are used for GPS / UWB tightly coupled vehicle collaborative positioning.

[0079] More specifically, S1 divides the dataset into a training set and a test set. The training set is used to learn and train model parameters, while the test set is used to evaluate the model's prediction accuracy. Secondly, the data is normalized using MinMaxScaler, mapping the input and output features of the sample data to [-1, 1]. This eliminates the impact of data dimensionality and improves the model's convergence speed.

[0080] Furthermore, the deep learning model includes: CNN module, BiGRU module, attention mechanism module and AOO optimization module;

[0081] CNN module, which is used to capture local features between different time steps in GPS observation data and share local features at different time points;

[0082] BiGRU module, used to obtain the bidirectional dependency relationship of time series data between local features;

[0083] Attention mechanism module, used to obtain key features based on bidirectional dependencies;

[0084] AOO optimization module is used to optimize the hyperparameters of the initial deep learning model.

[0085] Furthermore, the CNN module includes: input layer, convolution layer, pooling layer, fully connected layer and output layer;

[0086] Input layer, used to input GPS observation data;

[0087] Convolutional layers and pooling layers are used to extract local features from GPS observation data;

[0088] The fully connected layer is used to integrate local features and map them to the sample label space;

[0089] The output layer is used to output the processed local features.

[0090] Specifically, CNN is used to capture local correlations between different time steps and share local features at different time points to enhance the model's ability to extract spatiotemporal information.

[0091] Furthermore, the BiGRU module includes: an input layer, a forward hidden layer, a backward hidden layer, and an output layer;

[0092] Input layer, used to input processed local features;

[0093] The forward hidden layer is used to process data in chronological order and capture information features from the past to the present;

[0094] The backward hidden layer is used to process data in reverse chronological order, modeling from the future to the past and supplementing the information not captured by the forward processing;

[0095] The output layer is used to output the bidirectional dependency relationship of time series data between local features.

[0096] Specifically, BiGRU can efficiently model time series data in two opposite directions, fully explore the long-term dependencies in the data, and thus capture more comprehensive feature information.

[0097] At time t, the hidden state output of BiGRU is determined by two independent GRUs: Among them, the GRU(·) function represents the nonlinear transformation of the input time data, x t For input, and Represent the forward hidden layer state output and the reverse hidden layer state output, W f and V f represents the weight, b t Represents the bias corresponding to the BiGRU hidden layer state at time t.

[0098] Furthermore, the key features used to obtain the following based on the bidirectional dependency include:

[0099] According to the bidirectional dependency, calculate the attention distribution and calculate the weight coefficient;

[0100] According to the calculation results, key features are obtained;

[0101] Among them, the weight coefficient is:

[0102]

[0103] Among them, α z is the weight coefficient, s(x z ,q) is the attention score function, Softmax is the normalized exponential function, x z is the zth input vector, q is the query vector, used to calculate thez similarity, z is the zth input element, and n is the total number of input elements.

[0104] Specifically, the constructed CNN-BiGRU-Attention model uses two 3×1 convolutional layers and a maximum pooling layer to extract features from GPS observation data. The ReLU function is selected as the activation function for the convolutional layer, and the maximum pooling function is used as the activation function for the pooling layer. After the feature vector is processed by the convolutional and pooling layers, it passes through a fully connected layer structure and is activated using the sigmoid function to obtain the final output vector of the CNN layer.

[0105] Furthermore, the AOO optimization module is used to optimize the hyperparameters of the initial deep learning model, including:

[0106] Step 1: Set the model parameters of the initial deep learning model, N is the population size, T max_iter is the maximum number of iterations, Dim is the problem dimension, k is the current number of iterations, and r is a random number between 0 and 1;

[0107] Step 2: determine whether the model parameters meet the first constraint condition and obtain a first determination result;

[0108] Step 3: Perform biological parameter mapping on the first judgment result to obtain a mapping result;

[0109] Step 4: Determine whether the mapping result meets the second constraint condition and obtain a second determination result;

[0110] Step 5: Update the position according to the second judgment result to obtain the best position and suitable value;

[0111] Step 6: Determine whether the optimal position and the appropriate value meet the first constraint condition, and repeat steps 2 to 5 until the hyperparameter optimization is completed.

[0112] Specifically, the AOO algorithm uses a two-stage exploration and exploitation approach to simulate the propagation process of animated oat seeds involving three propagation modes:

[0113] 1. Seeds are spread through the influence or action of wind, water or animals.

[0114] 2. In the absence of obstacles during dispersal, seeds change position only by rolling caused by hygroscopic movement.

[0115] 3. During the spreading process, when encountering an obstacle, this embodiment assumes that the main awn will suddenly eject under the drive of stored energy. The entire seed spreading process can be regarded as a projectile motion mainly driven by ejection.

[0116] The initial population size of the AOO algorithm is set to n=3, and the maximum number of iterations is set to T=5.

[0117] Further, updating the position according to the second judgment result to obtain the optimal position and the suitable value includes:

[0118] When the second judgment result meets the second constraint condition, the exploration phase position update is performed:

[0119]

[0120] Among them, UB is the upper bound of the target problem; X k+1 (i) is the position of the i-th individual in the g+1th generation; X best is the position of the best individual in the group, W is the dynamic weight factor, k is the current number of iterations, c is the dynamic adjustment factor, r dim is a random matrix, X k (t) is the position of the i-th individual in the g-th generation;

[0121] According to the update results, the best position and fitness value are obtained;

[0122] When the second judgment result does not meet the second constraint condition, it is determined whether the second judgment result meets the third constraint condition, the third judgment result is obtained, and the optimal position and the suitable value are obtained according to the third judgment result.

[0123] Furthermore, according to the third judgment result, obtaining the optimal position and the suitable value includes:

[0124] If the second judgment result meets the third constraint condition, the position is updated by rolling caused by the moisture absorption movement, and the optimal position and the suitable value are obtained according to the updated result;

[0125] If the second judgment result does not meet the third constraint condition, the simplified projectile motion model is used to update the position, and the optimal position and fitness value are obtained according to the update result.

[0126] Furthermore, using GPS observations to perform GPS / UWB tightly integrated vehicle collaborative positioning includes:

[0127] Based on GPS observations, double difference processing is performed in combination with the vehicle's own pseudorange and Doppler frequency shift;

[0128] The processing results, UWB ranging values ​​and GPS observation values ​​are integrated to obtain the positioning results.

[0129] Specifically, a GPS / UWB tightly integrated vehicle collaborative positioning model was constructed using GPS and UWB sensors. The experimental platform consisted of two vehicles equipped with different navigation sensors: Vehicle A and Vehicle B were each equipped with a low-cost Xsens MTI-G INS and a UWB sensor. Vehicle A was equipped with a Leica GS10 receiver, while Vehicle B was equipped with a Novatel INS-LCI (integrated GNSS-INS) device. The experiment utilized Real-Time Kinematic (RTK) technology to obtain high-precision position estimates for both vehicles, thereby obtaining a ground reference value. Leveraging the advantages of UWB, such as high-precision ranging information and wireless communication capabilities, combined with predicted GPS observation data to achieve tightly integrated collaborative positioning, users can be provided with continuous and stable collaborative positioning results.

[0130] The present embodiment will be described in detail below with reference to the accompanying drawings:

[0131] like Figure 1 As shown in the figure, GPS observation data (including pseudorange and Doppler shift) and UWB ranging values ​​required for vehicle collaborative localization are obtained. This data is collected in real time by onboard sensors and provides basic input for subsequent processing. An AOO-CNN-BiGRU-Attention deep learning prediction model is constructed. Finally, the predicted GPS pseudorange and Doppler shift are used for GPS / UWB tightly integrated vehicle collaborative localization.

[0132] like Figure 2 As shown in Figure 1, the core of this tightly coupled GPS / UWB vehicle collaborative positioning system utilizes pseudoranges and Doppler shifts acquired by the receiver, along with inter-vehicle distances measured by UWB sensors, to achieve collaborative positioning. Vehicle A acquires GPS observation data (pseudoranges and Doppler shifts, etc.) transmitted by vehicle B in real time via a UWB communication link and performs double-difference processing based on its own pseudoranges and Doppler shifts. The UWB measurement sensor then acquires the relative distance between the two vehicles. These observations are then fused using the Extended Kalman Filter (EKF) algorithm to produce the final positioning result.

[0133] Reference Figure 2 The method for GPS / UWB tightly integrated vehicle collaborative positioning includes the following steps:

[0134] GPS pseudorange double difference formula: in, is the double difference between the pseudorange observed by the receiver and the satellite at time t, u h (t),u p(t) is the unit line-of-sight vector from receiver a or b to satellite h and satellite p, respectively, r ab (t) is the distance vector between receivers a and b, is the residual of the uncorrelated errors.

[0135] The double difference of the Doppler frequency shifts of satellites h and p at receivers a and b at time t Obtained by pseudorange partial derivative, the expression is: Where, λ is the wavelength of GPS L1 signal; v ab (t) is the relative velocity vector between vehicles a and b; is the observation noise.

[0136] The system state equation is defined as: X(t+τ)=FX(t)+GN(t). Where X(t) is the state vector; F is the state transition matrix; G is the process noise vector; N(t) is the Gaussian relative acceleration noise, which has a Gaussian distribution with zero mean and standard deviation σ; and τ is the observation period.

[0137] The system observation equation is defined as: Z(t) = h(X(t)) + ζ(t). Where Z(t) is the observation vector, which includes the pseudorange double difference Doppler shift Relative distances between receivers a and b based on UWB measurements h is a nonlinear function of X(t); ζ(t) is the observation noise.

[0138] like Figure 3 As shown in Figure 1, CNN is a multi-layer perceptron that is used to capture local correlations between different time steps and share local features at different time points to enhance the model's ability to extract spatiotemporal information. CNN typically consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0139] The input layer serves as the entry point for the entire network. The convolutional layer includes the convolution computation layer and the activation layer. Activation functions typically include sigmoid, tanh, ReLU, and their variants. The primary purpose of the pooling layer is to compress data, reducing the number of neural network parameters and computational complexity while preventing overfitting. A common pooling method is max pooling; other pooling methods include average pooling, global pooling, universal pooling, and random pooling. The fully connected layer integrates the distributed features learned by the network and maps them to the sample label space.

[0140] like Figure 4 As shown in the figure, the BiGRU model includes an input layer, a forward hidden layer, a backward hidden layer, and an output layer. By introducing a bidirectional structure, it can better capture the bidirectional dependency relationship of sequence data.

[0141] like Figure 5 As shown in the figure, the Attention mechanism assigns different weights to feature vectors to focus on important features and highlight key features, thereby improving model performance and prediction capabilities.

[0142] Let X=[X1,X2,...,X n ] is the input information vector, given a query vector q related to the task, calculate the attention distribution of the input information, and its weight coefficient α i : Among them, α z is the weight coefficient, s(x z ,q) is the attention score function, Softmax is the normalized exponential function, x z is the zth input vector, q is the query vector, used to calculate the z similarity, z is the zth input element, and n is the total number of input elements.

[0143] like Figure 6 As shown, this embodiment introduces the AOO algorithm for hyperparameter optimization.

[0144] AOO starts with a set of random solutions.

[0145] Among them, x i Represents the individuals in the ith subgroup and their corresponding position information; Dim represents the dimension of the problem; N represents the population size. i,j The calculation formula is: i,j =r×(UB j -LB j )+LB j ,i=1,…,N,j=1,…,Dim. Among them, r is a random number between 0 and 1; UB j and LB j are the upper and lower bounds of the j-th dimension of the given problem.

[0146] Biological parameter mapping formula:

[0147] Where m is the seed mass; L is the main awn length; dim is the dimension of the target problem; P is the population size; e is the eccentric rotation coefficient; k is the current number of iterations; T is the maximum number of iterations; and c is the dynamic adjustment factor.

[0148] The position update formula in the exploration phase is as follows:

[0149]

[0150] Among them, UB is the upper bound of the target problem; X k+1(i) is the position of the i-th individual in the g+1th generation; X best is the position of the best individual in the group, W is the dynamic weight factor, c is the dynamic adjustment factor, r dim is a random matrix, X k (t) is the position of the i-th individual in the g-th generation.

[0151] At this stage, the remaining animated oat seeds will be divided into two propagation methods according to whether they encounter obstacles. In addition, this embodiment assumes that the probability of the two situations is equal.

[0152] In the absence of obstacles, the change in seed position is achieved by hygroscopic rolling driven by the stress gradient induced by moisture. This example uses the critical curvature of snap buckling to simulate this movement. The rolling mechanism is mathematically represented by the eccentric rotation and torque formula:

[0153]

[0154] Among them, r dim (-A, A) is a random matrix whose values ​​range from -A to A, and whose dimensions match the target problem. In Levy flight, u represents the desired position or mean, which is typically used to adjust the step size and is a random value between 0 and 1; σ is the scale parameter of the step size; v represents the current velocity vector; β is the parameter of the stable distribution; and Γ represents the gamma function, which is used to calculate the factorial value of non-integer parameters.

[0155] When a seed encounters an obstacle during its dispersal process, this example assumes that the main awn undergoes a projectile motion driven by stored energy. The entire seed dispersal process is characterized by projectile motion. To simulate this process, this example uses a simplified projectile motion model for position updates, and the position update formula is as follows:

[0156]

[0157] Among them, r dim (-B, B) is a random matrix whose values ​​range from -B to B, and its dimension matches the target problem. f is the elastic coefficient of the seed's main awn; x is the change in the length of the main awn during the projectile energy storage process; θ is the angle between the ejection trajectory and the ground; α is the air resistance coefficient during the projectile motion; r is a random number between 0 and 1; and r′ is a random number between 0 and T.

[0158] Figure 7 The construction of the AOO-CNN-BiGRU-Attention prediction model is given. The GPS observation dataset used in the experiment is divided into a training set and a test set. The training set is used for optimizing and learning the model parameters, while the test set is used to evaluate the prediction accuracy and generalization performance of the model.

[0159] Data normalization is a crucial preprocessing step in building machine learning models. We use the MinMaxScaler function to normalize the data, linearly mapping the input and output features of GPS pseudorange and Doppler shift data to the range [-1, 1]. This normalization effectively eliminates the impact of data dimensionality on the model and significantly improves model convergence speed.

[0160] After data preprocessing, the normalized data is input into the model for training and prediction. CNN is used to extract local features from the data, while the BiGRU network, through its forward and backpropagation mechanisms, effectively overlays feature information, fully exploiting the temporal and spatial dependencies in the data. Furthermore, an attention mechanism is introduced to selectively aggregate feature vectors by calculating the distribution of attention weights, thereby highlighting and retaining the most significant feature information. The performance of deep learning models depends heavily on the selection and adjustment of hyperparameters. During model training, the AOO algorithm is used to optimize and update model hyperparameters such as convolution kernels, weights, convolution kernel channels, and fully connected layers. Ultimately, the attention layer generates predicted GPS pseudorange and Doppler shift data.

[0161] The experimental platform consisted of two vehicles equipped with different navigation sensors. Vehicle A and Vehicle B were each equipped with a low-cost Xsens MTI-G INS and UWB sensor. Vehicle A was equipped with a Leica GS10 receiver, and Vehicle B was equipped with a Novatel INS-LCI (integrated GNSS-INS) device. The experiment utilized Real-Time Kinematic (RTK) technology to obtain high-precision position estimates for both vehicles, thereby obtaining a ground reference value. To ensure synchronization between data in the tightly coupled algorithm, the experiment used GPS for time synchronization. The experimental data was collected at a frequency of 1 Hz, with a sampling time of 14 minutes and 841 sampling points.

[0162] This example uses the AOO-CNN-BiGRU-Attention prediction model for error compensation. All experiments were performed on a 2.42GHz quad-core processor. Core TM The system is running on i5-1135G7@, Windows 10 64-bit version, and Matlab 2024a.

[0163] The first 589 seconds of data are selected as the training set for the neural network, and the satellite signal is assumed to be interrupted between 590 and 841 seconds. At this time, the proposed algorithm is used for error compensation.

[0164] Two 3×1 convolutional layers and a maximum pooling layer are used to extract features from GPS observation data. The ReLU function is selected as the activation function of the convolutional layer, and the maximum pooling function is selected as the activation function of the pooling layer. After the feature vector is processed by the convolutional layer and the pooling layer, it passes through the fully connected layer structure and is activated using the Sigmoid function. Finally, the final output vector of the CNN layer is obtained. Considering the training time, the initial population size of the AOO algorithm is set to n = 3, and the maximum number of iterations is set to T = 5. The relative position error curves in each direction of collaborative positioning are shown as follows: Figure 8 (a)-(c) represent.

[0165] This embodiment aims to solve the problem of how to accurately locate vehicles in complex urban road environments. This embodiment can predict the observation value during the signal interruption by using historical GPS observation data, fill the positioning information gap, and improve the stability and reliability of vehicle positioning in complex environments.

[0166] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A vehicle collaborative positioning method based on multi-intelligent model fusion, characterized in that: include: Get a dataset containing GPS observation data and UWB ranging values; randomly dividing the data set into a first data set and a second data set; Building an initial deep learning model using the first dataset; The initial deep learning model is supervisedly learned using the second data set to obtain a deep learning model, wherein the deep learning model is used to predict GPS observation values, and the GPS observation values ​​are used to perform GPS / UWB tightly combined vehicle collaborative positioning.

2. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 1 is characterized in that: The deep learning model includes: CNN module, BiGRU module, attention mechanism module and AOO optimization module; The CNN module is used to capture local features between different time steps in GPS observation data and share the local features at different time points; The BiGRU module is used to obtain the bidirectional dependency relationship of the time series data between the local features; The attention mechanism module is used to obtain key features based on the bidirectional dependency relationship; The AOO optimization module is used to optimize the hyperparameters of the initial deep learning model.

3. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 2 is characterized in that: The CNN module includes: input layer, convolution layer, pooling layer, fully connected layer and output layer; The input layer is used to input GPS observation data; The convolution layer and the pooling layer are used to extract local features from the GPS observation data; The fully connected layer is used to integrate the local features and map them to the sample label space; The output layer is used to output the processed local features.

4. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 2 is characterized in that: The BiGRU module includes: an input layer, a forward hidden layer, a backward hidden layer and an output layer; The input layer is used to input processed local features; The forward hidden layer is used to process data in chronological order and capture information features from the past to the present; The backward hidden layer is used to process data in reverse chronological order, model the future and the past, and supplement information not captured by the forward processing; The output layer is used to output the bidirectional dependency relationship of time series data between local features.

5. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 2 is characterized in that: The key features used to obtain the following according to the bidirectional dependency relationship include: Calculating the attention distribution and the weight coefficient according to the bidirectional dependency relationship; Obtaining the key features according to the calculation results; Wherein, the weight coefficient is: Among them, α z is the weight coefficient, s(x z ,q) is the attention score function, Softmax is the normalized exponential function, x z is the zth input vector, q is the query vector, used to calculate the z similarity, n is the total number of input elements.

6. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 2 is characterized in that: The AOO optimization module is used to optimize the hyperparameters of the initial deep learning model, including: Step 1: Set the model parameters of the initial deep learning model, wherein the model parameters include: population size, maximum number of iterations, problem dimension, current number of iterations, and a random number between 0 and 1; Step 2: determine whether the model parameters meet the first constraint condition and obtain a first determination result; Step 3: Perform biological parameter mapping on the first judgment result to obtain a mapping result; Step 4: Determine whether the mapping result meets the second constraint condition and obtain a second determination result; Step 5: Update the position according to the second judgment result to obtain the best position and suitable value; Step 6: Determine whether the optimal position and the suitable value meet the first constraint condition, and repeat steps 2 to 5 until the hyperparameter optimization is completed.

7. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 6 is characterized in that: Updating the position according to the second judgment result to obtain the optimal position and the suitable value includes: When the second judgment result meets the second constraint condition, the exploration phase position update is performed: Among them, UB is the upper bound of the target problem; X k+1 (i) is the position of the i-th individual in the g+1th generation; X best is the position of the best individual in the group, W is the dynamic weight factor, k is the current number of iterations, c is the dynamic adjustment factor, r dim is a random matrix, X k (t) is the position of the i-th individual in the g-th generation; According to the update result, the optimal position and the suitable value are obtained; When the second judgment result does not meet the second constraint condition, it is determined whether the second judgment result meets the third constraint condition, a third judgment result is obtained, and the optimal position and the suitable value are obtained according to the third judgment result.

8. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 7 is characterized in that: According to the third judgment result, obtaining the optimal position and the suitable value includes: If the second judgment result meets the third constraint condition, updating the position by rolling caused by moisture absorption movement, and obtaining the optimal position and the suitable value according to the update result; If the second judgment result does not meet the third constraint condition, a simplified projectile motion model is used to update the position, and the optimal position and fitness value are obtained according to the update result.

9. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 1 is characterized in that: Using the GPS observation value to perform GPS / UWB tightly integrated vehicle collaborative positioning includes: Based on the GPS observation value, double difference processing is performed in combination with the vehicle's own pseudorange and Doppler frequency shift; The processing results, UWB ranging values ​​and GPS observation values ​​are integrated to obtain the positioning results.

10. The vehicle collaborative positioning method based on multi-intelligent model fusion according to claim 9 is characterized in that: Double difference processing includes: in, is the double difference between the pseudorange observed by the receiver and the satellite at time t, u h (t),u p (t) is the unit line-of-sight vector from receiver a or b to satellite h and satellite p, respectively, r ab (t) is the distance vector between receivers a and b, is the residual of the uncorrelated errors.

Images