Inertial navigation error correction method based on multi-head attention convolution inertial network

Abstract

The application relates to the technical field of inertial navigation, and particularly discloses an inertial navigation error correction method based on a multi-head attention convolution inertial network, which comprises the following steps: acceleration measurement values and gyro measurement values are calculated to obtain relative displacement estimation values through a multi-head attention convolution inertial network; a MACIN position measurement vector is constructed by using the relative displacement estimation values; a weighted position measurement vector is obtained by weighting and fusing the MACIN position measurement vector and a GNSS position measurement vector; and the error state of inertial navigation system measurement data is estimated and corrected through extended Kalman filtering by using the weighted position measurement vector. Through the multi-head attention convolution inertial network, the application enhances the extraction capability of IMU data features, realizes accurate modeling and efficient correction of inertial navigation errors, and improves the robustness and positioning accuracy of the navigation system in combination with an extended Kalman filtering framework.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of inertial navigation, and in particular to an inertial navigation error correction method based on a multi-head attention convolution inertial network. BACKGROUND

[0002] As one of the core technologies in the field of modern positioning and navigation, inertial navigation technology has been widely applied in complex scenarios such as intelligent driving, aerospace, and robots. However, inertial navigation systems (INS) have the problem of error accumulation during long-term operation, which leads to a gradual decline in positioning accuracy, becoming a key bottleneck restricting the performance improvement. Therefore, how to effectively correct the inertial navigation error, especially in the demand for high-precision positioning in complex dynamic environments, has become a research hotspot.

[0003] A Chinese invention patent with publication number CN110702104A proposes an inertial navigation error correction method based on vehicle zero speed detection. This method collects wheel speed sensor data and INS data, uses the Riemann Pearson criterion for zero speed detection, and corrects the speed, angular velocity, and gyroscope zero offset in the zero speed state to suppress error accumulation. However, this method mainly relies on zero speed detection, and is limited to vehicles and other carriers that have zero speed states. Its error correction ability for non-zero speed states is limited. At the same time, this method does not fully utilize the advantages of multi-source data fusion, and is difficult to deal with error drift in complex dynamic environments.

[0004] A Chinese invention patent with publication number CN110553668A proposes an error correction system and method for inertial navigation in intelligent driving. It calculates the distance deviation by combining the actual road position with the inertial navigation information, and corrects it when the deviation exceeds the threshold. Although this method can correct the inertial navigation error in time, its correction accuracy is highly dependent on the accuracy of external reference data, and in the absence of high-precision external data, the correction effect may be significantly reduced. At the same time, this method does not fully consider the internal error propagation mechanism of the inertial navigation system, which may lead to incomplete error compensation in the correction process, affecting the overall positioning accuracy.

[0005] The deficiencies of the above-mentioned prior art show that the current inertial navigation error correction methods still have obvious limitations in terms of adaptability to complex dynamic environments, multi-source data fusion capability, and internal error propagation mechanism modeling. Specifically, existing methods are difficult to achieve efficient error correction in non-zero speed states, lack deep mining of IMU (Inertial Measurement Unit) data features, and cannot maintain high-precision positioning in the absence of GNSS or external reference data. SUMMARY

[0006] The present application aims to solve the above problems. To this end, the present application provides an inertial navigation error correction method based on a multi-head attention convolution inertial network, which enhances the extraction capability of IMU data features by introducing a multi-head attention mechanism and a convolution neural network, realizes accurate modeling and efficient correction of inertial navigation errors. At the same time, the present application combines the extended Kalman filter framework to construct a hybrid neural-inertial fusion filter mechanism, fully utilizes the advantages of multi-source data fusion, and improves the robustness and positioning accuracy of the navigation system, so as to meet the high-precision navigation demand in complex scenes.

[0007] The present application provides an inertial navigation error correction method based on a multi-head attention convolution inertial network, which adopts the technical scheme as follows: comprising the following steps:

[0008] S1: obtaining inertial navigation system measurement data and GNSS measurement data;

[0009] S2: according to the accelerometer measurement value and the gyroscope measurement value in the inertial navigation system measurement data, the relative displacement estimation value is calculated through the multi-head attention convolution inertial network;

[0010] S3: using the relative displacement estimation value, combining the yaw angle and the position initial condition, the MACIN position measurement vector is constructed;

[0011] S4: the MACIN position measurement vector and the GNSS position measurement vector in the GNSS measurement data are weighted and fused to obtain the weighted position measurement vector;

[0012] S5: using the weighted position measurement vector, the error state of the inertial navigation system measurement data is estimated and corrected through the extended Kalman filter.

[0013] Further, the multi-head attention convolution inertial network comprises an input layer, a first Transformer layer, a plurality of convolution layers, a second Transformer layer and a full connection layer stacked in turn.

[0014] Further, in step S3, the MACIN position measurement vector is expressed as:

[0015]

[0016]

[0017] wherein, is the x-direction component of the position at time k, is the y-direction component of the position at time k; is the x-direction component of the position at time k-1, is the y-direction component of the position at time k-1, and initial condition of corresponding position; relative displacement estimation value at k moment; yaw angle at k moment, is a matrix transpose.

[0018] Further, in step S4, the MACIN position measurement vector and the GNSS position measurement vector are weighted and summed to obtain a weighted position measurement vector.

[0019] Further, in step S4, the weight is calculated based on the residual entropy of the GNSS measurement data, and the calculation formula is:

[0020]

[0021] wherein λ is a historical mean value of the residual entropy, is a GNSS weight, is a MACIN weight, is a residual entropy at k moment.

[0022] Further, in step S4, when the GNSS signal is interrupted, the MACIN position measurement vector is taken as the weighted position measurement vector.

[0023] Further, in step S5, the state update equation of the extended Kalman filter is:

[0024]

[0025] wherein, is a state posteriori estimation value, is a state priori estimation value, is a Kalman gain, is a measurement matrix, is a weighted position measurement vector.

[0026] Further, in step S5, the relative displacement estimation value is included in the state vector, and the state vector is expressed as:

[0027]

[0028] wherein, is a position error in x direction, is a position error in y direction, is a velocity error in x direction, is a velocity error in y direction, is a yaw rate error, is a relative displacement estimation value, is a matrix transpose.

[0029] Further, in step S5, a memory factor is introduced, and the covariance prediction formula is extended as:

[0030]

[0031] wherein, is the covariance prior estimate value at k moment, F is a state transition matrix, is the covariance posterior estimate value at k-1 moment, is the covariance posterior estimate value at k-2 moment, is a process noise covariance matrix, is a memory factor, is a matrix transpose.

[0032] Further, in step S5, the memory factor is calculated according to the GNSS interruption duration.

[0033] The one or more technical solutions in the embodiments of the application have at least one of the following technical effects:

[0034] The application significantly improves the precision, robustness and environmental adaptability of inertial navigation error correction by deep fusion of multi-head attention mechanism and convolutional neural network, combined with a hybrid neural-inertial filtering framework. The specific advantages are as follows:

[0035] 1. IMU data feature mining ability is significantly enhanced

[0036] Long sequence dependence modeling: The multi-head attention mechanism of the Transformer layer can capture the long-distance dependence relationship in the IMU data across time steps, and is better at analyzing the global features of complex motion patterns compared to traditional convolutional networks.

[0037] Local feature fine extraction: The convolutional layer extracts the local change trend of acceleration and angular velocity (such as transient features such as sudden acceleration and turning) through sliding window operation, which is complementary to the global modeling of Transformer, realizing multi-scale characterization of error propagation law.

[0038] 2. Error correction precision is improved in complex environments

[0039] Cooperative optimization when GNSS is available: By adding the MACIN prediction value as a state variable, the active estimation and correction of the "data-driven model error" in the neural-inertial fusion system is realized for the first time, forming a double closed-loop correction mechanism of "physical error + model error", which theoretically breaks through the dependence of traditional filtering on "ideal measurement source", and provides theoretical support for high-precision navigation in complex scenarios.

[0040] Continuous correction during GNSS interruption: dynamically weighting historical covariance difference by memory factor, enabling the filter to "learn" the law of historical errors, fundamentally solving the modeling defect of traditional EKF on error nonlinear accumulation, providing key support for navigation continuity in complex environment.

[0041] 3. Robustness enhancement of fusion framework

[0042] Noise robustness: EKF update residual check mechanism effectively suppresses the impact of sensor burst noise.

[0043] Multi-source complementarity: heterogeneous data fusion of GNSS and MACIN makes up for the defects of single data source, for example, when GNSS multipath error is large in urban canyons, MACIN's inertial feature modeling can provide reliable correction, and vice versa.

[0044] 4. Engineering application value

[0045] Cost advantage: no need to rely on high-precision external equipment (such as lidar, depth camera), only low-cost IMU can achieve high-precision correction, suitable for cost-sensitive scenarios such as logistics robots.

[0046] Deployment flexibility: modular design can seamlessly embed existing INS / GNSS fusion system without reconfiguring hardware architecture, facilitating engineering migration.

[0047] In summary, the present application breaks through the dependence on external data and specific scenarios of existing technologies through the deep fusion of intelligent algorithms and traditional filtering, providing a solution with precision, robustness and real-time for inertial navigation error correction in complex dynamic environments, with significant academic innovation value and engineering application potential.

[0048] Additional aspects and advantages of the application will be described in part below, some will become apparent from the following description, or will be understood by those skilled in the art through practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0050] Figure 1 is a flowchart of the method provided by the application.

[0051] Figure 2 is an architectural schematic diagram of the multi-head attention convolution inertial network provided by the application.

[0052] Figure 3 is a positioning error comparison chart in a normal scenario of GNSS provided by the present application.

[0053] Figure 4 is an error comparison chart before and after GNSS interruption provided by the present application. DETAILED DESCRIPTION

[0054] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be clearly and completely described below in combination with the drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application. The following embodiments are used to illustrate the present application, but cannot be used to limit the scope of the present application.

[0055] In the description of the present application, the description of the terms “one embodiment”, “some embodiments”, “example”, “specific example” or “some examples” means that the specific features, structures or characteristics described in combination with the embodiment or example are contained in at least one embodiment or example of the embodiments of the present application. In the present application, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, the different embodiments or examples described in the present application and the features of the different embodiments or examples can be combined and combined by those skilled in the art without contradiction.

[0056] The inertial navigation system (INS) has an error accumulation problem in long-time operation, resulting in a decline in positioning accuracy. Existing error correction methods mainly rely on external reference data or zero-speed detection, but have insufficient adaptability in complex dynamic environments and limited multi-source data fusion capability. Therefore, the present application introduces a multi-head attention mechanism and a convolutional neural network (CNN) to extract deep features and model non-linearly from inertial measurement unit (IMU) data, thereby accurately capturing error propagation rules and combining an extended Kalman filter (EKF) framework to implement error compensation, significantly improving the positioning accuracy and robustness of the navigation system.

[0057] The present application will be further described below in combination with Figures 1 to 4 A multi-head attention convolutional inertial network-based inertial navigation error correction method is described.

[0058] In the present embodiment, as shown in Figure 1 A multi-head attention convolutional inertial network-based inertial navigation error correction method is provided, comprising the following steps:

[0059] S1: Obtain inertial navigation system measurement data and GNSS measurement data.

[0060] S2: According to the accelerometer measurement value and the gyroscope measurement value in the inertial navigation system measurement data, a relative displacement estimation value is calculated through a multi-head attention convolutional inertial network.

[0061] The embodiment designs a multi-head attention convolutional inertial network (MACIN), which adopts an alternating stacking structure of "Transformer-convolution-Transformer", forms a "global-local-global" feature extraction link, fully gives play to the advantages of the multi-head attention mechanism and the convolution operation, and enhances the network's extraction capability for IMU data features.

[0062] As shown in Figure 2 , the multi-head attention convolutional inertial network includes an input layer, a first Transformer layer, a plurality of convolutional layers, a second Transformer layer and a fully connected layer stacked in sequence.

[0063] The data received by the input layer is denoted as , which is a tensor with a shape of Tx6x1, wherein T represents a time step, that is, window data containing T IMU measurement values, for example, T can be 200; 6 corresponds to 6 measurement dimensions of three-axis accelerometers and three-axis gyroscopes; and 1 represents a channel number. These data reflect the motion state information of the platform within a certain time, capture long time sequence dependence through the first Transformer layer, extract local features through L convolutional layers, and finally integrate global features through the second Transformer layer.

[0064] The input is input into the first Transformer layer, and the Transformer layer is composed of a multi-head attention mechanism (Multi-Head Attention) and a feed-forward neural network (Feed-Forward Neural Network). In the multi-head attention mechanism, query vectors , key vectors and value vectors are obtained through linear transformation, and the formula for calculating attention weights is as follows:

[0065]

[0066] wherein is the key vector dimension of the matrix, is a normalization function, is a matrix transpose.

[0067] The multi-head attention mechanism calculates in parallel through multiple different attention heads, captures data features from different angles, and finally splices the results of each head and performs linear transformation to obtain the output . The feedforward neural network further processes to obtain the output of the first Transformer layer, enhancing the non-linear expression ability of the network. The first Transformer layer can effectively capture the long-distance dependency relationship of the IMU data at different time steps through the self-attention mechanism, and mine the potential features in the data.

[0068] into L convolutional layers. Taking the first convolutional layer as an example, the input is , and the calculation method of the output feature map is as follows:

[0069]

[0070] wherein, is the convolution kernel weight matrix of the first convolutional layer, represents a convolution operation, is the bias vector of the first convolutional layer, is the input of the first convolutional layer, is the output of the first convolutional layer, is an activation function, and the ReLU activation function is selected in this embodiment. The value of i is 1, 2, …, L, when i = 1, The convolution kernel size, step length and other parameters of each convolutional layer are determined through optimization, and the local features in the IMU data, such as the velocity and acceleration change trend of the platform motion, are further extracted through the convolution operation. After processing through multiple convolutional layers, the final output of the convolutional layer is obtained. The convolutional layer extracts local features such as the change trend of the motion state through a sliding window, enhancing the adaptability of the network to complex dynamic environments. The alternating stacking design of the Transformer layer and the convolutional layer not only can capture the long-distance dependency relationship of the IMU data, but also can effectively extract local features, thereby improving the expression ability of the model.

[0071] ​​​ The second Transformer layer is entered. Similarly, in the multi-head attention mechanism of the second Transformer layer, The query vector, the key vector and the value vector are obtained by linear transformation, and are calculated according to the above attention weight calculation formula. After multi-head splicing and linear transformation, the output of the second Transformer layer is obtained after being processed by the feedforward neural network The second Transformer layer uses the multi-head attention mechanism and the feedforward neural network to globally integrate and optimize the features extracted by the convolutional layer, further excavate the correlation between the features, and enhance the adaptability of the network to complex motion patterns.

[0072] Finally, After the flattening operation, the full connection layer is input for processing. The full connection layer integrates the feature vectors processed by the Transformer layer and the convolutional layer, and outputs a vector with a dimension of 1 through the full connection calculation between neurons, corresponding to the relative displacement estimation value. The estimation value is used as a key parameter of the platform position change, which is used for subsequent navigation calculation.

[0073] The data set is constructed in this embodiment, and the specific process is as follows: collect inertial sensor data, including accelerometer measurement values and gyroscope measurement values, and obtain high-precision true trajectory data by using GNSS RTK (Global Navigation Satellite System Real-Time Kinematic). In order to eliminate the influence of sensor noise and uneven data distribution, the above data is preprocessed such as normalization, so that the mean value of the data is 0 and the variance is 1. Subsequently, the training set and the test set are divided according to a certain proportion, which are used for subsequent model training and performance evaluation.

[0074] Model training process: using the above data set, the MACIN is trained by using the ADAM optimizer, and the learning rate is dynamically adjusted by using the scheduling strategy. The training is performed with a window size of 200 samples and a batch size of 64, and the training process is performed for a total of 70 epochs. During training, the model input is the IMU measurement value window data, and the output is the distance change prediction value. The prediction value is compared with the corresponding distance change true value in the true trajectory data, and the mean square error loss function MSE is calculated. Through the back propagation algorithm, the gradient of the loss function is transmitted to each layer of the network, the network parameters are updated, and the network structure is optimized, so that the model can accurately regress the distance change of the platform. And the adaptive learning rate scheduling strategy dynamically adjusts the learning rate to speed up the convergence and prevent overfitting. At the same time, through the Dropout regularization technology, part of the neurons are randomly discarded, and the generalization ability of the model is improved.

[0075] S3: Construct MACIN position measurement vector with relative displacement estimate, combined with yaw angle and position initial condition.

[0076] The relative displacement estimate output by MACIN, combined with the yaw angle and position initial condition provided by the attitude and heading reference system, can be used to calculate the platform position:

[0077]

[0078] where, is the x-direction component of the position at time k, is the y-direction component of the position at time k, used to construct the MACIN position measurement vector; is the x-direction component of the position at time k-1, is the y-direction component of the position at time k-1, corresponding to the position initial condition; is the relative displacement estimate at time k; is the yaw angle at time k.

[0079] Thus, the MACIN position measurement vector is constructed as , .

[0080] S4: Weighted fusion of MACIN position measurement vector and GNSS position measurement vector in GNSS measurement data to obtain weighted position measurement vector.

[0081] Residual sequence of GNSS measurement data is recorded using a sliding window, and residual entropy is calculated:

[0082]

[0083] where, is the residual entropy at time k, is the probability of occurrence, is the i-th residual sample in the residual sequence, is obtained by kernel density estimation or histogram statistics, and n is the number of residual samples in the sliding window, usually 100-200. Based on the residual entropy, the fusion weight of GNSS and MACIN is calculated:

[0084]

[0085] where λ is the historical mean of the residual entropy, used for normalization,

[0086] is the GNSS weight, is the MACIN weight. When the GNSS signal is disturbed, such as multipath effect, the dispersion of residual distribution increases, increases, ​Automatically reduce.

[0087] Constructing a weighted position measurement vector :

[0088]

[0089] wherein, is a GNSS position measurement vector, obtained from GNSS measurement data.

[0090] When GNSS signal is interrupted, GNSS position measurement vector is unavailable, at this time only use MACIN position estimate as external update source, MACIN position measurement vector is directly used as weighted position measurement vector, that is .

[0091] In summary, the weighted position measurement vector when GNSS is available and interrupted can be expressed as:

[0092] .

[0093] S5: Using the weighted position measurement vector, the error state of the inertial navigation system measurement data is estimated and corrected by extended Kalman filtering.

[0094] This step is based on the prediction-update two-stage framework of extended Kalman filtering (EKF), which deeply integrates the MACIN neural network output and GNSS measurement data, and realizes the full-process correction of inertial navigation error through dynamic state space expansion, adaptive measurement update and robustness enhancement algorithm.

[0095] In the prediction stage, this embodiment uses the relative displacement estimate to expand the state prediction equation.

[0096] In traditional vehicle navigation systems, the INS state vector only contains position, velocity and attitude error, and does not consider the prediction error of neural network output. However, as a data-driven model, the relative displacement estimate output by MACIN has inherent errors, such as generalization bias of the model to complex dynamic scenes, prediction deviation caused by sensor noise coupling, etc. If it is directly used as "error-free" measurement value for filtering update, it will introduce unmodeled error sources, resulting in a decrease in fusion accuracy. This embodiment innovatively includes the relative displacement estimate of MACIN in the state vector, constructing a 6-dimensional state vector :

[0097]

[0098] wherein, is the position error in the x direction, is the position error in the y direction, is the velocity error in the x direction, is the velocity error in y direction, is the yaw rate error, is the relative displacement estimate.

[0099] By including the relative displacement estimate of MACIN in the state vector, the dynamic modeling of neural network error is achieved. Through the state transition equation, the prediction error of MACIN is estimated in real time and is used as a closed-loop correction of the internal state of the system, rather than simply considering it as an “ideal measurement value”. In addition, the coupling relationship between neural-inertial error can be established, and the prediction error of MACIN is dynamically related to the physical quantity error of INS, for example, the velocity error of INS will cause the IMU data to deviate, and then affect the feature extraction accuracy of MACIN. By including the state vector, the coupling relationship can be quantified through the state transition matrix F to realize the collaborative correction of multi-source errors.

[0100] The state transition equation is:

[0101]

[0102] In the formula, is the state prior estimate at time k, is the state posterior estimate at time k-1, is the process noise gain matrix, 6x3 dimension, corresponding to the three-axis noise input of IMU, is the process noise vector, 3 dimension, containing the white noise of accelerometer and gyroscope; F is the state transition matrix, 6x6 dimension, and the structure is:

[0103] .

[0104] The first two rows of F correspond to the integral relationship of position error and velocity error, and the last four rows represent the zero self-transition of other state quantities.

[0105] The covariance prediction of the traditional extended Kalman filter only depends on the current state transition and process noise, and does not consider the cumulative trend of historical errors. In the GNSS interruption scene, the error of INS presents a nonlinear cumulative characteristic with time, such as the quadratic growth of position error caused by accelerometer error. Only through the covariance of the current and last time, the long-term error propagation law cannot be accurately described, and covariance underestimation (leading to filter divergence) or overestimation (leading to correction lag) is prone to occur.

[0106] The historical error memory factor is introduced in this embodiment, and the covariance prediction formula is extended to:

[0107]

[0108] wherein, ​is the covariance prior estimate at time k, a 6x6 matrix; is the covariance posterior estimate at time k-1; is the covariance posterior estimate at time k-2; is the process noise covariance matrix, 6x6, representing the effect of IMU noise on the state; is the memory factor. is the historical error compensation term, which can capture the error accumulation trend, reflects the covariance difference between the last time and the last two times, quantifying the error growth rate; By weighting this difference, the "memory" of historical errors is injected into the current covariance estimate, enabling the filter to predict the evolution of errors over time, especially suitable for GNSS long-term interruption scenarios.

[0109] According to the duration of GNSS interruption, the calculation formula is:

[0110]

[0111] wherein, is the duration of GNSS interruption, in seconds, is the minimum value. The shorter the interruption time, such as less than 100 seconds, is smaller, taking 0.1~0.8, mainly relying on the current state transition; the longer the interruption time, such as more than 100 seconds, is fixed at 0.8, strengthening the historical error compensation, avoiding the underestimation of actual errors due to long-term interruption.

[0112] In the update phase, the embodiment uses a weighted position measurement vector to estimate and correct the error state of the inertial navigation system measurement data through the state update equation of EKF.

[0113] The state update equation of EKF is:

[0114]

[0115] wherein, is the state posterior estimate, is the state prior estimate, is the Kalman gain, is the residual, To measure the matrix, the sub-matrix corresponding to GNSS and MACIN is combined, when the GNSS signal is interrupted, the measurement matrix H is degraded into the sub-matrix corresponding to MACIN, and the positioning accuracy of the navigation system is maintained through the state update equation. In addition, in order to ensure the stability of the filter, the hybrid neural-inertial fusion filter also includes a noise amplification factor checking mechanism, which monitors abnormal observation values in real time and eliminates their influence by calculating the covariance matrix of the residual.

[0116] The MACIN architecture combining the multi-head attention mechanism and the convolutional neural network is innovatively introduced, which can accurately model the inertial navigation error propagation law and significantly improve the error correction accuracy. At the same time, the hybrid neural-inertial fusion filter based on the EKF framework fully utilizes the advantages of MACIN and GNSS, state space expansion and memory covariance compensation, suppresses the error growth when GNSS is interrupted and maintains the convergence of the filter, and the dynamic weighting mechanism suppresses the influence of GNSS multipath error on the integrated navigation error, effectively meeting the navigation demand in complex dynamic environment.

[0117] In order to verify the actual effect of the embodiment, a group of experimental scenes are selected for testing. The experiment is carried out in a city road containing complex dynamic environment, and the test vehicle is equipped with high-precision IMU and GNSS RTK equipment. During the experiment, first, the IMU and GNSS data are collected, and after normalization processing, they are input into MACIN for training. After training, MACIN can accurately output the relative displacement estimation value of the platform position change. Then, the output of MACIN is combined with the GNSS measurement value, and input into the EKF framework for hybrid neural-inertial fusion filtering.

[0118] (1) State expansion test results

[0119] Directly taking the MACIN output as the measurement value and ignoring its prediction error will cause the fusion result to be "polluted" in the GNSS multipath scene. The MACIN error in the state vector is tracked in real time by the embodiment, which reduces the positioning root mean square error (RMSE) from 0.72 meters to 0.58 meters, a reduction of 19.4%, as shown in Figure 3

[0120] (2) Covariance compensation test results

[0121] As shown in Figure 4 ​​As shown, the GNSS signal is interrupted at 200 seconds, and the experimental results show that, in the case of normal GNSS signal, after covariance compensation, the positioning error of the system is further reduced and is obviously smaller than the GNSS error, and the average error is controlled within 0.5 meters; and in the case of GNSS signal interruption, the system can still rely on the output of MACIN to maintain high positioning accuracy, and the error does not exceed 1.5 meters. This fully proves the robustness and high-precision characteristics of the method in complex dynamic environments.

[0122] In summary, by introducing the MACIN architecture combining the multi-head attention mechanism and the convolutional neural network, the inertial navigation error propagation law can be accurately modeled, and the error compensation can be realized in combination with the EKF framework. The application not only significantly improves the positioning accuracy of the navigation system, but also effectively meets the navigation requirements in complex dynamic environments.

[0123] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A method for correcting inertial navigation error based on a multi-head attention convolution inertial network, characterized in that, The method comprises the following steps: S1: acquiring inertial navigation system measurement data and GNSS measurement data; S2: calculating a relative displacement estimation value by a multi-head attention convolution inertial network according to accelerometer measurement values and gyroscope measurement values in the inertial navigation system measurement data; S3: constructing a MACIN position measurement vector by combining the relative displacement estimation value, a yaw angle and a position initial condition; S4: weighting and fusing the MACIN position measurement vector and a GNSS position measurement vector in the GNSS measurement data to obtain a weighted position measurement vector; S5: estimating and correcting an error state of the inertial navigation system measurement data by an extended Kalman filter using the weighted position measurement vector.

2. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 1, wherein, The multi-head attention convolution inertial network comprises an input layer, a first Transformer layer, a plurality of convolution layers, a second Transformer layer and a full connection layer stacked in sequence. 3.The method of claim 1, wherein, In step S3, the MAC IN position measurement vector is denoted as: wherein, is the x-direction component of the position at time k, is the y-direction component of the position at time k; is the x-direction component of the position at time k-1, is the y-direction component of the position at time k-1, and corresponding position initial conditions; is the relative displacement estimate at time k; is the yaw angle at time k, is the matrix transpose.

4. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 1, wherein, In step S4, the MACIN position measurement vector and the GNSS position measurement vector are weighted and summed to obtain the weighted position measurement vector.

5. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 4, wherein, In step S4, the weight is calculated based on residual entropy of the GNSS measurement data, and a calculation formula is: where λ is the historical mean of the residual entropy, is the GNSS weight, is the MACIN weight, is the residual entropy at time k.

6. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 4, characterized in that, In step S4, when the GNSS signal is interrupted, the MACIN position measurement vector is taken as the weighted position measurement vector.

7. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 1, wherein, In step S5, a state update equation of the extended Kalman filter is: wherein, is a state posterior estimate, is a state prior estimate, is a Kalman gain, is a measurement matrix, is a weighted position measurement vector. 8.The method of claim 1, wherein, In step S5, the relative displacement estimate is incorporated into the state vector, the state vector is represented as: wherein is a position error in the x-direction, is a position error in the y-direction, is a velocity error in the x-direction, is a velocity error in the y-direction, is a yaw rate error, is a relative displacement estimate, is a matrix transpose.

9. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 1, wherein, In step S5, a memory factor is introduced, and a covariance prediction formula is extended to: wherein, is the covariance prior estimate at time k, F is the state transition matrix, is the covariance posterior estimate at time k-1, is the covariance posterior estimate at time k-2, is the process noise covariance matrix, is the memory factor, is the matrix transpose.

10. The inertial navigation error correction method based on the multi-head attention convolution inertial network according to claim 9, wherein, The memory factor is calculated according to a GNSS interruption duration.

Images