A method for GNSS-aided inertial navigation alignment in complex environments

Abstract

The application belongs to the technical field of inertial navigation, and particularly discloses a method for GNSS-aided inertial navigation alignment in a complex environment. In the case that GNSS observation conditions are poor, the application is used for realizing fast and stable alignment of INS in an unstable dynamic environment, so as to guarantee the real-time performance and reliability of subsequent high-precision positioning measurement. An initial alignment error prediction model of an inertial navigation system based on BiLSTM-Attention is built. The time-synchronized GNSS and INS measurement positions and velocity differences are extracted as feature vectors, and the feature vectors are input into the initial alignment error prediction model of the inertial navigation system. The model takes the minimization of the measurement deviation between GNSS and INS as the target, and through iterative optimization in the training process, the error variance and the loss function are converged to the predetermined accuracy range, and finally the INS three-axis misalignment angle estimation value and the three-axis velocity deviation estimation value are output, so as to realize high-precision calibration of the inertial navigation system in the motion state of the carrier.

Description

Technical Field

[0001] This invention belongs to the field of inertial navigation technology and relates to a method for GNSS-assisted inertial navigation alignment in complex environments. Background Technology

[0002] In complex urban environments, tall buildings block and reflect GNSS signals, leading to signal quality degradation and exacerbated multipath effects, making it difficult for the system to obtain stable and reliable positioning results. Although systems equipped with GNSS+INS can perform short-term positioning independently when there is no GNSS signal, the error rapidly increases to hundreds or even thousands of meters once the usable short-term range of the inertial navigation system is exceeded. Simultaneously, unstable GNSS signals make it difficult to correct the INS in time, further resulting in significant positioning errors and making it impossible to consistently provide high-precision positioning. Therefore, in complex urban environments, it is essential to address the positioning accuracy problem under poor GNSS signal quality. Initial alignment of the inertial navigation system and rapid calibration after receiving a valid GNSS signal again are particularly important. Initial alignment aims to obtain the initial attitude of the inertial navigation system, mainly including the setting of initial position and initial velocity, the initial alignment of the mathematical platform, and the drift measurement and calibration of the inertial instruments.

[0003] Currently, various filtering models are mainly used for error estimation in inertial navigation system (INS) initial alignment. Nonlinear filters such as Kalman filtering and unscented Kalman filtering are used for model construction. However, these algorithms suffer from high computational complexity, significant data computation difficulties, and challenges in suppressing filter divergence. Utilizing external sensors for auxiliary INS initial alignment further increases computational complexity, requiring more consideration of error elimination from these new sensors, and also increases system cost and integration complexity. Furthermore, the issue of GNSS signal obstruction in complex urban environments is not adequately considered, failing to effectively guarantee high-precision alignment of the INS under low GNSS signal quality conditions. This leads to large error accumulation due to insufficient INS correction by unstable GNSS, reducing the overall navigation positioning accuracy.

[0004] Currently, neural network learning models are gradually becoming an important direction for assisting integrated navigation research and improving positioning accuracy in complex environments. Among them, Long Short-Term Memory (LSTM) neural networks, as a deep learning model that is good at processing time-dependent data, can effectively capture the temporal characteristics of inertial navigation sensor data and adapt to the characteristics of nonlinear systems by virtue of their unique gating mechanism, providing a new path for solving the technical pain points of inertial navigation initial alignment.

[0005] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0006] To address the initial alignment and recalibration errors of inertial navigation systems (INS) in complex urban environments with unstable GNSS signals, this invention proposes a method for GNSS-assisted inertial navigation alignment in complex environments. This method constructs an INS initial alignment error prediction model based on BiLSTM-Attention. Under stable GNSS observation conditions, the model is continuously updated and trained using corresponding historical data. When GNSS observation conditions are unstable, the INS initial alignment error prediction model obtained from the previous stage is used to estimate the INS initial alignment error with high accuracy. This improves the INS reconvergence efficiency under unstable GNSS signal conditions, as well as the estimation efficiency and robustness of the INS initial alignment.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for GNSS-assisted inertial navigation alignment in complex environments includes the following steps: Step 1. Initial alignment data acquisition for the inertial navigation system; The initial alignment is performed on two parts: the correction of pitch angle, roll angle and heading angle of the moving vehicle. Step 2. Realignment using a GNSS-assisted inertial navigation system in complex environments; First, we build an inertial navigation initial alignment error prediction model based on BiLSTM-Attention. This inertial navigation initial alignment error prediction model includes a BiLSTM layer, an Attention layer, and a fully connected regression layer. The model's input is constructed from the difference between the position and velocity data measured by the GNSS and INS of the moving vehicle; the model's output is the inertial navigation alignment error vector, namely the estimated values ​​of the INS three-axis misalignment angle and the three-axis velocity deviation. BiLSTM is a bidirectional LSTM neural network that performs bidirectional time-series dependency modeling on the GNSS and INS difference input data. The forward LSTM network establishes the forward hidden states of the previous and current time steps. The inverse LSTM neural network establishes the backward hidden states of the next time step and the current time step. This yields the fused feature vector at each time step. ; Then, the attention layer is used to fuse the feature vectors output by the BiLSTM at each time step. Calculate the weights and the fused feature vector at each time step. Weighted fusion is performed to obtain the global vector; The weighted and fused global vector is input into a fully connected regression network to predict the inertial navigation alignment error vector. Under stable GNSS observation conditions, the inertial navigation initial alignment error prediction model is updated in real time, i.e. trained, based on historical data under these stable GNSS observation conditions, in order to obtain the model parameters of the inertial navigation initial alignment error prediction model. When GNSS is under unstable observation conditions, the model trained under the previous stable GNSS observation conditions is used to predict the inertial navigation alignment error vector in real time, and finally realize the realignment process of the inertial navigation.

[0008] Furthermore, based on the aforementioned method for GNSS-assisted inertial navigation alignment in complex environments, this invention also proposes a corresponding system for GNSS-assisted inertial navigation alignment in complex environments, the scheme of which is as follows: A GNSS-assisted inertial navigation alignment system for complex environments includes the following modules: The inertial navigation system coarse alignment module performs coarse alignment for two parts: pitch angle, roll angle correction, and yaw angle correction of the moving vehicle. The precision alignment and calibration module is used for realignment using a GNSS-assisted inertial navigation system in complex environments. First, we build an inertial navigation initial alignment error prediction model based on BiLSTM-Attention. This inertial navigation initial alignment error prediction model includes a BiLSTM layer, an Attention layer, and a fully connected regression layer. The model's input is constructed from the difference between the position and velocity data measured by the GNSS and INS of the moving vehicle; the model's output is the inertial navigation alignment error vector, namely the estimated values ​​of the INS three-axis misalignment angle and the three-axis velocity deviation. BiLSTM is a bidirectional LSTM neural network that performs bidirectional time-series dependency modeling on the GNSS and INS difference input data. The forward LSTM network establishes the forward hidden states of the previous and current time steps. The inverse LSTM neural network establishes the backward hidden states of the next time step and the current time step. This yields the fused feature vector at each time step. ; Then, the attention layer is used to fuse the feature vectors output by the BiLSTM at each time step. Calculate the weights and the fused feature vector at each time step. Weighted fusion is performed to obtain the global vector; The weighted and fused global vector is input into a fully connected regression network to predict the inertial navigation alignment error vector. Under stable GNSS observation conditions, the inertial navigation initial alignment error prediction model is updated in real time, i.e. trained, based on historical data under these stable GNSS observation conditions, in order to obtain the model parameters of the inertial navigation initial alignment error prediction model. When GNSS is under unstable observation conditions, the model trained under the previous stable GNSS observation conditions is used to predict the inertial navigation alignment error vector in real time, and finally realize the realignment process of the inertial navigation.

[0009] Furthermore, based on the GNSS-assisted inertial navigation alignment method in the aforementioned complex environment, this invention also proposes a computer device, which includes a memory and one or more processors.

[0010] The executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the GNSS-assisted inertial navigation alignment method described above in complex environments.

[0011] Furthermore, based on the above-mentioned method for GNSS-assisted inertial navigation alignment in complex environments, this invention also proposes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the above-mentioned method for GNSS-assisted inertial navigation alignment in complex environments.

[0012] The present invention has the following advantages: As described above, this invention proposes a GNSS-assisted inertial navigation alignment method for complex environments. This invention utilizes GNSS data and a BiLSTM-Attention-based inertial navigation initial alignment error prediction model to assist in INS recalibration. GNSS data has the advantages of providing absolute position and high accuracy compared to other carrier sensors, and it does not require system initialization. Under good signal conditions, it can guarantee continuous, stable, and high-precision positioning results. Simultaneously, the deep learning model ensures the correction of large positioning errors caused by data loss due to GNSS signal obstruction in complex environments. Compared to conventional methods using Kalman filtering to process the realignment estimation of the inertial navigation system, this invention employs a BiLSTM-Attention neural network model, which is more suitable for temporal features and better conforms to data patterns and motion characteristics. The model algorithm is simple and easy to implement, conforms to data processing rules, and has higher practicality. The BiLSTM-Attention neural network model is better able to capture the positioning and attitude differences between the GNSS and INS systems within a time period, and reasonably analyzes the temporal changes of observed values, thereby achieving higher accuracy in alignment estimation. Attached Figure Description

[0013] Figure 1 This is a flowchart of a GNSS-assisted inertial navigation alignment method in a complex environment, as described in an embodiment of the present invention.

[0014] Figure 2 This is a network structure diagram of the inertial navigation initial alignment error prediction model constructed in this embodiment of the invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 This embodiment 1 describes a method for GNSS-assisted inertial navigation alignment in complex environments. This method addresses the problem of initial alignment and recalibration errors in the INS integrated navigation system with unstable GNSS signals in complex urban environments. It proposes an INS initial alignment error prediction model that utilizes the sensitivity of BiLSTM-Attention to time-related data to achieve high-precision estimation of INS initial alignment errors. This reduces the computational load of the algorithm, improves the INS reconvergence efficiency under unstable GNSS signals, and enhances the estimation efficiency and robustness of the INS initial alignment.

[0016] like Figure 1 As shown in this embodiment, the GNSS-assisted inertial navigation alignment method in complex environments includes the following steps: Step 1. Coarse alignment of the inertial navigation system.

[0017] Coarse alignment is performed on two parts: the correction of pitch angle, roll angle, and heading angle of the moving vehicle.

[0018] The coarse alignment of the pitch and roll angles of the moving vehicle is mainly achieved by using the output information of the accelerometer to level the vehicle, obtain a rough estimate of the two angles, and complete the leveling and alignment process.

[0019] In addition, accelerometers are generally affected by environmental and equipment vibrations, resulting in small-scale irregular vibrations. It is necessary to average the acceleration data over a period of time before calibration calculation.

[0020] Accelerometer data is used to estimate pitch and roll angles. Heading angle correction is mainly achieved through coarse calibration using gyroscopes. High-precision gyroscopes can sensitively detect the Earth's rotation and clearly identify very small angular changes.

[0021] Because of the gyroscope's sensitivity to angle changes, some noise will have a more significant impact on the gyroscope, so it is necessary to perform average processing on the gyroscope data over a period of time.

[0022] Step 2. Realignment using a GNSS-assisted inertial navigation system in complex environments.

[0023] Fine calibration during motion. To address the significant errors that still exist during coarse alignment, the moving carrier is more susceptible to noise and environmental interference, which can easily lead to large deviations in the inertial navigation system (INS) and even static alignment failure. Therefore, more precise INS alignment is required when the carrier is in motion to avoid significant deviations in subsequent carrier positioning.

[0024] The alignment accuracy of the carrier in motion directly affects the accuracy of subsequent integrated navigation calculations and positioning results. Therefore, a high-precision alignment method is required to achieve precise inertial navigation alignment in motion.

[0025] Under stable GNSS observation conditions, GNSS can be used to observe the instantaneous three-dimensional velocity of a dynamic vehicle, and the vehicle's heading angle can be determined. With pitch angle This completes the initial calibration of the inertial navigation system, as shown in the following formula: ; .

[0026] in, , , The three-axis velocities obtained by GNSS measurement under GNSS observation conditions.

[0027] Estimate the initial roll angle based on IMU measurements. The formula is as follows: .

[0028] in express, This represents the velocities along the y and z axes measured by INS. It represents the acceleration due to gravity.

[0029] However, in complex environments, the instability of GNSS signals and significant deviations in GNSS observation speed lead to errors in the estimated initial IMU attitude heading and pitch angles, significantly increasing the error in subsequent integrated navigation and positioning, resulting in inaccurate positioning. To improve the efficiency and accuracy of initial alignment of the inertial navigation system, this invention provides a high-precision and efficient inertial navigation initial alignment system using a BiLSTM-Attention neural network in INS and GNSS integrated navigation modes, as well as an INS recalibration and reconvergence system in complex urban environments. Under poor GNSS observation conditions, to achieve fast and stable INS alignment in unstable dynamic environments, fine alignment needs to be completed within a short time to ensure the real-time performance and reliability of subsequent high-precision positioning measurements. This invention constructs an inertial navigation initial alignment error prediction model based on BiLSTM-Attention. By extracting the position and velocity differences between time-synchronized GNSS and INS measurements as feature vectors, the model aims to minimize the measurement deviation between GNSS and INS. Through iterative optimization during training, the error variance and loss function converge to a predetermined accuracy range. Ultimately, the trained inertial navigation initial alignment error prediction model can output the estimated values ​​of the INS three-axis misalignment angle and the three-axis velocity deviation, thereby achieving high-precision calibration of the inertial navigation system under the motion state of the carrier.

[0030] like Figure 2 As shown, the inertial navigation initial alignment error prediction model includes a BiLSTM layer, an attention layer, and a fully connected regression layer.

[0031] The model input is constructed from the difference between the position and velocity data measured by GNSS and INS of the moving vehicle; the input of the inertial navigation initial alignment error prediction model is assumed to be the system observation vector. The calculation process is as follows: ; in For position observation vectors, This is the velocity observation vector; , These are GNSS position and velocity observations; , This is for the position and velocity data output by the INS.

[0032] The difference between GNSS and INS measured position and velocity is used as the input feature. (i.e., the single-epoch feature below): ;in for and The difference between them , , Components in three directions, for and The difference between them , , Components in three directions.

[0033] By constructing an inertial navigation initial alignment error prediction model, the difference between the obtained GNSS position and velocity values ​​and INS values ​​is used as the input value of the BiLSTM-Attention neural network, and the inertial navigation alignment error is used as the output value of the network model.

[0034] Inertial navigation alignment error The estimated values ​​of the INS triaxial misalignment angle and triaxial velocity deviation are expressed by the following formulas: ; in( ) represents the estimated value of the INS three-axis misalignment angle. ) represents the estimated value of the three-axis velocity deviation.

[0035] This invention utilizes a BiLSTM-Attention neural network model for rapid alignment of inertial navigation system fine calibration.

[0036] BiLSTM is a bidirectional LSTM neural network that models the temporal dependency of GNSS and INS difference input data in two directions. The forward LSTM network establishes the forward hidden state between the previous and current time steps, while the backward LSTM network establishes the backward hidden state between the next and current time steps, thus obtaining the fused feature vector at each time step. .

[0037] Then, the attention mechanism is used to calculate the weights of the features at each time step of the BiLSTM output. The attention mechanism enhancement process is as follows: ; ; Where H is the output of BiLSTM and α is the attention weight vector. This is the weighted fusion feature vector. The resulting global vector is a weighted fusion of the key feature vectors.

[0038] The weight of important moments with good motion status and stable GNSS signal throughout the entire motion time is increased, while the weight of moments with error detection or incomplete data acquisition is decreased.

[0039] This is the attention intermediate feature matrix after nonlinear mapping, which mainly represents the nonlinear expressive power used to enhance the output features of BiLSTM, enabling the model to better distinguish the contribution of different epochs within the time window to the prediction of inertial navigation alignment error.

[0040] This is the weight matrix of the attention layer. Its initial value can be randomly set and is updated through backpropagation along with the parameters of the BiLSTM layer and the fully connected layer during model training.

[0041] The softmax function is used to convert the attention scores of each epoch into non-negative weights, so that the sum of the weights of each epoch is 1.

[0042] Using the learnable attention parameter vector ω to apply attention to the intermediate feature matrix Attention scores are calculated for each epoch feature and normalized using the softmax function to obtain the attention weight vector.

[0043] The BiLSTM-Attention model includes updating parameters through backpropagation during the training phase, dynamically adjusting the output correction amount and attention mechanism during the prediction phase, and realizing dynamic information transfer and weight update through temporal iteration.

[0044] Since inertial navigation error is not determined by the observation difference at a single moment, but is related to the error change trend over multiple epochs, this invention forms a time window sequence from the input features of L consecutive epochs.

[0045] The time window sequence is then input into the BiLSTM-Attention model. Thus, the model input is no longer an isolated single-point observation, but a temporal sequence of GNSS and INS position and velocity deviations over a continuous period of time. The BiLSTM layer is used to extract bidirectional temporal features from the aforementioned difference time window.

[0046] The forward LSTM reads the position difference and velocity difference sequences sequentially in the forward direction of time to learn the variation law of INS error propagation from past epochs to the current epoch, such as the growth law of position deviation caused by the accumulation of velocity error over time, and the influence law of attitude deviation caused by gyroscope zero bias on velocity calculation, etc.

[0047] The inverse LSTM reads the same difference sequence from the end of the time window to the beginning, which is used to supplement the correlation information between the current epoch and subsequent short-term states, and enhance the model's ability to identify error trends, abrupt changes and abnormal observations within a short time window.

[0048] For the i-th epoch within the time window, the forward LSTM outputs the forward hidden state, and the backward LSTM outputs the backward hidden state. The two are concatenated to obtain the fused temporal feature vector for that epoch.

[0049] This yields the BiLSTM output feature sequence over the entire time window, which includes not only the GNSS / INS position and velocity differences at each epoch, but also the trend information, continuity information, and inertial navigation error accumulation characteristics of the differences over time.

[0050] After obtaining the fused temporal features of each epoch using BiLSTM, an attention mechanism is further introduced to adaptively weight the temporal features of different epochs. Specifically, the attention layer calculates the importance of the hidden features of each epoch and normalizes them using the Softmax function to obtain the attention weights. Through this attention mechanism, the model can increase the weight of epochs with relatively stable GNSS observations, representative motion states, and obvious continuous changes in GNSS / INS differences; while reducing the weight of epochs with GNSS jumps, multipath interference, short-term occlusion, or anomalous changes in differences. Thus, in complex environments, the model can pay more attention to the key moments that effectively contribute to the estimation of inertial navigation alignment errors and suppress the interference of low-quality observations on the prediction results.

[0051] (1) Forward propagation Initialize the weights and biases of each layer (input gate weights, forget gate weights, output gate weights, candidate memory weights), and propagate the input time-series signal from the input layer through the LSTM hidden layer step by step. The specific steps are as follows: Forget gate calculation: Based on the current input and the hidden state of the previous time step, calculate the forget gate output to determine the retention ratio of the cell state of the previous time step.

[0052] Input gate calculation: Determines the update weight of the current input information and generates candidate memories.

[0053] Cell state update: Combine the forget gate and the input gate to update the cell state.

[0054] Output gate calculation: The output gate controls the proportion of the cell state output to the hidden state, resulting in the final hidden state. The hidden state is mapped through the output layer to obtain the prediction result, providing an error basis for backpropagation.

[0055] (2) Backpropagation Based on the prediction error from forward propagation, the Backpropagation Through Time (BPTT) algorithm is employed, combined with the chain rule to calculate the gradient of each weight. Starting from the loss function, the error is propagated backward along the time steps, focusing on correcting the gating weights and biases. The parameters are updated according to the learning rate using gradient descent, thus solving the gradient vanishing / exploding problem of recurrent neural networks.

[0056] This feature vector comprehensively characterizes the temporal evolution of the position and velocity differences between GNSS and INS within the current time window. It also highlights key epochs with stable GNSS observations, effective motion states, and obvious error change patterns through attention weights, while suppressing the adverse effects of GNSS jumps, multipath interference, short-term occlusion, or abnormal observation epochs on the prediction results.

[0057] This invention utilizes an attention mechanism to enhance the calculation of weights for the features at each time step of the BiLSTM output, dynamically focusing on key frames / key features (such as sharp turns, acceleration / deceleration, and GNSS abrupt changes).

[0058] After BiLSTM extracts the bidirectional temporal features of the GNSS / INS difference sequence, it further introduces an attention mechanism to adaptively assign weights to features at each epoch within the time window. Since inertial navigation alignment errors are not equally observable at all times, during motion states such as sharp turns, accelerations, decelerations, uphill / downhill slopes, or significant attitude changes, the position and velocity differences between GNSS and INS more readily reflect the influence of gyroscope bias, initial misalignment, and velocity errors on the INS solution results. These epochs typically contain strong error identification information. The attention mechanism can increase the weight of corresponding features at these key epochs, making the model pay more attention to temporal segments that contribute significantly to the prediction of inertial navigation alignment errors.

[0059] Meanwhile, in complex environments such as urban canyons, tall building obstructions, multipath effects, or short-term signal quality degradation, GNSS observations may exhibit jumps, outliers, or local discontinuities. If the features at each time point are directly input into the regression layer with equal weights, abnormal GNSS observations will interfere with the model's prediction results, leading to instability in INS alignment error estimation. Through an attention mechanism, the model can assign smaller weights to epochs that are suspected of being abnormal or have low contributions, based on the consistency of features within the time window and the error evolution pattern, thereby reducing the impact of GNSS abrupt changes, multipath interference, and noisy observations on the INS alignment error prediction results.

[0060] Therefore, the introduction of the attention mechanism enables the model not only to utilize BiLSTM's ability to model sequential dependencies, but also to further filter out key moments and features that are more conducive to error estimation from within the time window.

[0061] This attention mechanism helps enhance the model's adaptability to complex dynamic environments, improves the prediction accuracy of INS three-axis misalignment angle estimates and three-axis velocity corrections, shortens the inertial navigation realignment time after GNSS observation quality deterioration, and improves the robustness and continuous positioning reliability of the integrated navigation system under unstable GNSS conditions.

[0062] This invention utilizes a BiLSTM-Attention neural network for data processing in initial and re-alignment of inertial navigation systems (INS) in complex environments. The model possesses the advantage of capturing temporal data relationships. Under favorable GNSS signal conditions, it can accurately estimate the differences between the GNSS observation vehicle's position and attitude and the INS accelerometer and gyroscope observation data over time, ensuring the accuracy of the initial INS calibration. In complex environments with unstable GNSS signals, it can perform temporal estimation based on these differences, reasonably predicting the INS alignment deviation within a given time period. This ensures the recalibration and rapid reconvergence of the INS in integrated navigation, guaranteeing the positioning accuracy of subsequent integrated navigation systems.

[0063] In addition, the ReLU function was chosen as the overall activation function of the model to improve training efficiency and ensure training accuracy.

[0064] The ReLU function replaces Tanh in some LSTM variants to alleviate the vanishing gradient problem, making it suitable for long-term combined navigation data. Because the ReLU function always assigns a gradient of 1 to positive inputs, it doesn't compress the gradient to a very small value like sigmoid / tanh, effectively modeling complex and unstable GNSS environments. It is naturally compatible with the Attention mechanism. Attention weights are output through softmax and are inherently non-negative. The non-negative output of the ReLU function makes its integration with attention weighting more natural, thus avoiding a large number of negative values ​​that could lead to abnormal weight distribution. This results in more stable attention weights, making the model easier to train.

[0065] The loss function for BiLSTM-Attention needs to be determined based on the task type. In regression tasks, the weighted root mean square error (RMSE) is commonly used to define the loss function.

[0066] For the estimated misalignment angle of the INS three-axis gyroscope, a gyroscope misalignment angle prediction loss term is constructed. : ; For the triaxial velocity deviation estimates, a velocity deviation estimate prediction loss term is constructed. : ; Taking into account the combined effects of gyroscope misalignment angle estimation error and velocity correction error on the initial alignment accuracy of the inertial navigation system, a weighted RMSE loss function is constructed: .

[0067] .

[0068] in This represents the estimated value of the INS three-axis misalignment angle. This indicates the true value of the INS three-axis misalignment angle. This represents the estimated value of the triaxial velocity deviation. This represents the true value of the three-axis velocity deviation. This represents the L2 norm of a vector. The number of training samples. The attitude error weights; This represents the speed error weight.

[0069] Since gyroscope zero bias directly affects attitude calculation and further causes the accumulation of velocity and position errors, the alignment accuracy can be appropriately increased during model training according to the required alignment precision. The weight.

[0070] When more attention is paid to the short-term velocity continuity of the moving vehicle and the re-convergence speed of the integrated navigation system, it is also appropriate to increase... The weight.

[0071] During model training, with the goal of minimizing the aforementioned loss function L, the network parameters of the BiLSTM layer, attention layer, and fully connected regression layer are updated through the backpropagation algorithm, enabling the model to gradually learn the nonlinear mapping relationship between the time series sequences of GNSS / INS position difference and velocity difference and the INS gyroscope zero bias and velocity correction.

[0072] When the loss function converges to a preset threshold or reaches the maximum number of training iterations, the trained BiLSTM-Attention inertial navigation initial alignment error prediction model is obtained. This invention's model comprises two stages: During the training phase, the feature data X of the training dataset is input, and the time correlation sequences of position error and velocity error of GNSS and INS are established. After repeated iterative training, the initial alignment deviation of IMU is output.

[0073] During the prediction phase, the model outputs the initial alignment deviation of the INS in real time, which is used to correct the INS solution and suppress drift divergence. In subsequent processing, the results are fused with the extended Kalman filter to improve the stability of the filter.

[0074] Stable GNSS observation conditions do not simply mean that the GNSS has output, but rather that the GNSS position, velocity, time synchronization, and quality indicators are all relatively reliable. The screening criteria set in this invention are as follows: ; ; ; in: Indicates the number of visible satellites; Indicates the minimum number of satellites; Indicates the position precision factor; This is the threshold for the position accuracy factor; Indicates the standard deviation of GNSS position; Indicates the standard deviation of GNSS position; Indicates the standard deviation of GNSS velocity; This represents the threshold for the standard deviation of GNSS velocity.

[0075] If all the above screening conditions are met, then the observation is considered to be under stable GNSS conditions. In this embodiment... Set the minimum number of satellites to 7; Set to 2.0. Set to 0.05m; Set to 0.02 m / s.

[0076] During the screening process, data segments with severe cycle skipping, multiple paths, occlusion, or obvious location jumps should be removed.

[0077] Under stable GNSS observation conditions, the inertial navigation initial alignment error prediction model is updated in real time, i.e. trained, based on historical data under these stable GNSS observation conditions, in order to obtain the model parameters of the inertial navigation initial alignment error prediction model.

[0078] Under stable GNSS observation conditions, the model training process is as follows: The acquired input samples are fed into the inertial navigation initial alignment error prediction model based on BiLSTM-Attention. The model predicts the estimated values ​​of the three-axis misalignment angle and the three-axis velocity deviation.

[0079] The model input samples are constructed using time-synchronized GNSS and INS measurements of position and velocity deviations.

[0080] Therefore, in the first Each epoch can be used to construct: ; The most basic single-epoch input features are: .

[0081] BiLSTM does not take a single point of data as input, but rather a time series. Let the time window length be... Then the first The training samples are: .

[0082] in ; The length of the time window; The feature dimension is 1 epoch; S represents the interval between the starting positions of two adjacent time windows; The dimension of an input sample.

[0083] If a 5-second time window is used, then , = 1, each epoch inputs 6-dimensional features, then the input sample dimensions are: .

[0084] A sliding window can be generated as follows: ; ; .

[0085] Of course, in this embodiment, the step size can also be set to 5 or 10 to reduce sample overlap.

[0086] Model output samples This refers to the inertial navigation system (INS) alignment error vector that the model needs to predict. This method can quickly complete the recalibration task and allow the model to directly learn the INS velocity error for subsequent compensation. The training dataset generation process is as follows: Under stable GNSS observation conditions, the training dataset is generated according to the following procedure.

[0087] 1. Collect data under multiple operating conditions.

[0088] It should cover different motion states: stationary; uniform linear motion; acceleration; deceleration; turning; uphill / downhill; vibration; low-speed motion; and high-dynamic motion. Although GNSS is stable during the training phase, the motion conditions should not be too singular.

[0089] 2. Construct initial alignment error samples.

[0090] To enable the model to learn the "INS error evolution caused by the initial misalignment angle", multiple sets of initial attitude errors can be set.

[0091] Combined navigation data from different times, road sections, and scenarios is collected, and then mechanically orchestrated using INS to obtain an INS trajectory with initial alignment error. This trajectory is then compared with GNSS or reference ground truth to construct inputs and labels.

[0092] The training process of the BiLSTM-Attention model is as follows: 1. Model Input. The input is a time window sequence: Its formula is expressed as follows: ; in Each row in the text represents a single epoch feature, such as the one mentioned above. That is, single-epoch characteristics.

[0093] 2. BiLSTM temporal feature extraction.

[0094] BiLSTM consists of a forward LSTM and a backward LSTM. BiLSTM is suitable for offline training and short time window realignment because it can utilize both forward and backward temporal information within the training window.

[0095] The essence of model training, in terms of training objective design, is to minimize the error between the predicted correction and the actual correction. ; in: These are the parameters of the BiLSTM-Attention model; The model predicts the output; This is a real label.

[0096] In the inertial navigation initial calibration error prediction model, the inertial navigation alignment error that needs to be predicted includes the estimated values ​​of the three-axis misalignment angle and the three-axis velocity deviation. Therefore, the influence of a single variable cannot be considered alone.

[0097] A multi-task weighted setting is established for the predicted inaccuracy angle and velocity deviation to balance the impact of the two errors on the overall prediction.

[0098] This invention designs a multi-task weighted loss function, the formula of which is expressed as follows: Because the initial alignment quality has a significant impact on subsequent INS error propagation, and since this invention focuses on fine alignment, the attitude error weight can be appropriately increased to achieve the desired result. .

[0099] During model training, the direct optimization objective is to minimize the error between the estimated INS triaxial misalignment angles and triaxial velocity corrections predicted by the model and the corresponding true calibration labels in the training samples. Through iterative calculations during model training, the loss function calculation results are made to meet the preset prediction accuracy, thus completing the model training.

[0100] This invention fully utilizes the strong temporal correlation prediction capability of the BiLSTM-Attention model to enhance the prediction accuracy of GNSS data under unstable GNSS signal conditions, reduce the convergence time of the inertial navigation fine calibration process, and improve the estimation accuracy.

[0101] For the test dataset, the training dataset has formed a high-precision nonlinear mapping relationship between INS and positioning error, forming a deep learning model structure. This structure is then used to test the test dataset and output the INS alignment deviation.

[0102] Under unstable GNSS observation conditions, the model trained under the previous stable GNSS observation conditions is used to predict the estimated values ​​of the three-axis misalignment angle and the three-axis velocity deviation in real time, and finally realize the realignment process of the inertial navigation system.

[0103] Specifically, the BiLSTM-Attention inertial navigation initial alignment error prediction model, which was trained or updated in real time in the previous GNSS stable observation phase, is used to predict the INS alignment error in the current GNSS unstable observation phase in real time. The model input is constructed based on the difference between real-time acquired GNSS data and INS data; then, the inertial navigation alignment error is predicted using the model trained under the previous stable GNSS observation conditions. The model is based on real-time input and outputs predicted INS triaxial misalignment angle estimates and triaxial velocity deviation estimates.

[0104] Specifically, the process of realigning the inertial navigation system is as follows: The corrected INS three-axis gyroscope bias correction value and three-axis velocity correction value are added to the GNSS+INS integrated navigation system's recalibration process in complex environments with unstable GNSS signals, completing the INS calibration, as shown in the following formula: ; in , , for , , Estimated triaxial misalignment angles after axis correction , , This indicates the measured value of the triaxial misalignment angle; , , This refers to the corrected triaxial speed adjustment. , , This indicates the measured speed value.

[0105] This invention addresses the problem of reduced initial alignment accuracy in integrated navigation systems due to degraded GNSS observation quality under complex environments such as GNSS signal obstruction, multipath interference, or reduced satellite count. It proposes a GNSS-assisted inertial navigation realignment method based on BiLSTM-Attention. Instead of directly training the neural network with low-quality GNSS data when observation quality is poor, this invention first determines the GNSS observation status. When GNSS is under stable observation conditions, time-synchronized GNSS data and INS solution data are used to construct GNSS / INS position and velocity difference time series as model input samples. These are combined with the corresponding INS correction values ​​as supervision labels to train or update the BiLSTM-Attention network, enabling the model to continuously learn the nonlinear mapping relationship between the GNSS / INS difference sequence and the INS initial alignment error.

[0106] When the quality of GNSS observations deteriorates, to avoid the degradation of network parameters caused by low-quality GNSS data participating in model training, the system stops model updates and keeps the model parameters obtained from the previous stable observation phase unchanged. Simultaneously, a real-time GNSS / INS difference time window is constructed using GNSS data still available in the current unstable observation phase and INS data as model input. Only forward prediction is performed, and real-time estimates of the INS three-axis misalignment angle and three-axis velocity deviation are output. These prediction results are then fed back into the INS solution process, enabling rapid realignment of the inertial navigation system.

[0107] The BiLSTM network utilizes time-series samples of GNSS / INS differences accumulated during the stable observation phase to learn the evolution of inertial navigation errors over time, thus fully leveraging historical observation information to describe the cumulative characteristics of inertial navigation errors. The Attention mechanism further adaptively selects key epochs that contribute significantly to error prediction within a time window, improving the utilization rate of effective observation information and reducing the impact of local anomalies and short-term noise on the prediction results. Therefore, the BiLSTM-Attention network is not built to handle invalid GNSS data, but rather to establish an error prediction model using reliable GNSS / INS differences obtained during the stable observation phase. This model is then used to predict and compensate for INS alignment errors when GNSS observation quality deteriorates.

[0108] Once the GNSS observation quality stabilizes, the system reuses the GNSS / INS difference samples from the stabilization phase to update the model online, enabling the model to continuously adapt to different environments and motion states, thus achieving a closed-loop working mechanism of "continuous learning during the stable observation phase, prediction and compensation during the observation quality degradation phase, and re-updating during the stabilization phase".

[0109] This invention does not require the establishment of a complex and accurate inertial navigation error model. Instead, it utilizes a neural network model that is continuously updated during the stable observation phase. This model can still predict INS initial alignment errors and complete error compensation during periods of deterioration in GNSS observation quality, thereby ensuring that the integrated navigation system still has good initial alignment accuracy, continuous navigation capability, and robustness in complex environments.

[0110] Example 2 This embodiment 2 describes a system for GNSS-assisted inertial navigation alignment in complex environments, which is based on the same inventive concept as the method for GNSS-assisted inertial navigation alignment in complex environments in embodiment 1.

[0111] The GNSS-assisted inertial navigation alignment system in this embodiment for complex environments includes the following modules: The inertial navigation system coarse alignment module performs coarse alignment for two parts: pitch angle, roll angle correction, and yaw angle correction of the moving vehicle. The precision alignment and calibration module is used for realignment using a GNSS-assisted inertial navigation system in complex environments. First, we build an inertial navigation initial alignment error prediction model based on BiLSTM-Attention. This inertial navigation initial alignment error prediction model includes a BiLSTM layer, an Attention layer, and a fully connected regression layer. The model's input is constructed from the difference between the position and velocity data measured by the GNSS and INS of the moving vehicle; the model's output is the inertial navigation alignment error vector, namely the estimated values ​​of the INS three-axis misalignment angle and the three-axis velocity deviation. BiLSTM is a bidirectional LSTM neural network that performs bidirectional time-series dependency modeling on the GNSS and INS difference input data. The forward LSTM network establishes the forward hidden states of the previous and current time steps. The inverse LSTM neural network establishes the backward hidden states of the next time step and the current time step. This yields the fused feature vector at each time step. ; Then, the Attention layer is used to calculate the weights of the fused feature vector output by BiLSTM at each time step, and the weights of the fused feature vector at each time step are calculated. Weighted fusion is performed to obtain the global vector; The weighted and fused global vector is input into a fully connected regression network to predict the inertial navigation alignment error vector. Under stable GNSS observation conditions, the inertial navigation initial alignment error prediction model is updated in real time, i.e. trained, based on historical data under these stable GNSS observation conditions, in order to obtain the model parameters of the inertial navigation initial alignment error prediction model. When GNSS is under unstable observation conditions, the model trained under the previous stable GNSS observation conditions is used to predict the inertial navigation alignment error vector in real time, and finally realize the realignment process of the inertial navigation.

[0112] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0113] Example 3 This embodiment 3 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the GNSS-assisted inertial navigation alignment method in complex environments described in embodiment 1 above.

[0114] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0115] Example 4 This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the GNSS-assisted inertial navigation alignment method in complex environments described in embodiment 1 above.

[0116] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0117] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for GNSS-assisted inertial navigation alignment in complex environments, characterized in that, Includes the following steps: Step 1. Initial alignment data acquisition for the inertial navigation system; The initial alignment is performed on two parts: the correction of pitch angle, roll angle and heading angle of the moving vehicle. Step 2. Realignment using a GNSS-assisted inertial navigation system in complex environments; First, we build an inertial navigation initial alignment error prediction model based on BiLSTM-Attention. This inertial navigation initial alignment error prediction model includes a BiLSTM layer, an Attention layer, and a fully connected regression layer. The model's input is constructed from the difference between the position and velocity data measured by the GNSS and INS of the moving vehicle; the model's output is the inertial navigation alignment error vector, namely the estimated values ​​of the INS three-axis misalignment angle and the three-axis velocity deviation. BiLSTM is a bidirectional LSTM neural network that performs bidirectional time-series dependency modeling on the GNSS and INS difference input data. The forward LSTM network establishes the forward hidden states of the previous and current time steps. The inverse LSTM neural network establishes the backward hidden states of the next time step and the current time step. This yields the fused feature vector at each time step. ; Then, the attention layer is used to fuse the feature vectors output by the BiLSTM at each time step. Calculate the weights and the fused feature vector at each time step. Weighted fusion is performed to obtain the global vector; The weighted and fused global vector is input into a fully connected regression network to predict the inertial navigation alignment error vector. Under stable GNSS observation conditions, the inertial navigation initial alignment error prediction model is updated in real time, i.e. trained, based on historical data under these stable GNSS observation conditions, in order to obtain the model parameters of the inertial navigation alignment error prediction model. When GNSS is under unstable observation conditions, the model trained under the previous stable GNSS observation conditions is used to predict the inertial navigation alignment error vector in real time, and finally realize the realignment process of the inertial navigation.

2. The method for GNSS-assisted inertial navigation alignment in complex environments according to claim 1, characterized in that, In step 1, the pitch and roll angles of the moving vehicle are coarsely aligned by using the output information of the accelerometer to level them and obtain a rough estimate of the two angles, thus completing the leveling and alignment process; the heading angle is corrected by coarse calibration using a gyroscope.

3. The method for GNSS-assisted inertial navigation alignment in complex environments according to claim 1, characterized in that, The input to the inertial navigation initial alignment error prediction model is the system observation vector. The calculation process is as follows: ； in For position observation vectors, This is the velocity observation vector; , Position and velocity observations measured by GNSS; , For position and triaxial velocity data measured by INS; The triaxial position and velocity differences measured by GNSS and INS are used as input features. : ； in for and The difference between them , , Components in three directions, for and The difference between them , , Components in three directions; Model output The formula is expressed as follows: ; in,( ) represents the estimated value of the INS three-axis misalignment angle. ) represents the estimated value of the three-axis velocity deviation.

4. The method for GNSS-assisted inertial navigation alignment in complex environments according to claim 3, characterized in that, In step 2, the process of training the model under stable GNSS observation conditions is as follows: I. Construct time-synchronized GNSS and INS measurement position and velocity deviations as input samples for model training; since the BiLSTM receives time-series sequences, the time window length is set to... Then the first training input samples for: ； in It includes Individual epoch features; For single-epoch feature dimensions; The sliding step size represents the interval between the starting positions of two adjacent time windows; The dimension of an input sample; No. training input samples Its formula expands as follows: ； in Each row represents a single epoch feature; II. Input the acquired training input samples into the inertial navigation initial alignment error prediction model based on BiLSTM-Attention. The model predicts the estimated values ​​of the three-axis misalignment angle and the three-axis velocity deviation. During model training, a multi-task weighted loss function is designed, and its formula is expressed as follows: ； in The attitude error weights; As the speed error weight; This represents the estimated value of the INS three-axis misalignment angle. This indicates the true value of the INS three-axis misalignment angle. This represents the estimated value of the triaxial velocity deviation. This represents the true value of the triaxial velocity deviation. This represents the L2 norm of a vector.

5. The method for GNSS-assisted inertial navigation alignment in complex environments according to claim 1, characterized in that, In step 2, under unstable GNSS observation conditions, the input of the inertial navigation initial alignment error prediction model is constructed based on the difference between the real-time acquired GNSS data and INS data. Then, the inertial navigation alignment error is predicted using the model trained under the previous GNSS stable observation conditions; the model is based on real-time input and outputs the predicted INS three-axis misalignment angle estimate and the three-axis velocity deviation estimate.

6. The method for GNSS-assisted inertial navigation alignment in complex environments according to claim 3, characterized in that, In step 2, the realignment process of the inertial navigation system is as follows: The estimated values ​​of the three-axis misalignment angles and the estimated values ​​of the three-axis velocity deviations are added to the INS calibration process during the recalibration of the GNSS+INS integrated navigation system in complex environments with unstable GNSS signals, and the correction is performed as follows: ； in , , for , , Estimated triaxial misalignment angles after axis correction , , This indicates the measured value of the triaxial misalignment angle; , , This refers to the corrected triaxial speed adjustment. , , This indicates the measured speed value.

7. The method for GNSS-assisted inertial navigation alignment in complex environments according to claim 1, characterized in that, In step 2, the definition is... Indicates the number of visible satellites. Indicates the position precision factor. Indicates the standard deviation of GNSS position. This represents the standard deviation of GNSS velocity, and the following filtering criteria are set: ； ； ； in, Indicates the minimum number of satellites; This is the threshold for the position accuracy factor; Indicates the standard deviation of GNSS position; Indicates the threshold for the standard deviation of GNSS velocity; If all of the above screening conditions are met, the observation is considered to be under stable GNSS conditions; otherwise, it is considered to be under unstable observation conditions.

8. A GNSS-assisted inertial navigation alignment system for complex environments, characterized in that, Includes the following modules: The inertial navigation system coarse alignment module performs coarse alignment for two parts: pitch angle, roll angle correction, and yaw angle correction of the moving vehicle. The precision alignment and calibration module is used for realignment using a GNSS-assisted inertial navigation system in complex environments. First, we build an inertial navigation initial alignment error prediction model based on BiLSTM-Attention. This inertial navigation initial alignment error prediction model includes a BiLSTM layer, an Attention layer, and a fully connected regression layer. The model's input is constructed from the difference between the position and velocity data measured by the GNSS and INS of the moving vehicle; the model's output is the inertial navigation alignment error vector, namely the estimated values ​​of the INS three-axis misalignment angle and the three-axis velocity deviation. BiLSTM is a bidirectional LSTM neural network that performs bidirectional time-series dependency modeling on the GNSS and INS difference input data. The forward LSTM network establishes the forward hidden states of the previous and current time steps. The inverse LSTM neural network establishes the backward hidden states of the next time step and the current time step. This yields the fused feature vector at each time step. ; Then, the attention layer is used to fuse the feature vectors output by the BiLSTM at each time step. Calculate the weights and the fused feature vector at each time step. Weighted fusion is performed to obtain the global vector; The weighted and fused global vector is input into a fully connected regression network to predict the inertial navigation alignment error vector. Under stable GNSS observation conditions, the inertial navigation initial alignment error prediction model is updated in real time, i.e. trained, based on historical data under these stable GNSS observation conditions, in order to obtain the model parameters of the inertial navigation initial alignment error prediction model. When GNSS is under unstable observation conditions, the model trained under the previous stable GNSS observation conditions is used to predict the inertial navigation alignment error vector in real time, and finally realize the realignment process of the inertial navigation.

9. A computer device comprising a memory and one or more processors; executable code is stored in the memory; characterized in that, When the processor executes the executable code, it performs the steps for implementing the method for GNSS-assisted inertial navigation alignment in a complex environment as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program is used to implement the steps of the GNSS-assisted inertial navigation alignment method in a complex environment as described in any one of claims 1 to 7.

Images