A combined navigation method based on mamba and attention mechanism

Abstract

The application discloses a kind of based on Mamba and attention mechanism combined navigation method, belong to GNSS / INS combined navigation positioning technical field.The method includes: using InsMamba module to the IMU measurement sequence is iteratively modeled in two-way state space, and the IMU state variation is output;IMU state variation is mapped as query feature, and GNSS observation ephemeris interval variation is mapped as key feature and value feature, and through cross attention mechanism, heterogeneous feature nonlinear fusion is carried out;Using GNSS signal quality index generates confidence factor by modulating network, and the displacement and velocity variation of current ephemeris are obtained by adaptively weighting fusion of fusion feature and IMU state variation;The current positioning result is obtained by combining the navigation information recursion of last time.The application significantly improves the continuity and precision of combined navigation under the dynamic change environment of GNSS signal quality.

Description

Technical Field

[0001] This invention belongs to the field of GNSS / INS integrated navigation and positioning technology, specifically relating to an integrated navigation method based on Mamba and attention mechanisms. Background Technology

[0002] Global Navigation Satellite System (GNSS) calculates receiver position by measuring satellite signal propagation time. It features global coverage, all-weather operation, and high accuracy, playing a vital role in transportation, surveying, and consumer electronics. However, GNSS signals rely on open environments. In urban canyons, tunnels, indoor spaces, or under conditions of strong interference, signals are prone to interruption, attenuation, or multipath effects, leading to a significant decrease in positioning accuracy or even failure.

[0003] Inertial Navigation Systems (INS) do not rely on external signals. They measure acceleration and angular velocity through an Inertial Measurement Unit (IMU) and integrate these measurements to calculate the vehicle's position, velocity, and attitude. INS offers strong autonomy and high short-term accuracy, providing continuous navigation information even in environments with limited GNSS signals. However, a fundamental drawback of INS is that errors accumulate over time, leading to a divergent decline in positioning accuracy, making it unsuitable for long-term independent operation.

[0004] Given their complementarity, GNSS / INS integrated navigation has become the mainstream solution for high-precision navigation. Traditional integrated navigation mainly uses Kalman filtering (KF) and its derivative algorithms for data fusion. Extended Kalman Filter (EKF) handles nonlinear problems through local linearization, Unscented Kalman Filter (UKF) uses unscented transformation to more accurately propagate the mean and covariance, and Particle Filter (PF) handles arbitrary nonlinear non-Gaussian systems using Monte Carlo methods. These methods have achieved great success in the field of integrated navigation.

[0005] However, in real-world complex navigation environments, GNSS signal obstruction, multipath effects, and transient interference often result in observation noise exhibiting long-tailed, non-Gaussian, or even multi-peaked distribution characteristics. Traditional Kalman filtering methods, based on the linear Gaussian assumption, struggle to accurately model such complex errors, leading to decreased filter performance or even divergence. Neural networks, as nonlinear modeling tools, can adaptively learn complex mapping relationships from massive amounts of data without requiring pre-defined precise physical models or noise distribution assumptions, providing a new approach to overcoming the limitations of traditional integrated navigation algorithms. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a combined navigation method based on Mamba and an attention mechanism. This method first proposes an InsMamba model to deeply process IMU data, efficiently extracting and updating the vehicle's motion state to generate IMU state representations. Subsequently, these IMU states serve as queries (Q) for the cross-attention mechanism, while GNSS observation data, after encoding, serves as keys (K) and values ​​(V), enabling intelligent fusion and updating of IMU predictions with GNSS measurements. Finally, the output of the cross-attention network is dynamically fused with the IMU states generated by InsMamba. The adaptive fusion network dynamically generates a modulation parameter α by inputting the quality characteristics of the GNSS signal—signal-to-noise ratio and geometrical dilution of precision (GDOP)—into a multilayer perceptron and a sigmoid function. This parameter is used to weight GNSS features (weight α) and IMU features (weight 1-α), thereby adaptively adjusting the fusion ratio according to the real-time reliability of the GNSS signal, ensuring that the system consistently outputs accurate, continuous, and robust combined navigation and positioning results even in complex and harsh environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A combined navigation method based on Mamba and attention mechanisms includes:

[0009] Step 1: Input the position and velocity changes between observation epochs of the Global Navigation Satellite System, as well as the specific force and angular acceleration sequences measured by the inertial measurement unit, into the neural network;

[0010] Step 2: Use the bidirectional Mamba module to perform state-space iterative modeling on the specific force and angular acceleration sequences, and output the position and velocity changes calculated by the inertial navigation system;

[0011] Step 3: Generate query features from the position and velocity changes calculated by the inertial navigation system through linear mapping; generate key and value features from the position and velocity changes between observation epochs of the global satellite navigation system through linear mapping; and fuse the query features, key features, and value features through a cross-attention mechanism to obtain fused features.

[0012] Step 4: Generate confidence factors based on the quality characteristics of the global satellite navigation system, and use the confidence factors to adaptively weight and fuse the fused features with the position and velocity changes calculated by the inertial navigation system to obtain the position and velocity changes at the current observation epoch.

[0013] Step 5: Add the navigation position and velocity from the previous moment to the position and velocity changes at the current observation epoch to obtain the positioning result and vehicle velocity at the current moment.

[0014] Furthermore, in step 2, the bidirectional Mamba module dynamically generates a discretized step size, an input projection matrix, and an output projection matrix based on the input specific force and angular acceleration data at the current moment, and uses the discretized step size, the input projection matrix, and the output projection matrix to perform selective memorization and forgetting on the specific force and angular acceleration sequence.

[0015] Furthermore, the bidirectional Mamba module includes a forward Mamba module and a reverse Mamba module. The forward Mamba module processes the specific force and angular acceleration sequences in ascending time order, while the reverse Mamba module processes the specific force and angular acceleration sequences in reverse time order. The output features of the forward Mamba module and the output features of the reverse Mamba module are added together, and after layer normalization and residual connection processing, the position change and velocity change calculated by the inertial navigation system are output.

[0016] Furthermore, in step 3, the cross-attention mechanism determines the fusion weight of the value feature by matching the similarity between the query feature and the key feature, thereby obtaining the fused feature.

[0017] Furthermore, in step 4, the confidence factor is generated based on the quality characteristics of the global navigation satellite system as follows: the signal-to-noise ratio, geometric accuracy factor, and number of visible satellites of the global navigation satellite system are input into the modulation network, and the confidence factor is generated through a multilayer perceptron and a sigmoid function. The confidence factor is used to characterize the confidence weight of the observation information of the global navigation satellite system.

[0018] Furthermore, in step 4, the confidence factor is used as a weight to sum the results of the fused features and the position and velocity changes calculated by the inertial navigation system after being mapped by the multilayer perceptron, respectively, to obtain the position and velocity changes of the current observation epoch.

[0019] Furthermore, the method also includes freezing network parameters other than the bidirectional Mamba module, and calculating the mean square error between the position and velocity changes calculated by the inertial navigation system output by the bidirectional Mamba module and the position and velocity changes of the high-precision reference system, in order to train the parameters of the bidirectional Mamba module.

[0020] Furthermore, after the bidirectional Mamba module is trained, the mean square error of the position and velocity changes of the current observation epoch output by the neural network is calculated with the position and velocity changes of the high-precision reference system, and the parameters of the entire neural network are trained in an end-to-end manner.

[0021] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned combined navigation method based on Mamba and attention mechanism.

[0022] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned combined navigation method based on Mamba and attention mechanism.

[0023] The beneficial effects of this invention are as follows:

[0024] This invention uses the InsMamba model to perform bidirectional state-space modeling of high-frequency motion of IMU. Its parameters are dynamically generated from the input data, enabling selective memory and forgetting. It can efficiently capture complex temporal dependencies and motion accumulation errors in long IMU sequences. Compared with traditional LSTM, Transformer and other models, it provides better inertial estimation capabilities with linear computational complexity.

[0025] This invention employs a cross-attention mechanism to achieve deep nonlinear fusion of heterogeneous features from IMU and GNSS. Using IMU status as the query and GNSS observations as the key and value, adaptive information selection is achieved through dynamic alignment and weighting. This overcomes the dependence of traditional Kalman filtering on the linear Gaussian assumption and improves the fusion accuracy in complex noise environments.

[0026] This invention is the first to explicitly introduce GNSS signal quality indicators (signal-to-noise ratio, geometric accuracy factor, and number of visible satellites) into the fusion network. By dynamically generating confidence factors through the modulation network and adaptively adjusting the fusion weights of GNSS and IMU, data-driven GNSS quality control is achieved, which significantly improves the robustness and positioning continuity of the system in complex environments where signal quality changes dynamically.

[0027] This invention adopts an end-to-end learning framework, which eliminates the need for manual feature engineering and complex model tuning. Its modular design facilitates the integration of more sensors (such as wheel speedometers, lidar, etc.), and it has good versatility and scalability. Attached Figure Description

[0028] Figure 1This is a flowchart of a combined navigation method based on Mamba and attention mechanism according to the present invention;

[0029] Figure 2 This is a diagram of the bidirectional Mamba network structure of the present invention;

[0030] Figure 3 This is a diagram of the adaptive fusion network structure of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] This invention proposes a combined navigation method based on Mamba and attention mechanisms, specifically, as follows: Figure 1 As shown, the method includes:

[0033] Step 1: Input the position and velocity changes between observation epochs of the Global Navigation Satellite System, as well as the specific force and angular acceleration sequences measured by the inertial measurement unit, into the neural network;

[0034] Step 2: Use the bidirectional Mamba module to perform state-space iterative modeling on the specific force and angular acceleration sequences, and output the position and velocity changes calculated by the inertial navigation system;

[0035] Step 3: Generate query features from the position and velocity changes calculated by the inertial navigation system through linear mapping; generate key and value features from the position and velocity changes between observation epochs of the global satellite navigation system through linear mapping; and fuse the query features, key features, and value features through a cross-attention mechanism to obtain fused features.

[0036] Step 4: Generate confidence factors based on the quality characteristics of the global satellite navigation system, and use the confidence factors to adaptively weight and fuse the fused features with the position and velocity changes calculated by the inertial navigation system to obtain the position and velocity changes at the current observation epoch.

[0037] Step 5: Add the navigation position and velocity from the previous moment to the position and velocity changes at the current observation epoch to obtain the positioning result and vehicle velocity at the current moment.

[0038] Furthermore, in step 1:

[0039] The navigation data received by the GNSS / INS integrated navigation system is the navigation data calculated from the GNSS signal using pseudorange and Doppler. , ,in express The location of the carrier in real-time GNSS positioning. This indicates the corresponding velocity; IMU observation information is... , Indicates triaxial force information. This represents the three-axis angular acceleration. The amount of IMU observation information between a single observation epoch is related to the sampling frequency.

[0040] Considering the dead reckoning principle of IMUs, the measurement information from an IMU within a single epoch can only reflect the changes in velocity, position, and attitude within that epoch. Therefore, GNSS observation information is converted into changes in position and velocity between observation epochs, constructing input data with the same logic as the IMU. In this case, the GNSS input data consists of the changes in position and velocity between epochs. , The input data for the IMU is the sequence of specific force and angular acceleration measurements within the current epoch. The above information is input into the neural network. The core module of the neural network in this invention is the Mamba module.

[0041] Furthermore, such as Figure 2 As shown, in step 2:

[0042] To fully exploit the high-frequency dynamics, temporal dependencies, and inherent motion patterns in IMU navigation information, this invention utilizes the Mamba model for fitting and reconstruction. IMU data is input into the InsMamba module to achieve inertial navigation mechanical orchestration and updates. The kinematics of an IMU is essentially a continuous-time dynamic system. The InsMamba network (i.e., bidirectional Mamba), through its state-space model structure, can learn and fit the inherent dynamics of this physical system, constructing a stable navigation information processing model. The IMU measurement sequence... The InsMamba module outputs the fitted position and velocity changes. , Specifically:

[0043] First, through the linear layer Map the input specific force and angular acceleration into high-dimensional features. :

[0044] (1)

[0045] Obtain high-dimensional features Then, the forward I-Mamba module performs successive scanning of the input information and iterative updates of the state-space model. High-dimensional features. Within the forward I-Mamba, the linear layers first traverse different branches:

[0046] (2)

[0047] (3)

[0048] and These represent the outputs of the first and second branches, respectively.

[0049] Then, the output and input of the first and second branches are respectively activated by the Silu activation functions of their respective branches:

[0050] (4)

[0051] (5)

[0052] This indicates the result of the output of the first branch passing through the corresponding Silu activation function; This indicates the result of the output of the second branch passing through the corresponding Silu activation function;

[0053] Will The input state-space model is used for iterative modeling of time-series data, and it serves as the input. Unlike traditional fixed-parameter SSMs, Mamba's discretization step size... Input projection matrix and output projection matrix It is dynamically calculated from the current IMU input data, enabling the model to adaptively adjust its memory, writing, and retrieval methods based on the specific content and context of the IMU measurements. The state-space model data processing flow is as follows:

[0054] (6)

[0055] (7)

[0056] in, Indicates time The hidden state, Indicates hidden state The first derivative with respect to time, Represents continuous-time variables in an IMU measurement sequence; It is the state transition matrix, used to describe the evolution of the hidden state over time; It is the input projection matrix, responsible for inputting the current IMU features. It is mapped to the hidden state space and participates in the hidden state update. It is the output projection matrix, used to project the hidden state. Mapped to the output feature space. It skips connection parameters, allowing IMU feature input. Skipping the hidden state directly affects the output, which helps to enhance the model's expressive power and gradient propagation. The right-hand side of equation (7) This represents the state-space output item generated from the hidden state. This represents the skipped connection output term directly generated from the current IMU input. The sum of the two yields the state-space model at time [time]. Output .

[0057] For a practical discrete IMU sampling system, equations (6) and (7) can be discretized as follows:

[0058] (8)

[0059] (9)

[0060] (10)

[0061] (11)

[0062] in, This represents the discretization step size, used to convert a continuous-time state-space model into a discrete-time state-space model. This represents the discretized state transition matrix, used to describe the recursive relationship of the hidden states between adjacent IMU sampling times; This represents the continuous-time input projection matrix. This represents the discretized input projection matrix, used to map the current IMU feature input to the hidden state space; Represents the identity matrix; Indicates matrix exponentiation; Indicates the first Discrete hidden states at each IMU sampling time This represents the discrete hidden state at the previous IMU sampling time. Indicates the first Discrete IMU feature inputs at each IMU sampling time; Indicates the first State-space model output at each IMU sampling time; This represents the discretized output projection matrix, used to map the hidden state to the output feature space; The skip connection parameters represent the discretized values ​​used to allow the current IMU feature input to directly participate in the output computation. These discrete equations describe how the state transition space iteratively updates its internal state and generates the output at each IMU sampling point, based on the memory of the previous time step and the current input. By learning dynamic parameters, Mamba can better fit the noise characteristics and kinematic cumulative error patterns in IMU data, providing smoother and more accurate inertial extrapolation, and offering high-quality query features for subsequent fusion. The output is obtained. Then with Performing element-wise multiplication, also known as the Hadamard product. :

[0063] (12)

[0064] then The positive Mamba output features are obtained through a linear layer. :

[0065] (13)

[0066] The reverse I-Mamba (Reversed) is the counterpart to forward I-Mamba. The data is input in reverse order to an I-Mamba network with the same structure to obtain the reverse Mamba output, characterized by: Forward I-Mamba processes the IMU sequence in chronological order, capturing dependencies from the past to the present. Reverse I-Mamba reverses the IMU sequence and processes it to capture dependencies from the future to the past.

[0067] The outputs from both directions are then fused to generate a feature representation for each IMU data point that contains richer contextual information. This bidirectional processing approach aligns particularly well with the logic of inertial navigation systems for processing incremental information, enabling the model to more accurately infer the vehicle's trajectory between two GNSS observation points. Specifically, the forward Mamba output features... Features of inverse Mamba output Addition, normalization Normalized features were then obtained. Then, after passing through residual blocks and linear layers, the final calculated position and velocity changes of the inertial navigation system are obtained. , :

[0068] (14)

[0069] (15)

[0070] Furthermore, such as Figure 3 As shown, in step 3:

[0071] The epochal variation of GNSS measurements is obtained through the feature mapping module. , With InsMamba output , These are mapped to the feature space respectively. In the cross-attention mechanism module, , The mapped features are used as query Q. , The mapped features serve as keys K and values ​​V, and an attention mechanism is used to achieve nonlinear adaptive fusion of the two through dynamic weighting. Specifically, Q is generated by linear mapping from the IMU recursive information output by the InsMamba module, representing the IMU's prediction of the current position and velocity based on its own calculations, implicitly requiring external absolute correction. , Generated from GNSS positioning observation information through linear mapping, it includes absolute position / velocity information provided by GNSS and its corresponding context or potential mass information:

[0072] (16)

[0073] (17)

[0074] (18)

[0075] The core of the cross-attention mechanism is the dynamic alignment and weighting of information from two different sources: IMU (Q) and GNSS (K, V). The query features of the IMU are matched with the key features of the GNSS based on similarity. The higher the similarity, the greater the weight of the corresponding GNSS value feature. This allows IMU recursive information to intelligently select and incorporate necessary absolute positioning corrections from GNSS features. It's not a simple linear combination, but rather a non-linear fusion based on the complex relationship between the two features. For example, when IMU calculations and GNSS observations are highly consistent, the model assigns greater correction weight to GNSS; conversely, when there are significant differences (e.g., when GNSS changes abruptly), the model can learn to reduce its reliance on GNSS and place more trust in the smooth calculations from the IMU. , , The input cross-attention mechanism module performs the feature fusion steps:

[0076] (19)

[0077] in, Key feature matrix Feature dimensions, For the obtained fusion features, the superscript T denotes matrix transpose, and softmax denotes the normalization exponential function, used to normalize the similarity calculation results to obtain non-negative weights that sum to 1. Specifically, This represents the scaled similarity between query features and key features. After softmax processing, attention weights are obtained and then compared with the value feature matrix. Multiply to obtain the fusion feature .

[0078] Furthermore, in step 4:

[0079] Obtain fusion features Subsequently, GNSS quality characteristics are introduced, and the adaptive fusion module regulates the fusion positioning process of the integrated navigation features based on the adaptive factors generated by the quality characteristics. Finally, the position and velocity changes at the current observation epoch are obtained. , Specifically:

[0080] Utilizing GNSS quality characteristics: Geometric Precision Factor (GDOP), Signal-to-Noise Ratio Number of visible satellites As input features, generate quality-related confidence factors. This characterizes the confidence weight of the GNSS fusion features at this point, reflecting the current reliability of the GNSS observations.

[0081] (20)

[0082] In the formula, This represents the activation function. This represents a multilayer perceptron.

[0083] Confidence factors generated It will be used to dynamically control the fusion process of the cross-attention mechanism or its subsequent fusion layers. For example, when the GNSS signal quality is good, the confidence factor... The system allows GNSS to perform strong corrections to the IMU; when GNSS signal quality deteriorates (e.g., in urban canyons or areas with signal obstruction), the confidence factor... As the GNSS level decreases, the system will correspondingly reduce its reliance on GNSS, relying more on IMU calculations to maintain positioning continuity and smoothness. This data-driven adaptive adjustment ensures the optimal fusion strategy under various complex environments. The feature fusion steps are then as follows:

[0084] (twenty one)

[0085] In the formula, This represents the change in position and velocity at the current observation epoch; This represents the fusion features obtained through the cross-attention mechanism; , This represents the joint vector composed of the position and velocity changes calculated by the InsMamba module. and Both represent the mapping function of the multilayer perceptron, where Used to map fused features to changes in position and velocity. Used to map inertial calculation results to output results of the same dimension; This is the confidence factor generated from the GNSS quality characteristics in formula (20). The higher the current GNSS observation quality, the better. The larger the value, the greater the contribution of the fused features to the final output; the lower the current GNSS observation quality, the greater the contribution of the fused features to the final output. The smaller the value, the greater the contribution of the inertial calculation result to the final output.

[0086] Furthermore, in step 5:

[0087] Positioning results at time t and carrier velocity , :

[0088] (twenty two)

[0089] The InsMamba network is trained in the following way in this invention:

[0090] Step 1: Train the InsMamba network to fit INS state information:

[0091] First, the InsMamba network parameters are trained to fit the state information of INS. The parameters of other parts of the overall network are frozen. The loss function is... :

[0092] (twenty three)

[0093] in, This represents the mean squared error. These represent the changes in position and velocity of the high-precision reference system at time t during data acquisition.

[0094] Step 2: Train the overall network for integrated navigation and positioning:

[0095] At this point, the overall network parameters are trained, and the loss function is... :

[0096] (twenty four)

[0097] In summary, this invention constructs a combined navigation method based on Mamba and an attention mechanism. It utilizes a bidirectional Mamba module to selectively model the high-frequency time-series data of the inertial measurement unit (IMU), combines this with a cross-attention mechanism to achieve deep nonlinear fusion of inertial estimation and satellite observation, and introduces global navigation satellite system (GNSS) signal quality indices for adaptive weighted adjustment. This effectively overcomes the limitations of traditional Kalman filtering methods under the linear Gaussian assumption. In complex environments with dynamically changing GNSS signal quality, this invention significantly improves the positioning accuracy, continuity, and robustness of the combined navigation system, providing reliable technical support for high-precision navigation applications such as autonomous driving, drones, and precision agriculture.

[0098] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned combined navigation method based on Mamba and attention mechanism.

[0099] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned combined navigation method based on Mamba and attention mechanism.

[0100] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A combined navigation method based on Mamba and attention mechanisms, characterized in that, include: Step 1: Input the position and velocity changes between observation epochs of the Global Navigation Satellite System, as well as the specific force and angular acceleration sequences measured by the inertial measurement unit, into the neural network; Step 2: Use the bidirectional Mamba module to perform state-space iterative modeling on the specific force and angular acceleration sequences, and output the position and velocity changes calculated by the inertial navigation system; Step 3: Generate query features from the position and velocity changes calculated by the inertial navigation system through linear mapping; generate key and value features from the position and velocity changes between observation epochs of the global satellite navigation system through linear mapping; and fuse the query features, key features, and value features through a cross-attention mechanism to obtain fused features. Step 4: Generate confidence factors based on the quality characteristics of the global satellite navigation system, and use the confidence factors to adaptively weight and fuse the fused features with the position and velocity changes calculated by the inertial navigation system to obtain the position and velocity changes at the current observation epoch. Step 5: Add the navigation position and velocity from the previous moment to the position and velocity changes at the current observation epoch to obtain the positioning result and vehicle velocity at the current moment.

2. The combined navigation method based on Mamba and attention mechanism according to claim 1, characterized in that, In step 2, the bidirectional Mamba module dynamically generates a discretized step size, an input projection matrix, and an output projection matrix based on the input specific force and angular acceleration data at the current moment, and uses the discretized step size, the input projection matrix, and the output projection matrix to perform selective memorization and forgetting on the specific force and angular acceleration sequence.

3. The combined navigation method based on Mamba and attention mechanism according to claim 2, characterized in that, The bidirectional Mamba module includes a forward Mamba module and a reverse Mamba module. The forward Mamba module processes the specific force and angular acceleration sequences in forward time order, while the reverse Mamba module processes the specific force and angular acceleration sequences in reverse time order. The output features of the forward Mamba module and the output features of the reverse Mamba module are added together, and after layer normalization and residual connection processing, the position change and velocity change calculated by the inertial navigation system are output.

4. The combined navigation method based on Mamba and attention mechanism according to claim 1, characterized in that, In step 3, the cross-attention mechanism determines the fusion weight of the value feature by matching the similarity between the query feature and the key feature, thereby obtaining the fused feature.

5. The combined navigation method based on Mamba and attention mechanism according to claim 1, characterized in that, In step 4, the confidence factor is generated based on the quality characteristics of the global navigation satellite system as follows: the signal-to-noise ratio, geometric accuracy factor, and number of visible satellites of the global navigation satellite system are input into the modulation network, and the confidence factor is generated through a multilayer perceptron and a sigmoid function. The confidence factor is used to characterize the confidence weight of the observation information of the global navigation satellite system.

6. The combined navigation method based on Mamba and attention mechanism according to claim 1, characterized in that, In step 4, the confidence factor is used as a weight to sum the results of the fusion features and the position and velocity changes calculated by the inertial navigation system after being mapped by the multilayer perceptron, respectively, to obtain the position and velocity changes of the current observation epoch.

7. The combined navigation method based on Mamba and attention mechanism according to claim 1, characterized in that, The method further includes freezing network parameters other than the bidirectional Mamba module, and calculating the mean square error between the position and velocity changes calculated by the inertial navigation system output by the bidirectional Mamba module and the position and velocity changes of the high-precision reference system, in order to train the parameters of the bidirectional Mamba module.

8. The combined navigation method based on Mamba and attention mechanism according to claim 7, characterized in that, After the bidirectional Mamba module is trained, the position and velocity changes of the current observation epoch output by the neural network are compared with the position and velocity changes of the high-precision reference system to calculate the mean square error, thereby training the parameters of the entire neural network in an end-to-end manner.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the combined navigation method based on Mamba and attention mechanism as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the combined navigation method based on Mamba and attention mechanism as described in any one of claims 1-8.

Images