Integrated navigation performance compensation method under GNSS (Global Navigation Satellite System) interruption

Abstract

The invention relates to the technical field of unmanned aerial vehicle navigation, in particular to an integrated navigation performance compensation method under GNSS (Global Navigation Satellite System) interruption, which comprises the following steps: constructing an initial deep learning model by using a first data set, the initial deep learning model comprising an extended long and short term memory module and a Transformer module which are connected; performing supervised learning on the initial deep learning model by using the second data set to obtain a deep learning model, the deep learning model being used for predicting a pseudo measurement value of the global navigation satellite system, inputting the pseudo measurement value, output data of the inertial navigation system and a distance measurement value output by the ultra wide band sensor into a federated Kalman filter, calculating to obtain an error, and calculating the error; correcting output data of the inertial navigation system according to the error; and realizing integrated navigation of the unmanned aerial vehicle according to the corrected output data of the inertial navigation system. According to the method, the influence on the navigation performance under the GNSS interruption condition is compensated.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation technology, and more specifically to a method for compensating for the performance of integrated navigation under GNSS interruption. Background Technology

[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, its applications in military and civilian fields are becoming increasingly widespread. Navigation and positioning, as a core component of UAV flight, directly impacts mission performance in terms of accuracy and reliability. Currently, UAV navigation systems primarily employ a combination of Inertial Navigation Systems (INS) and Global Navigation Satellite Systems (GNSS). GNSS provides absolute positioning information, but its signals are susceptible to environmental interference and have a relatively low output frequency (typically 1 Hz). In contrast, INS, based on an Inertial Measurement Unit (IMU), achieves autonomous navigation, featuring high-frequency output (typically above 50 Hz) and short-term high accuracy. However, due to its reliance on integral calculations, errors accumulate over time, making long-term positioning accuracy difficult to guarantee. Therefore, GNSS / INS integrated navigation, through complementary advantages, has become the mainstream solution for high-precision UAV positioning.

[0003] However, in complex environments such as urban canyons, tunnels, or tree-lined roads, GNSS signals are susceptible to obstruction or multipath effects, leading to decreased positioning accuracy or even complete interruption. In such cases, the GNSS / INS integrated system degenerates into pure INS navigation, with errors accumulating rapidly and severely impacting the positioning reliability of the UAV. To compensate for this deficiency, existing technologies introduce ultra-wideband (UWB) sensors, utilizing their high-precision ranging capabilities to provide relative positioning information, forming a GNSS / INS / UWB integrated navigation system. However, UWB coverage is limited, and when GNSS is interrupted for extended periods, INS errors continue to diverge due to the lack of absolute position correction, ultimately degrading navigation performance. Existing methods struggle to effectively suppress error accumulation in scenarios with prolonged GNSS signal loss, failing to meet the continuous high-precision positioning requirements of UAVs. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for compensating for integrated navigation performance during GNSS interruptions. The invention primarily utilizes an extended long short-term memory module and a neural network architecture fused with a Transformer to predict spurious measurements during GNSS interruptions. These spurious measurements are then fused with INS / UWB data using a federated Kalman filter. This method continuously corrects accumulated INS errors and maintains the high-precision integrated navigation performance of the UAV when GNSS signals are missing.

[0005] The technical means employed in this invention are as follows: Compared with the prior art, the present invention has the following advantages: 1. This invention achieves high-precision continuous positioning of UAV navigation systems in the event of GNSS interruption by deeply integrating with deep learning and federated Kalman filtering techniques.

[0006] 2. The extended long short-term memory and Transformer fusion neural network architecture provided by this invention achieves high-precision prediction of GNSS pseudo-measurements by efficiently extracting INS temporal features and accurately capturing sequence dependencies.

[0007] 3. The sensor collaborative compensation mechanism provided by this invention achieves effective compensation for the performance of UAV integrated navigation system in complex environments by fusing predicted GNSS pseudo-measurement values ​​with INS / UWB data through federated Kalman filtering.

[0008] 4. The federated filter data fusion scheme provided by this invention achieves stable maintenance of UAV positioning accuracy during GNSS outages by processing GNSS pseudo-measurement values ​​and UWB relative ranging information in parallel.

[0009] 5. The dual-stage training and compensation mechanism provided by this invention achieves an intelligent solution to the complex environment combined navigation problem of UAVs by training the model when the signal is normal and applying prediction when the signal is interrupted.

[0010] Based on the above reasons, this invention can be widely applied in fields such as drone navigation. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a method for compensating the performance of integrated navigation under GNSS interruption according to the present invention.

[0013] Figure 2 This is a schematic diagram of the deep learning model of the present invention.

[0014] Figure 3 This is a schematic diagram of the flight trajectory of the UAV based on the present invention.

[0015] Figure 4 This is a schematic diagram of the prediction results based on the xLSTM-transformer model in an embodiment of the present invention.

[0016] Figure 5 This is a diagram showing the distance error results of the UAV in an embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] like Figure 1 As shown, this invention provides a method for compensating for the performance of integrated navigation under GNSS interruption, the specific steps of which are as follows: S1. Obtain the dataset, which includes observation data from the Global Navigation Satellite System, output data from the Inertial Navigation System, and ranging values ​​output by the Ultra-Wideband Sensor.

[0020] S2. Randomly divide the dataset into the first dataset and the second dataset.

[0021] The dataset is normalized and divided proportionally into a training set (first dataset) and a test set (second dataset). The training set is used for optimizing and learning model parameters, while the test set is used to evaluate the model's prediction accuracy and generalization performance. Next, the Min-max normalization method is used to process the data, mapping the input and output features of the sample data to […]. [1, 1]. This process can eliminate differences in the dimensions of different features and maintain data scale consistency, thereby laying the foundation for subsequent model training and analysis and improving the model's performance and generalization ability.

[0022] S3. Construct an initial deep learning model using the first dataset. The initial deep learning model includes a connected extended long short-term memory module and a Transformer module.

[0023] Deep learning models combine the advantages of xLSTM and transformers, enabling them to better capture the feature information and dependencies of data and improve the model's predictive performance. An extended long short-term memory module is used for data processing and feature extraction, while the Transformer module is used for processing sequence data and capturing the dependencies within the sequence through a multi-head self-attention mechanism.

[0024] like Figure 2 As shown, the workflow of the initial deep learning model includes: The input data is fed into the Extended Long Short-Term Memory (LSTM) module, which includes a first LSTM module and a second LSTM module, to obtain the output of the LSTM module. The output of the LSTM module is then fed into the Transformer module to obtain pseudo-measured values.

[0025] The workflow of the first LSTM module includes: The first step is to receive the current input vector and the hidden state from the previous time step, process them through three gating structures (input gate, forget gate, and output gate), and generate candidate cell states.

[0026] The second step is to update the current cell state based on the outputs of the forget gate and the input gate.

[0027] The third step is to calculate the current hidden state based on the updated cell state and output gate.

[0028] Fourth step: Pass the current hidden state and cell state to the next time step, and use the current hidden state as the output of the first LSTM module.

[0029] The first LSTM module (sLSTM module) is a short-term memory module that combines scalar memory, scalar updates, and new memory.

[0030] Furthermore, the sLSTM module includes: an input gate, a forget gate, an output gate, cell states, and hidden states. The input gate controls the flow of input information into the storage cells. The forget gate determines the amount of information to retain or forget from the previous storage cell state. The output gate controls the flow of information from the storage cells to the hidden states. The cell states store long-term memory information. The hidden states represent the output of the sLSTM at each time step.

[0031] sLSTM enhances the model's ability to extract data by enabling the model to modify storage decisions and improve the performance of tasks that require state tracking.

[0032] The workflow of the second LSTM module includes: The first step is to linearly map the input vector into a query vector, a key vector, and a value vector.

[0033] The second step is to update the cell state matrix to obtain the matrix memory unit.

[0034] The third step is to generate input gates and forget gates through an independent gating network.

[0035] The fourth step is to use the query vector to retrieve relevant information from the matrix memory unit and generate the output vector.

[0036] Step 5: Modulate the output vector through the output gate, and combine the modulated vector with the input vector to generate the final hidden state.

[0037] Step 6: Use the final hidden state as the output of the second LSTM module.

[0038] Furthermore, the second LSTM module (mLSTM module) includes: an input gate, a forget gate, an output gate, cell states, keys, values, and queries. The input gate controls the flow of input information into the memory cells. The forget gate determines the amount of information to retain or forget from the previous memory cell state. The output gate controls the flow of information from the memory cells to the hidden states. The cell states store long-term memory information as a matrix. Keys and values ​​represent the input information to be stored in the memory matrix. Queries retrieve information from the memory matrix.

[0039] Specifically, mLSTM enhances the model's ability to extract data by increasing storage capacity and enabling parallel computing.

[0040] The combination of the first and second LSTM modules effectively extracts the temporal correlations between time points and the long-term dependencies between time periods, allowing the model to capture both local and global temporal features.

[0041] The Transformer module consists of an encoder and a decoder connected in sequence. The encoder uses a multi-head attention mechanism. The encoder processes the query Q, key K, and value V through h linear mapping layers to obtain the transformed vector. The similarity weights are calculated using the transformed vectors, and the similarity weights are weighted and summed to obtain the encoder output.

[0042] The calculation process for the encoder's relevant attention mechanism is as follows:

[0043]

[0044] in, d k Let be the dimension of the key vector. It is the output of attention. This indicates that attention should be paid to the number of heads. Indicates the first The number of heads to pay attention to. The outputs of multiple attention heads are concatenated according to their feature dimensions to form a long vector. This is the linear mapping matrix used for training. The feedforward layer performs a non-linear transformation on the output of the multi-head attention layer, enhancing the model's learning and expressive capabilities.

[0045] The decoder identifies dependencies in the encoder's output, processes non-linear relationships, and obtains the output of the Transformer module.

[0046] Furthermore, the xLSTM-Transformer model combines the temporal processing capabilities of xLSTM with the transformer's ability to better capture temporal and long-range dependencies when processing sequence data. Using the constructed xLSTM-transformer model, GNSS pseudo-measurements are predicted under GNSS outage conditions.

[0047] S4. Supervised learning of the initial deep learning model is performed using the second dataset to obtain a deep learning model. The deep learning model is used to predict the pseudo-measurement values ​​of the global navigation satellite system. The pseudo-measurement values, the output data of the inertial navigation system, and the ranging values ​​output by the ultra-wideband sensor are input into the federated Kalman filter to calculate the error. The output data of the inertial navigation system is corrected based on the error.

[0048] Pseudo-measurements are a well-known term in the art for GNSS observations predicted by deep learning models. In this invention, pseudo-measurements are the model's predictions of UAV location information during GNSS outages.

[0049] An xLSTM-transformer neural network model is used to predict GNSS pseudo-measurements during GNSS interruption. The GNSS pseudo-measurements predicted by the model are combined with INS / UWB integrated navigation to correct INS errors and compensate for the impact of GNSS signal interruption on navigation performance.

[0050] Specifically, in the event of GNSS interruption, the model-predicted GNSS pseudo-measurements are combined with INS / UWB technology to compensate for the performance of GNSS / INS / UWB integrated navigation, correcting INS errors and compensating for the impact of GNSS signal interruption on UAV navigation performance. An experimental platform for UAV cooperative navigation was built, consisting of three master node UAVs, three slave node UAVs, and a ground control station. The master nodes are equipped with high-precision GNSS / INS integrated navigation equipment and UWB ranging equipment, while the slave nodes are equipped with low-precision GNSS and low-precision INS navigation equipment. The UWB ranging equipment on the slave nodes enables ranging between different master and slave nodes. Therefore, UAV navigation only processes the sensor data on the slave nodes, thereby improving navigation accuracy and reliability.

[0051] Furthermore, in the event of GNSS signal interruption, a trained xLSTM-transformer model is used to predict position information. Subsequently, the position information is input into a federated Kalman filter to estimate the position error. Finally, the position error is used to correct the INS error, yielding the integrated navigation output, thus solving the UAV integrated navigation problem in complex environments.

[0052] This invention proposes to combine the advantages of GNSS, INS and UWB, adopting a GNSS / INS / UWB integrated navigation system and using the federated Kalman filter algorithm as the data fusion algorithm to achieve high-precision continuous positioning of UAVs.

[0053] The GNSS / INS / UWB integrated navigation system employs federated filters as a means of multi-source information fusion, establishes state equations and measurement equations, and designs federated filters.

[0054] Specifically, the state equation of a GNSS / INS / UWB integrated navigation system can be defined as:

[0055] in, Represents system state variables. , , and These are position error, velocity error, and attitude error, respectively. and These are the gyroscope and accelerometer deviations, respectively. Represents a process function. Represents the process noise distribution matrix. Represents the noise vector. This is the output of the system state equations.

[0056] The specific expressions for the correlation matrix and vectors are as follows:

[0057]

[0058]

[0059]

[0060] in, The state transition moment for the INS error term, , and The noise of the gyroscope in three directions. , and For the noise of the accelerometer in three directions, , and These are yaw angle, pitch angle, and roll angle, respectively. This is the transformation matrix from the navigation coordinate system to the body coordinate system. The transition matrix representing the relationship between the inertial device and navigation parameters is as follows:

[0061] The federated filter used has two sub-filters, namely the first sub-filter and the second sub-filter. Therefore, the measurement equation can be composed of two parts.

[0062] The first error vector is obtained by subtracting the output data of the inertial navigation system from the pseudo-measurement value. A first sub-filter is then constructed using this first error vector. Kalman filtering is then performed based on this first sub-filter to update the measurements and obtain the first error. The measurement equation for the first sub-filter is as follows:

[0063] in, The output of the measurement equation for the first sub-filter. This is the first measurement matrix. This is the first error vector. To measure noise, the first sub-filter is used for absolute navigation calculation of the slave device. It adopts a loose combination strategy of GNSS and INS to achieve data fusion, and uses the position difference between INS and GPS as the observation.

[0064] Based on the high-precision position of the host computer, the output data of the inertial navigation system, and the ranging value output by the ultra-wideband sensor, a second error vector is calculated. A second sub-filter is constructed based on the second error vector. Kalman filtering is then performed on the measurement update based on the second sub-filter to obtain the second error. The measurement equation of the second sub-filter is as follows:

[0065] in, The output of the measurement equation for the second sub-filter. This is the second measurement matrix. This is the second error vector. It is Gaussian noise.

[0066] Based on the independent Kalman filtering prediction and measurement updates performed by the two sub-filters, their respective error estimates and covariances are obtained. Then, the federated filter fuses the error estimates of the two sub-filters into a global error using covariance weighting to correct the INS error, thereby obtaining a more accurate and reliable integrated navigation error compensation result. The first and second covariance matrices are fused into a final covariance matrix. Using this final covariance matrix, the first and second errors are fused into the final error using covariance weighting. The formula for calculating the final covariance matrix is ​​as follows:

[0067] in, The final covariance matrix, Let be the first covariance matrix. Let be the second covariance matrix, and the formula for calculating the error is:

[0068] in, For error, The first error, The second error is calculated using the first sub-filter, which obtains the first error and the first covariance matrix based on the INS output and the pseudo-measured values. The second sub-filter obtains the second error and the second covariance matrix based on the host high-precision position, INS output, and UWB ranging data.

[0069] The GNSS / INS / UWB integrated navigation system uses the federated Kalman filter algorithm as the data fusion algorithm. The detailed operation steps of the algorithm are as follows: estimation of covariance and initialization of noise covariance; time update of each sub-filter in the main filter; independent measurement update of each sub-filter; global optimal fusion; information allocation; and information fusion at the next time step after information allocation.

[0070] UAV 4 was selected as the experimental subject. The xLSTM-transformer model was used to predict GNSS pseudo-measurements of UAV 4 and to compensate the integrated navigation system in case of GNSS interruption. A schematic diagram of the flight trajectory of UAV 4 is shown below. Figure 3 As shown. For comparison, For speed, and For the yaw angle, select f n , v n , The three parameters are used as input to the model. For ease of representation, f n , v n , Combined into input data Output results This is GNSS location information.

[0071] When GNSS signals are available For input data, The output data is input into the model, then the xLSTM-transformer model is trained and saved, and the results are found. f n , v n , and The relationship between them.

[0072] When GNSS signals are lost, a trained xLSTM-transformer model is used to predict position information. This position information is then fed into a federated Kalman filter to estimate the position error. Finally, the position error is used to correct the INS error, yielding the integrated navigation output.

[0073] S5. Based on the output data of the corrected inertial navigation system, realize the integrated navigation of the UAV.

[0074] Example This embodiment provides a method for compensating for the performance of integrated navigation under GNSS interruption. (Used in this embodiment) The dataset is divided into a training set and a test set. The training set is used for optimizing and learning the model parameters, while the test set is used to evaluate the model's prediction accuracy and generalization performance.

[0075] In the process of building machine learning models, data normalization is a crucial preprocessing step. The min-max normalization method is used to process the data, transforming... Data features linearly mapped to [ Within the interval [1, 1]. Scale the data to the range [0, 1] using the following formula:

[0076] in, It is normalized data. It is the raw data. and These are the minimum and maximum values ​​in the dataset, respectively. After the prediction model outputs and the prediction results are displayed, the prediction results should be denormalized. The denormalization formula is as follows:

[0077] in, for t Predicted value at time, and These are the minimum and maximum values ​​in the dataset, respectively. x This is the output obtained through inverse normalization. Normalization effectively eliminates the impact of data dimensionality differences on the model, while significantly improving the model's convergence speed.

[0078] After data preprocessing, the normalized data is input into the model for training and prediction. Features are extracted using an xLSTM model, and a Transformer model is used to capture dependencies within the sequence of data. Finally, the model-predicted GNSS pseudo-measurements are combined with INS / UWB navigation to correct INS errors.

[0079] This embodiment uses a total of six UAVs. UAVs 1, 2, and 3 are designated as master units, and UAVs 0, 4, and 5 as slave units. The GNSS / INS / UWB integrated navigation and positioning system is mainly responsible for the acquisition of information from each sensor and the solution of the integrated navigation and positioning algorithm. It consists of a micro-electro-mechanical system (MEMS) inertial sensor, a UWB ranging sensor, a GNSS receiver, communication sensors, and a core processor. The MEMS unit uses an ADIS16460 IMU. The GNSS unit uses a ublox ZED-F9P RTK GNSS module manufactured by Beitian Communication. The UWB unit uses a Decawave 1000. The UAV sampling time is 10 minutes. The system is time-synchronized based on the 1-second pulse (PPS) signal output by the satellite timing module. The INS sampling rate is 50 Hz, the UWB sampling rate is 10 Hz, and the GNSS sampling rate is 1 Hz.

[0080] This embodiment employs an xLSTM-transformer prediction model for INS error compensation. The xLSTM model extracts features from the data, while the transformer model captures dependencies within the sequence. All experiments were conducted on a 2.42 GHz quad-core Intel® Core processor. TMRunning on an i5-1135G7 processor, Windows 10 64-bit, with Matlab 2025a as the operating system.

[0081] When GNSS signals are lost, a trained xLSTM-transformer model is used for GNSS position prediction. At this time, the proposed method is employed for INS error compensation.

[0082] This embodiment uses two mLSTM modules and one sLSTM module to extract features from the data, selects the ReLU function as the activation function, and chooses Adam as the optimizer. The order of the modules is pseudo-input-mLSTM-sLSTM-mLSTM-output. The regularization parameter is 0.001, the learning rate is 0.01, and the maximum number of training iterations is 150.

[0083] The mean absolute error (MAE), root-mean-square error (RMSE), mean square error (MSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) were used. 2 Indicators such as ( ) are used to evaluate the predictive performance of the model. The calculation formula is as follows:

[0084]

[0085]

[0086]

[0087]

[0088] in, This represents the predicted values ​​from the first data point to the nth data point. This represents the actual values ​​of the first to nth data points. n The sample size is represented by the value of MAE, RMSE, MSE, and MAPE. Lower values ​​for MAE, RMSE, MSE, and MAPE indicate higher prediction accuracy. (Coefficient of Determination) R 2 The closer the value is to 1, the better the model fits the sample data. A diagram illustrating the prediction results based on the xLSTM-transformer model is shown below. Figure 4 As shown.

[0089] This embodiment aims to solve the problem of error divergence in the GNSS / INS / UWB integrated navigation system when the GNSS system is interrupted. For example, the distance error curve of a UAV is shown in the figure below. Figure 5 This embodiment demonstrates that the xLSTM-transformer model can predict GNSS pseudo-measurements during GNSS interruptions, correct INS errors, and solve the problem of UAV integrated navigation in complex environments.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for compensating for the performance of a combined navigation under GNSS outage, characterized in that, The method comprises the following steps: obtaining a data set comprising observation data of a global navigation satellite system, output data of an inertial navigation system and ranging values output by an ultra-wideband sensor; randomly dividing the data set into a first data set and a second data set; constructing an initial deep learning model using the first data set, the initial deep learning model comprising connected extended long short-term memory modules and a Transformer module; performing supervised learning on the initial deep learning model using the second data set to obtain a deep learning model, the deep learning model being used to predict pseudo-measurement values of the global navigation satellite system, the pseudo-measurement values, the output data of the inertial navigation system and the ranging values output by the ultra-wideband sensor being input into a federated Kalman filter to calculate errors, and the output data of the inertial navigation system being corrected according to the errors; realizing combined navigation of a UAV according to the corrected output data of the inertial navigation system.

2. The method of claim 1, wherein, The working process of the initial deep learning model comprises: inputting input data into the extended long short-term memory modules, the extended long short-term memory modules comprising a first LSTM module and a second LSTM module to obtain output of the extended long short-term memory modules; inputting the output of the extended long short-term memory modules into the Transformer module to obtain pseudo-measurement values.

3. The method of claim 2, wherein, The working process of the first LSTM module comprises: receiving a current input vector and a hidden state of a previous time step, processing through three gate structures of an input gate, a forget gate and an output gate, and generating a candidate cell state; updating a current cell state according to outputs of the forget gate and the input gate; calculating a current hidden state based on the updated cell state and the output gate; passing the current hidden state and the cell state to a next time step, and taking the current hidden state as output of the first LSTM module.

4. The method of claim 2, wherein, The working process of the second LSTM module comprises: linearly mapping an input vector into a query vector, a key vector and a value vector at the same time; updating a cell state matrix to obtain a matrix memory unit; generating an input gate and a forget gate through an independent gating network; retrieving relevant information from the matrix memory unit using the query vector to generate an output vector; modulating the output vector through an output gate, combining the modulated vector with the input vector, and generating a final hidden state; taking the final hidden state as output of the second LSTM module.

5. The method of claim 1, wherein, The Transformer module comprises a sequentially connected encoder and a decoder, the encoder adopts a multi-head attention mechanism, the encoder processes queries, keys and values through a linear mapping layer to obtain transformed vectors, calculates similarity weights using the transformed vectors, performs weighted summation on the similarity weights to obtain output of the encoder; the decoder identifies dependency relationships in the output of the encoder, processes nonlinear relationships, and obtains output of the Transformer module.

6. The method of claim 1, wherein, The inputting of the pseudo-measurement values, the output data of the inertial navigation system and the ranging values output by the ultra-wideband sensor into the federated Kalman filter to calculate errors comprises: The federal Kalman filter comprises a first sub-filter and a second sub-filter; The output data of the inertial navigation system is subtracted from the pseudo-measurement value to obtain a first error vector, the first error vector is used to construct the first sub-filter, measurement update of Kalman filtering based on the first sub-filter is performed to obtain a first error and a first covariance matrix, and a measurement equation of the first sub-filter is as follows: wherein, is an output of a measurement equation of the first sub-filter, is a first measurement matrix, is a first error vector, is a measurement noise; A second error vector is calculated based on the high-precision position of the host, the output data of the inertial navigation system and the ranging value output by the ultra-wideband sensor, a second sub-filter is constructed based on the second error vector, measurement update of Kalman filtering based on the second sub-filter is performed to obtain a second error and a second covariance matrix, and a measurement equation of the second sub-filter is as follows: wherein, is an output of a measurement equation of the second sub-filter, is a second measurement matrix, is a second error vector, is a Gaussian noise; The first covariance matrix and the second covariance matrix are fused into a final covariance matrix, the first error and the second error are fused into the error through covariance weighting by using the final covariance matrix, and a calculation formula of the final covariance matrix is as follows: wherein, is the final covariance matrix, is the first covariance matrix, is the second covariance matrix, the error being calculated according to the formula: wherein is an error, is a first error, is a second error.

7. The method of claim 1, wherein, The method comprises a training phase and a compensation phase, The training phase comprises: when global navigation satellite system signals are available, using output data of an inertial navigation system as input of a deep learning model and using measurement values of the global navigation satellite system as supervision labels to train the deep learning model; The compensation phase comprises: when global navigation satellite system signals are interrupted, inputting pseudo-measurement values predicted by the trained deep learning model into a federal Kalman filter to correct errors of the inertial navigation system and compensate navigation performance.

Images