Cascade learning type multi-source fusion positioning method and system under satellite denial

Abstract

The invention discloses a cascade learning type multi-source fusion positioning method and system under satellite denial. The method comprises the following steps: firstly, constructing a system sensor mathematical model which comprises an INS error model, a WSS / INS integrated navigation model and a GPS / INS integrated navigation model; then designing a neural network model based on a cascade learning CL algorithm; when the satellite is available, training a neural network model based on a cascade learning CL algorithm; when the satellite is denied, the neural network model which is completely trained and is based on the cascade learning CL algorithm is utilized to predict the pseudo GPS position, and the GPS blank is filled up. The system comprises a sensor mathematical model building module, a CL algorithm model building module, a training module and a positioning prediction module. The method improves the positioning precision of the vehicle in the GPS failure period, is excellent in performance especially in a complex environment and under long-time interruption, reduces the positioning error, improves the overall performance, and has important application value for vehicle positioning in the technical fields of intelligent transportation and automatic driving.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle navigation in satellite denial scenarios, and in particular to a cascade learning multi-source fusion positioning method and system under satellite denial. Background Art

[0002] Amid the booming development of modern intelligent transportation systems and autonomous driving technologies, the accuracy and reliability of vehicle positioning have become crucial technical components. Vehicle positioning technology is primarily used to determine a vehicle's precise location while driving, providing critical data support for navigation, route planning, safety control, and other aspects. Currently, the most widely used vehicle positioning system is a combination of the Global Positioning System (GPS) and the Inertial Navigation System (INS). Examples include invention patent CN202210108243.1, which discloses a positioning method and device for an in-vehicle integrated navigation system, and invention patent CN202311532333.4, which discloses an in-vehicle fusion navigation method based on inertial satellite odometry and kinematic constraints. While GPS can provide high-precision global location information, it has significant limitations. In challenging environments such as urban streets with tall buildings, tunnels, and underground parking lots, GPS signals are susceptible to obstruction, reflection, and interference, resulting in poor signal reception or even loss. When the GPS signal fails, the INS can independently provide vehicle location information in a short period of time. However, INS uses an integral calculation to estimate position, and its positioning error accumulates over time.

[0003] In order to overcome the positioning problem caused by GPS failure, researchers have made many attempts. A common method is to add additional sensors. For example, sensors such as wheel speed sensors, radars, cameras and ultra-wideband are used to compensate for the information missing from GPS, such as the laser-inertial-satellite multi-source fusion navigation method and device disclosed in the invention patent CN202410514710.X. However, this method has many disadvantages. First, adding additional sensors will lead to a significant increase in system costs. Secondly, there is the problem of spatiotemporal synchronization between different sensors, and it is very complicated to ensure the consistency of data from each sensor in the same time and space reference system. Moreover, in order to fuse these multi-source sensor data, complex algorithms need to be designed. In addition, the performance of some sensors, such as cameras, is easily affected by weather conditions (such as heavy rain, heavy snow, dense fog, etc.), and they cannot provide reliable data under severe weather conditions.

[0004] Zero-speed update technology is also a solution. This technology can suppress the cumulative error of INS and has a relatively low cost. For example, the invention patent CN201910924281.2 discloses an inertial navigation error correction method based on vehicle zero-speed detection. However, the traditional method of detecting the stationary state of a vehicle has major flaws. Due to the nonlinear relationship between INS data and the stationary state of the vehicle, it is difficult to set an accurate threshold to determine whether the vehicle is stationary, which leads to low detection accuracy and inability to effectively achieve zero-speed update, limiting the role of this technology in improving vehicle positioning accuracy.

[0005] In recent years, with the rise of artificial intelligence technology, data-driven neural network prediction models have been used to provide pseudo-GPS information. For example, invention patent CN202211549582.X discloses an intelligent autonomous navigation method under satellite denial, invention patent CN202211503193.3 discloses a vehicle layered multi-module fusion precise positioning method under satellite system denial, and invention patent CN202211053671.5 discloses a UAV multi-source navigation information processing system based on factor graphs. However, most of these existing methods have shortcomings. Most methods rely on neural networks to directly map INS data to GPS positions or increments, but due to the divergence of INS data, the long-term positioning accuracy of this method will be seriously affected over time. In addition, existing intelligent positioning methods often simplify the vehicle motion model and ignore the important link of vehicle parking and motion state classification, resulting in the inability to accurately locate the vehicle in practical applications.

[0006] In summary, existing technologies have certain limitations in solving the problem of vehicle positioning during GPS failure, and cannot meet the requirements of intelligent transportation and autonomous driving technologies for high-precision and high-reliability vehicle positioning. A new and more effective vehicle positioning method is urgently needed. Summary of the Invention

[0007] The purpose of the present invention is to provide a multi-source fusion positioning method and system with high positioning accuracy, which can effectively cope with complex environments and GPS failures and significantly improve the reliability and stability of vehicle positioning.

[0008] The technical solution for achieving the purpose of the present invention is: a cascade learning multi-source fusion positioning method under satellite denial, comprising the following steps:

[0009] Step 1: Construct the mathematical model of the system sensor, including the INS error model, the WSS / INS integrated navigation model and the GPS / INS integrated navigation model;

[0010] Step 2: Design a neural network model based on the cascade learning CL algorithm;

[0011] Step 3: When the satellite is available, train the neural network model based on the cascade learning CL algorithm;

[0012] Step 4: When satellites are denied, a well-trained neural network model based on the cascade learning CL algorithm is used to predict the pseudo-GPS position to fill the GPS gap.

[0013] A cascade learning multi-source fusion positioning system under satellite denial is used to implement the cascade learning multi-source fusion positioning method under satellite denial. The system includes a sensor mathematical model construction module, a CL algorithm model construction module, a training module, and a positioning prediction module, wherein:

[0014] The sensor mathematical model construction module is used to construct the sensor mathematical model including the inertial navigation system INS error model, the wheel speed sensor WSS / INS integrated navigation model and the global positioning system GPS / INS integrated navigation model;

[0015] The CL algorithm model building module is used to build a neural network based on the deep reinforcement learning algorithm;

[0016] The training module is used to train the neural network to obtain a trained neural network model;

[0017] The positioning prediction module is used to call the trained neural network model to output a pseudo GPS position to fill the gaps in GPS.

[0018] Compared with the existing technology, the significant advantages of the present invention are as follows: (1) Significantly improved positioning accuracy: Through a unique cascade learning model, the different motion states of the vehicle are accurately classified and targeted, effectively reducing positioning errors. During GPS failure, especially in complex environments and long-term interruptions, the positioning accuracy of the present invention far exceeds that of the existing technology, and can provide more reliable location information for vehicle navigation; (2) Strong ability to adapt to complex environments: The method based on multi-task cascade learning has stronger adaptability to the environment. Whether in urban high-rise buildings, tunnels and other areas where GPS signals are easily interfered with, or in complex driving conditions such as frequent parking and turning of vehicles, it can stably maintain a high positioning performance, ensure the continuity and accuracy of vehicle positioning, and provide strong guarantees for the safe driving of autonomous vehicles in complex urban traffic scenes; (3) Model design The advantages of the design are outstanding: the deep residual convolutional neural network DRCNN is used for vehicle stationary detection, which can better handle the nonlinear relationship between INS data and the vehicle's stationary state and improve the detection accuracy; the dual spatiotemporal adaptive neural network D-TSANN combines CNN, gated recurrent unit GRU and self-attention mechanism when predicting pseudo GPS position. Compared with a single-structure neural network, it can more effectively extract spatiotemporal features, enhance the understanding and prediction capabilities of vehicle motion patterns, and thus improve the overall positioning effect; (4) Good cost-effectiveness: There is no need for a large amount of additional hardware equipment investment. Relying on existing vehicle-mounted equipment such as INS, GPS and wheel speed sensor WSS, the positioning performance is improved through algorithm design, which reduces the system cost while improving the positioning accuracy. It has a high cost-effectiveness and is easier to promote and apply in actual vehicle positioning systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is a flow chart of a cascade learning multi-source fusion positioning method under satellite denial according to the present invention.

[0020] Figure 2 This is a flowchart of a vehicle stationary detection algorithm in a specific embodiment of the present invention.

[0021] Figure 3 This is a flowchart of a trajectory prediction algorithm based on TSANN in a specific embodiment of the present invention.

[0022] Figure 4 This is a physical picture of a sports car in an embodiment of the present invention.

[0023] Figure 5 This is a satellite trajectory diagram of the sports car experiment in an embodiment of the present invention.

[0024] Figure 6 2 is a comparison chart of the errors of various algorithms in the sports car experiment in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following combination Figures 1 to 6 The present invention is further described in detail with reference to the accompanying drawings and specific embodiments.

[0026] Combine Figure 1 The present invention provides a cascade learning multi-source fusion positioning method under satellite denial, comprising the following steps:

[0027] Step 1: Construct the mathematical model of the system sensor, including the INS error model, the WSS / INS integrated navigation model and the GPS / INS integrated navigation model;

[0028] Step 2: Design a neural network model based on the cascade learning CL algorithm;

[0029] Step 3: When the satellite is available, train the neural network model based on the cascade learning CL algorithm;

[0030] Step 4: When satellites are denied, a well-trained neural network model based on the cascade learning CL algorithm is used to predict the pseudo-GPS position to fill the GPS gap.

[0031] As a specific example, in step 1, a system sensor mathematical model is constructed, including an INS error model, a WSS / INS integrated navigation model, and a GPS / INS integrated navigation model, as follows:

[0032] Step 1.1: Construct the INS error model as follows:

[0033] The INS error model uses the geographic coordinate system of the northeast sky as the navigation coordinate system n, the local coordinate system on the right front as the carrier coordinate system b, the earth system e and the inertial system i. The INS velocity error equation is as follows:

[0034]

[0035] Where, f n and is the accelerometer vector and its bias vector in the n-frame; V n =[V e ,V n ,V u ] T and δV n =[δV e ,δV n ,δV u ] T are the velocity vector and its error vector in the n system respectively; and represents the Earth's rotation angular velocity vector and its error vector of the e system relative to the i system in the n system, and represents the rotation angular velocity vector and its error vector of the n-system relative to the e-system in the n-system; δg n represents the gravity error vector in the n-frame;

[0036] The INS attitude error equation is:

[0037]

[0038] Where, φ=[φ P ,φ R ,φ H ] T is the attitude angle error vector in the n system, φ P ,φ R ,φ H Represent the pitch angle error, roll angle error and heading angle error respectively; and is the angular velocity vector of system n relative to system i and its error vector, ε n is the gyroscope drift vector in the n coordinate system;

[0039] The INS position error equation is:

[0040]

[0041] Where R M and R N are the curvature radii in the meridian and principal vertical directions, L, λ and h represent latitude, longitude and altitude respectively, and δL, δλ and δh represent their corresponding errors.

[0042] Step 1.2: Construct the WSS / INS integrated navigation model and the GPS / INS integrated navigation model as follows:

[0043] Due to various noises in the inertial measurement unit, the independent INS system will quickly deviate from the true trajectory, so external measurement information is needed to control its error diffusion. The WSS / INS integrated navigation model and the GPS / INS integrated navigation model are constructed using the Kalman filter KF method. Both navigation models use a 15-dimensional state vector:

[0044]

[0045] Among them, ε bx ,ε by ,ε bz represents the gyroscope drift in frame b, represents the accelerometer bias in frame b;

[0046] The state equation of the KF system is:

[0047]

[0048] Among them, F k,k-1 and G k-1 Represent the state transfer matrix and system noise matrix respectively, W k-1 and V k They represent the process noise and measurement noise in the state transition process, respectively, and are both set to zero-mean Gaussian noise; Z k and H k are the measurement vector and measurement matrix, respectively. In the INS / GPS model, the measurement vector is the difference between the velocity and position measurements of INS and GPS; in the WSS / GPS model, the measurement vector is the difference between the northeast sky velocity measurements of INS and WSS, that is:

[0049] Z GPS / INS =[V INS -V GPS ,P INS -P GPS ] T (6)

[0050] Z WSS / INS =[V INS -V ENU ] T (7)

[0051] Where V INS and P INS Represent the speed and position values provided by INS, V GPS and P GPS They represent the velocity and position measurements provided by GPS, respectively. Since WSS only provides one-dimensional velocity V in the b system, WSS , in order to decompose it into the three-dimensional velocity V of the northeast sky in the n system ENU , coordinate transformation needs to be performed in combination with the attitude angle provided by INS:

[0052]

[0053] Among them, V E,WSS 、V N,WSS and V U,WSS Indicates V WSS Decompose the northeastern celestial velocity in the n-system, and Respectively represent the pitch angle and heading angle provided by INS, and the two models have different observation matrices:

[0054] H GPS / INS =[0 3×6 ,I 3×3 ,0 3×3 ] T (9)

[0055] H WSS / INS =[0 3×3 ,I 3×3 ,0 3×9 ] T (10)

[0056] Among them, I 3×3 represents the 3rd-order unit matrix, the others are zero matrices, and the subscript represents the matrix dimension;

[0057] The Kalman filter algorithm is divided into two steps: time update and measurement update:

[0058] (1) Time updating refers to using the state transition equation of the filter to predict the state of the system and update the prior estimate and the prior error covariance matrix:

[0059]

[0060] in, represents the prior estimate of the system state, P k,k-1 represents the prior error covariance matrix, P k-1 represents the covariance matrix of the posterior estimate at the previous moment, Q k-1 represents the process noise covariance matrix;

[0061] (2) Measurement update is the correction step in the Kalman filter algorithm, which is used to compare the new observation data with the predicted state and update the state estimate to optimize the estimate:

[0062]

[0063] P k =P k,k-1 -K k H k P k,k-1 (15)

[0064] Where K k represents the Kalman gain, R k represents the measurement noise covariance matrix, represents the a posteriori estimate of the system state.

[0065] As a specific example, in step 2, a neural network model based on the cascade learning CL algorithm is designed as follows:

[0066] Step 2.1: Use the deep residual convolutional neural network DRCNN to build a vehicle static detector. The angular velocity, specific force, northeast sky speed and heading angle are used as inputs. After network processing, they are converted into binary classification results of the vehicle static state through functions and thresholds. The binary cross entropy BCE algorithm is used as the training loss function, as shown in the following example: Figure 2 As shown, the details are as follows:

[0067] A robust parking detector can not only suppress INS divergence during GPS interruptions, but also correct the predicted values of the pseudo-GPS predictor, thereby improving the reliability of vehicle positioning and navigation. In existing research, researchers often use wheel speed signals and IMU data provided by wheel speed sensors as the core inputs of parking detection algorithms. Considering the "dead zone" phenomenon in wheel speed signals, that is, the sensor only outputs the effective vehicle speed when the vehicle reaches a certain speed, but cannot provide accurate speed data at low speeds, and the mutual coupling between the raw measurement data of the inertial navigation system, there is a strong nonlinear mapping relationship between the input features and the parking status detection results. Based on the traditional threshold classification method, there are problems such as difficult threshold setting and low detection accuracy. A classification model based on deep residual CNN is proposed, which has strong advantages in handling nonlinear problems and fast calculation.

[0068] Consider parking status detection as a binary classification problem P parC ={0,1},P parC Represents the result of parking state detection, 1 represents parking state, 0 represents moving state, and the input of CNN classification model at time t is They are angular rate, specific force, northeast velocity, and heading angle;

[0069] The deep residual CNN uses residual connections to fuse the outputs of two 1DConv+L2+BN+ReLU layers and one 1DConv+L2+BN+ReLU layer. The features are then flattened through a Flatten layer, and the classification results are obtained through two fully connected layers (FC). The L2 regularization coefficient is 0.001, and the softmax activation function is used at FC2 to output a score probability between [0, 1]. The number of neurons in FC1 and FC2 is 64 and 3, respectively.

[0070] Since the GPS frequency is lower than the WSS frequency, the length of the input feature vector of the CNN classification model at time t is set to m, that is, c(1),...,c(m), then the CNN network can be described as:

[0071]

[0072] Among them, P parS It is the probability of the CNN classification model outputting a score between (0,1), which needs to be calculated by the function f u And the threshold u converts it into a binary classification result of 0 or 1:

[0073]

[0074] The binary cross entropy BCE algorithm is used as the loss function of the CNN two-class model:

[0075]

[0076] Among them, P i Represents the label value, Represents the predicted value of the i-th sample.

[0077] Step 2.2: Build a steering detector based on the threshold method. Based on the heading angle change rate, the vehicle steering state detection is considered as a binary classification problem to determine whether the vehicle is turning or driving straight. The details are as follows:

[0078] When turning, the vehicle needs to adapt to different road conditions and environments, including curvature, slope, and road surface conditions. Furthermore, when turning, the vehicle is affected by lateral forces, which introduces more vibration and noise. This increases the noise in the output signals of each on-board sensor, leading to a more complex nonlinear mapping relationship between the original input and the position increment label value. Therefore, the vehicle motion is divided into two states: turning and straight-ahead, and vehicle turning detection is treated as a binary classification problem. The vehicle turning state is defined as 1, and the vehicle straight-ahead state is defined as 0, that is:

[0079] S tur ={0,1} (19)

[0080] Among them, S tur This is the result of the steering state detection;

[0081] The vehicle turning detector is modeled as follows:

[0082]

[0083]

[0084] Where, Indicates the heading angle provided by the WSS / INS integrated navigation system, in degrees; Indicates the absolute value of the difference between the heading angle at the current moment and the previous moment. represents the measured value of the vehicle's turning angle, t represents the sampling time of the heading angle value, τ represents the maximum steering angle within the vehicle sampling time interval, and η is the set turning detection threshold; the heading angle is set to north-east as positive, and the value range is 0 to 360 degrees.

[0085] Step 2.3: Use the dual spatiotemporal adaptive neural network D-TSANN to build a pseudo GPS position predictor. D-TSANN consists of two TSANNs with similar structures but different learning strategies, targeting turning and straight driving scenarios respectively. TSANN uses convolutional neural network CNN and gated recurrent unit GRU to extract spatiotemporal features, introduces self-attention mechanism to enhance the influence of important features, and outputs pseudo GPS position increment information through the fully connected layer. The loss function uses mean square error MSE, as shown in the following example: Figure 3 As shown, the details are as follows:

[0086] Step 2.3.1: When GPS is interrupted, the D-TSANN model is used to activate the corresponding TSANN model based on the vehicle's steering state detection results. Each TSANN module has an independent learning strategy for position estimation for steering or straight driving states.

[0087] The TSANN model shares the same input as the parking detector. It uses two 1DConv+L2+ReLU+MaxPooling modules to mine the spatial features of the input, and a GRU+dropout module to mine the temporal features of the input. It then uses the attention mechanism (SA) to learn the importance of each spatiotemporal feature and enhance the influence of important features. Finally, a fully connected layer outputs pseudo-GPS position increment information. The regularization coefficient of the L2 layer is set to 0.001, and the dropout rate of the dropout layer is set to 0.2.

[0088] Step 2.3.2: Because the original input contains multiple time steps and multiple feature vectors, using a CNN-GRU architecture can better extract spatiotemporal features and improve the model's nonlinear mapping capabilities compared to separate CNNs and GRUs. The GRU, a variant of the RNN, does not have the risk of vanishing and exploding gradients and can better handle long sequences. Compared to the LSTM, another RNN variant, the GRU has a more streamlined structure. It uses update gates instead of the input and forget gates in the LSTM, employing a simpler gating mechanism. Consequently, it requires fewer computational resources and converges faster.

[0089] h in the GRU module t-1 and h t Represent the hidden state of the previous moment and the hidden state of the current moment, x t Represents the input at the current moment, r t and z t They are the reset gate and update gate in GRU, Represents the estimated value of the hidden state, σ and tanh are activation functions that map linear combinations to nonlinear spaces;

[0090] Step 2.3.3: The self-attention mechanism interacts within the input sequence, capturing relationships between different positions in the sequence and weighting and integrating information from different positions based on these relationships. Applying this capability to time series forecasting and multivariate regression not only helps the model identify important information between multiple features but also helps it better understand the mapping relationship between multiple inputs and outputs.

[0091] The self-attention mechanism can be described as:

[0092]

[0093] Among them, the query space Q, key space K and value space V are input by SA through the mapping matrix W q 、W k and W v Linear transformation is obtained;

[0094] Then the dot product operation is performed on the query space and the transpose of the key space to calculate the similarity between Q and K, that is, the attention score, Used for scaling of attention scores to prevent gradient disappearance;

[0095] Then use the Softmax function to normalize the scores so that the sum of the scores is 1;

[0096] Finally, the attention score is used to perform weighted summation on V to obtain the attention output;

[0097] Step 2.3.4: Output pseudo GPS position increment information through the fully connected layer. The pseudo position increment prediction model TSANN can be described as:

[0098]

[0099] Step 2.3.5: Use MSE as the loss function of the regression model:

[0100]

[0101] As a specific example, in step 3, when the satellite is available, the neural network model based on the cascade learning CL algorithm is trained as follows:

[0102] Step 3.1, set the random seed;

[0103] Step 3.2: Initialize the neural network model hyperparameters, including the number of training rounds, loss function, optimization method, and learning rate.

[0104] Step 3.3: Based on the algorithm’s model hyperparameters, perform forward propagation and use gradient descent to update the parameters of the strategy neural network and the valuation neural network.

[0105] Step 3.4: Execute step 3.3 until the training of the neural network model is completed and save the trained neural network model.

[0106] Step 4: When satellites are denied, a well-trained neural network model based on the cascade learning (CL) algorithm is used to predict pseudo-GPS positions to fill the gaps in GPS positioning. The details are as follows:

[0107] The initial position P provided by GPS G0 The predicted pseudo GPS position increment ΔP clS Accumulate and get t e Pseudo GPS location information at this moment:

[0108]

[0109] The present invention also provides a cascade learning multi-source fusion positioning system under satellite denial, which is used to implement the cascade learning multi-source fusion positioning method under satellite denial. The system includes a sensor mathematical model construction module, a CL algorithm model construction module, a training module and a positioning prediction module, wherein:

[0110] The sensor mathematical model construction module is used to construct the sensor mathematical model including the inertial navigation system INS error model, the wheel speed sensor WSS / INS integrated navigation model and the global positioning system GPS / INS integrated navigation model;

[0111] The CL algorithm model building module is used to build a neural network based on the deep reinforcement learning algorithm;

[0112] The training module is used to train the neural network to obtain a trained neural network model;

[0113] The positioning prediction module is used to call the trained neural network model to output a pseudo GPS position to fill the gaps in GPS.

[0114] Example

[0115] This example uses a sports car experiment to evaluate the positioning accuracy of the proposed method during GPS outages. The vehicle is equipped with relevant sensors to obtain actual vehicle operating data. The vehicle positioning data obtained by this method is compared with reference data provided by a high-precision integrated navigation system to verify the positioning accuracy and reliability of this method in practical applications.

[0116] In order to evaluate the positioning accuracy of the proposed method when GPS is interrupted, this embodiment conducted a real road vehicle experiment at Nanjing University of Science and Technology, China on February 27, 2024. The test vehicle was equipped with a MEMS-IMU stim300, a GNSS receiver and dual antennas, and a WSS. Figure 4 The sensor performance parameters are shown in Table 1. The reference data comes from the IPMV produced by China Hi-Tech Corporation, which is a GNSS / INS integrated navigation system that can provide centimeter-level positioning accuracy.

[0117] Table 1

[0118]

[0119] The sports car experiment will start at 2:30 PM on February 27, 2024, and end at 3:33 PM on February 27, 2024, for a total of 1 hour and 3 minutes. The test path is as follows: Figure 5 As shown, there are no natural interruptions in the GPS trajectory. The first 2500 seconds are training data, namely the green road segments in the figure, and the last 1300 seconds are test data, namely the sky blue and red road segments in the figure. The GPS week second corresponding to 3:11:40 pm is 227518, at which time the cascade model enters the prediction phase. During the experiment, the GPS signal was relatively sufficient. Simulated GPS signal interruptions of 60 seconds, 120 seconds, 180 seconds, and 240 seconds were intentionally introduced on some test sections, as shown in Figure 1. Figure 5 As shown in Figure 2, road segments #1 and #2 are short-duration sections, with GPS interruptions lasting 60 seconds. Section #1 is a simple turning section, and section #2 is a straight-through stop section, with vehicles stopped for 33 seconds. Road segments #3, #4, and #5 are long-duration sections, with GPS interruptions lasting 120 seconds, 180 seconds, and 240 seconds, respectively. Section #3 is a simple turning section, section #4 is a complex turning section, and section #5 is a complex turning stop section, with vehicles stopped for a total of 50 seconds. Table 2 details the start and end times, duration, and road scene descriptions of the five GPS interruption sections.

[0120] Table 2

[0121]

[0122] In order to evaluate the effectiveness and advancement of the proposed algorithm on five GPS interruption sections, the following five methods were analyzed and compared.

[0123] 1)WSS / INS combination;

[0124] 2) GRU, using GRU to predict pseudo GPS position;

[0125] 3) TSANN, using TSANN network to predict pseudo GPS position;

[0126] 4) D-TSANN, which classifies turning and straight-ahead scenarios using a fixed threshold method and activates the corresponding TSANN to predict pseudo-GPS positions according to different scenarios;

[0127] 5) CL, the method proposed in this paper uses a parking detector, a turn detector and D-TSANN to predict the pseudo GPS position.

[0128] To further demonstrate the advanced nature of the proposed method, the proposed method was compared with the Multi-LSTM method and the MTL method on GPS interruption sections #3, #4, and #5. Because the Multi-LSTM method requires visual classification of vehicle motion scenes, the vehicle driving state classification method proposed in this invention was used as an approximate replacement. To verify the effectiveness of the proposed parking detection algorithm, the proposed deep residual CNN network was compared with the generalized likelihood ratio method.

[0129] The algorithm is implemented using TensorFlow 2.5 and Python 3.8. The maximum steering angle rate τ in formula (21) is set to 30 degrees per second, and the steering detection threshold η in formula (22) is set to 2.5 degrees per second.

[0130] Figure 6The vehicle positioning errors of the CL algorithm proposed in the present invention and the other four comparative algorithms are shown. When the vehicle enters a GPS interruption environment, WSS is used as an additional measurement value to correct the continuously accumulated errors of INS, so the positioning accuracy is not significantly reduced compared to when GPS is effective. The prediction accuracy of the deep learning algorithm for celestial position is significantly higher than that of the WSS / INS integrated system. This is because deep learning has a strong correction ability for errors with obvious trends. By comparing the error curves of the GRU and TSANN methods, the positioning error of the TSANN method is significantly smaller than that of the GRU method in most time periods, indicating that the TSANN method can effectively extract important information of the input features and better construct a nonlinear mapping relationship between the original input and the position increment. By comparing the error curves of the TSANN and D-TSANN methods, the positioning error of the D-TSANN method is significantly smaller, which shows that the modeling of vehicle turning and straight driving scenarios can better grasp the vehicle movement laws and improve the positioning accuracy and reliability of the vehicle in complex scenarios. Further comparison of the error curves of the D-TSANN and CL models shows that the CL method significantly improves positioning accuracy compared to the D-TSANN method, particularly in the parking sections of GPS-interrupted sections 2# and 5#. This demonstrates that the deep residual CNN can accurately detect the parking state of vehicles and effectively suppress the positioning errors of the deep learning model during parking periods. The CL algorithm also achieves higher positioning accuracy in non-parking sections, demonstrating that refining the motion scene can improve overall model performance.

[0131] Table 3 summarizes the positioning errors of five algorithms on five GPS-interrupted sections. We used two evaluation metrics: maximum error (ME) and root mean square error (RMSE). The results show that the TSANN and GRU methods achieve slightly higher positioning accuracy than the WSS / INS combination method. However, on sections 2# and 5#, both methods exhibit lower positioning accuracy, indicating that deep learning methods are more susceptible to vehicle stop events. D-TSANN achieves higher positioning accuracy on complex turning sections. For example, on the eastbound direction of section 4, D-TSANN's RMSE is 62.5% lower than that of GRU and 36.8% lower than that of TSANN. The CL method achieves higher positioning accuracy than the other four methods. On sections 4# and 5# with long GPS interruptions, the CL method's RMSE is reduced by 74.4%, 65.9%, 46.9%, and 36.8% compared to the WSS / INS combination, GRU, TSANN, and D-TSANN, respectively, by 54.6%, 55.6%, 44.8%, and 23.8%.

[0132] Table 3

[0133]

[0134] In summary, the cascade learning multi-source fusion positioning under satellite denial proposed in this invention overcomes many drawbacks of existing vehicle positioning technology in GPS failure scenarios. Through precise model construction and efficient algorithm flow, it can stably and accurately locate vehicles in complex environments and different driving conditions, providing solid and reliable technical support for the development of intelligent transportation and autonomous driving technologies.

[0135] The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A cascade learning multi-source fusion positioning method under satellite denial, characterized in that: The following steps are involved: Step 1: Construct the mathematical model of the system sensor, including the INS error model, the WSS / INS integrated navigation model and the GPS / INS integrated navigation model; Step 2: Design a neural network model based on the cascade learning CL algorithm; Step 3: When the satellite is available, train the neural network model based on the cascade learning CL algorithm; Step 4: When satellites are denied, a well-trained neural network model based on the cascade learning CL algorithm is used to predict the pseudo-GPS position to fill the GPS gap.

2. The cascade learning multi-source fusion positioning method under satellite denial according to claim 1, characterized in that: In step 1, the INS error model is constructed as follows: The INS error model uses the geographic coordinate system of the northeast sky as the navigation coordinate system n, the local coordinate system on the right front as the carrier coordinate system b, the earth system e and the inertial system i. The INS velocity error equation is: Where, f n and is the accelerometer vector and its bias vector in the n-frame; V n =[V e ,V n ,V u ] T and δV n =[δV e ,δV n ,δV u ] T are the velocity vector and its error vector in the n system respectively; and represents the Earth's rotation angular velocity vector and its error vector of the e system relative to the i system in the n system, and represents the rotation angular velocity vector and its error vector of the n-system relative to the e-system in the n-system; δg n represents the gravity error vector in the n-frame; The INS attitude error equation is: Where, φ=[φ P ,φ R ,φ H ] T is the attitude angle error vector in the n system, φ P ,φ R ,φ H Represent the pitch angle error, roll angle error and heading angle error respectively; and is the angular velocity vector of system n relative to system i and its error vector, ε n is the gyroscope drift vector in the n coordinate system; The INS position error equation is: Where R M and R N are the curvature radii in the meridian and principal vertical directions, L, λ and h represent latitude, longitude and altitude, respectively. δL, δλ and δh represent the corresponding errors of L, λ and h, respectively.

3. The cascade learning multi-source fusion positioning method under satellite denial according to claim 2, characterized in that: In step 1, the WSS / INS integrated navigation model and the GPS / INS integrated navigation model are constructed as follows: The WSS / INS integrated navigation model and the GPS / INS integrated navigation model are constructed by the Kalman filter KF method. Both navigation models use a 15-dimensional state vector: Among them, ε bx ,ε by ,ε bz represents the gyroscope drift in frame b, represents the accelerometer bias in frame b; The state equation of the KF system is: Among them, F k,k-1 and G k-1 Represent the state transfer matrix and system noise matrix respectively, W k-1 and V k They represent the process noise and measurement noise in the state transition process, respectively, and are both set to zero-mean Gaussian noise; Z k and H k are the measurement vector and measurement matrix, respectively. In the INS / GPS model, the measurement vector is the difference between the velocity and position measurements of INS and GPS; in the WSS / GPS model, the measurement vector is the difference between the northeast sky velocity measurements of INS and WSS, that is: With GPS / INS =[In INS -V GPS ,P INS -P GPS ] T (6) With WSS / INS =[In INS -V ENU ] T (7) Where V INS and P INS Represent the speed and position values provided by INS, V GPS and P GPS They represent the velocity and position measurements provided by GPS, respectively. Since WSS only provides one-dimensional velocity V in the b system, WSS , in order to V WSS Decomposed into the three-dimensional velocity V of the northeast sky in the n system ENU , coordinate transformation needs to be performed in combination with the attitude angle provided by INS: Among them, V E,WSS 、V N,WSS and V U,WSS Indicates V WSS Decompose the east, north and sky velocities in the n system, and Respectively represent the pitch angle and heading angle provided by INS, and the two models have different observation matrices: H GPS / INS =[0 3×6 ,I 3×3 ,0 3×3 ] T (9) H WSS / INS =[0 3×3 ,I 3×3 ,0 3×9 ] T (10) Among them, I 3×3 represents the 3rd-order unit matrix, the others are zero matrices, and the subscript represents the matrix dimension; The Kalman filter algorithm is divided into two steps: time update and measurement update: (1) Time updating refers to using the state transition equation of the filter to predict the state of the system and update the prior estimate and the prior error covariance matrix: in, represents the prior estimate of the system state, P k,k-1 represents the prior error covariance matrix, P k-1 represents the covariance matrix of the posterior estimate at the previous moment, Q k-1 represents the process noise covariance matrix; (2) Measurement update is the correction step in the Kalman filter algorithm, which is used to compare the new observation data with the predicted state and update the state estimate to optimize the estimate: P k =P k,k-1 -K k H k P k,k-1 (15) Where K k represents the Kalman gain, R k represents the measurement noise covariance matrix, represents the a posteriori estimate of the system state.

4. The cascade learning multi-source fusion positioning method under satellite denial according to claim 3, characterized in that: In step 2, a neural network model based on the cascade learning CL algorithm is designed as follows: Step 2.1: Use a deep residual convolutional neural network (DRCNN) to construct a vehicle stationary detector. The angular velocity, specific force, northeast celestial velocity, and heading angle are used as inputs. After network processing, the detector is converted into a binary classification result of the vehicle's stationary state using a function and a threshold. The binary cross entropy (BCE) algorithm is used as the training loss function. Step 2.2: Construct a steering detector based on the threshold method. Based on the heading angle change rate, the vehicle steering state detection is regarded as a binary classification problem to determine whether the vehicle is turning or driving straight. Step 2.3: Use the dual spatiotemporal adaptive neural network D-TSANN to construct a pseudo GPS position predictor. D-TSANN consists of two TSANNs with similar structures but different learning strategies, targeting turning and straight-line driving scenarios respectively. TSANN uses convolutional neural network (CNN) and gated recurrent unit (GRU) to extract spatiotemporal features, introduces a self-attention mechanism to enhance the influence of important features, and outputs pseudo GPS position increment information through a fully connected layer. The loss function uses mean square error (MSE).

5. The cascade learning multi-source fusion positioning method under satellite denial according to claim 4, characterized in that: The vehicle stationary detector is constructed using the deep residual convolutional neural network (DRCNN) described in step 2.

1. Angular velocity, specific force, northeast celestial velocity, and heading angle are used as inputs. After network processing, they are converted into binary classification results of the vehicle's stationary state through functions and thresholds. The binary cross entropy (BCE) algorithm is used as the training loss function, as follows: The parking state detection is regarded as a two-class problem P parC ={0,1},P parC Represents the result of parking state detection, 1 represents parking state, 0 represents moving state, and the input of CNN classification model at time t is They are angular rate, specific force, northeast celestial velocity, and heading angle; The deep residual CNN uses residual connections to fuse the outputs of two 1DConv+L2+BN+ReLU layers and one 1DConv+L2+BN+ReLU layer. The features are then flattened through a Flatten layer, and the classification results are obtained through two fully connected layers (FC). The L2 regularization coefficient is 0.001, and the softmax activation function is used at FC2 to output a score probability between [0, 1]. The number of neurons in FC1 and FC2 is 64 and 3, respectively. Since the GPS frequency is lower than the WSS frequency, the length of the input feature vector of the CNN classification model at time t is set to m, that is, c(1),...,c(m), then the CNN network is described as: Among them, P parS It is the probability of the CNN classification model outputting a score between (0,1), which needs to be calculated by the function f u And the threshold u converts it into a binary classification result of 0 or 1: The binary cross entropy BCE algorithm is used as the loss function of the CNN two-class model: Among them, P i Represents the tag value, Represents the predicted value of the i-th sample.

6. The cascade learning multi-source fusion positioning method under satellite denial according to claim 5, characterized in that: The threshold-based steering detector described in step 2.2 is constructed based on the heading angle change rate. The vehicle steering state detection is considered a binary classification problem to determine whether the vehicle is turning or driving straight. The specific details are as follows: The vehicle motion is divided into two states: turning and straight-ahead. The vehicle turning detection is regarded as a binary classification problem. The vehicle turning state is defined as 1 and the vehicle straight-ahead state is defined as 0, that is: S tur ＝{0,1} (19) Among them, S tur This is the result of the steering state detection; The vehicle turning detector is modeled as follows: Where, Indicates the heading angle provided by the WSS / INS integrated navigation system, in degrees; Indicates the absolute value of the difference between the heading angle at the current moment and the previous moment. represents the measured value of the vehicle's turning angle, t represents the sampling time of the heading angle value, τ represents the maximum steering angle within the vehicle sampling time interval, and η is the set turning detection threshold; the heading angle is set to north-east as positive, and the value range is 0 to 360 degrees.

7. The cascade learning multi-source fusion positioning method under satellite denial according to claim 6, characterized in that: As described in step 2.3, a dual spatiotemporal adaptive neural network (D-TSANN) is used to construct a pseudo GPS position predictor. D-TSANN consists of two TSANNs with similar structures but different learning strategies, targeting turning and straight-line driving scenarios respectively. TSANN uses a convolutional neural network (CNN) and a gated recurrent unit (GRU) to extract spatiotemporal features, introduces a self-attention mechanism to enhance the influence of important features, and outputs pseudo GPS position increment information through a fully connected layer. The loss function uses the mean square error (MSE), as follows: Step 2.3.1: When GPS is interrupted, the D-TSANN model is used to activate the corresponding TSANN model based on the vehicle's steering state detection results. Each TSANN module has an independent learning strategy for position estimation for steering or straight driving states. The TSANN model shares the same input as the parking detector. It uses two 1DConv+L2+ReLU+MaxPooling modules to mine the spatial features of the input, and a GRU+dropout module to mine the temporal features of the input. It then uses the attention mechanism (SA) to learn the importance of each spatiotemporal feature. Finally, it outputs pseudo GPS position increments through a fully connected layer. The regularization coefficient of the L2 layer is set to 0.001, and the dropout rate of the dropout layer is set to 0.

2. Step 2.3.2: Since the original input contains multiple time steps and multiple feature vectors, the CNN-GRU architecture is used to extract spatiotemporal features. t-1 and h t Represent the hidden state of the previous moment and the hidden state of the current moment, x t Represents the input at the current moment, r t and z t They are the reset gate and update gate in GRU, Represents the estimated value of the hidden state, σ and tanh are activation functions that map linear combinations to nonlinear spaces; Step 2.3.3: Introduce the self-attention mechanism to enhance the influence of important features, as follows: The self-attention mechanism is described as: Among them, the query space Q, key space K and value space V are input by SA through the mapping matrix W q 、W k and W v Linear transformation is obtained; Then the dot product operation is performed on the query space and the transpose of the key space to calculate the similarity between Q and K, that is, the attention score, Used for scaling of attention scores to prevent gradient disappearance; Then use the Softmax function to normalize the scores so that the sum of the scores is 1; Finally, the attention score is used to perform weighted summation on V to obtain the attention output; Step 2.3.4: Output pseudo GPS position increment information through the fully connected layer. The pseudo position increment prediction model TSANN can be described as: Step 2.3.5: Use MSE as the loss function of the regression model:

8. The cascade learning multi-source fusion positioning method under satellite denial according to claim 7, characterized in that: When the satellite described in step 3 is available, the neural network model based on the cascade learning CL algorithm is trained as follows: Step 3.1, set the random seed; Step 3.2: Initialize the neural network model hyperparameters, including the number of training rounds, loss function, optimization method, and learning rate. Step 3.3: Based on the algorithm’s model hyperparameters, perform forward propagation and use gradient descent to update the parameters of the strategy neural network and the valuation neural network. Step 3.4: Execute step 3.3 until the training of the neural network model is completed and save the trained neural network model.

9. The cascade learning multi-source fusion positioning method under satellite denial according to claim 8, characterized in that: During satellite denial described in step 4, a well-trained neural network model based on the cascade learning CL algorithm is used to predict the pseudo-GPS position to fill the GPS gap, as follows: The initial position P provided by GPS G0 The predicted pseudo GPS position increment ΔP clS Accumulate and get t e Pseudo GPS location information at this moment:

10. A cascade learning multi-source fusion positioning system under satellite denial, characterized in that: The system is used to implement the cascade learning multi-source fusion positioning method under satellite denial according to any one of claims 1 to 9. The system includes a sensor mathematical model construction module, a CL algorithm model construction module, a training module and a positioning prediction module, wherein: The sensor mathematical model construction module is used to construct the sensor mathematical model including the inertial navigation system INS error model, the wheel speed sensor WSS / INS integrated navigation model and the global positioning system GPS / INS integrated navigation model; The CL algorithm model building module is used to build a neural network based on the deep reinforcement learning algorithm; The training module is used to train the neural network to obtain a trained neural network model; The positioning prediction module is used to call the trained neural network model to output a pseudo GPS position to fill the gaps in GPS.

Images