A robust state estimation method for unmanned aerial vehicles in urban environments

Abstract

The present application relates to the technical field of unmanned aerial vehicle state estimation, and particularly relates to a robust state estimation method for unmanned aerial vehicles based on uncertainty-guided KalmanNet, comprising the following steps: 1) constructing an uncertainty-guided implicit KalmanNet hybrid estimation model; 2) training the implicit KalmanNet hybrid estimation model through unmanned aerial vehicle flight data in an urban environment to obtain an optimal state estimation model; 3) collecting sensor data in real time and generating an uncertainty vector, outputting Kalman gain through model inference, and combining a "prediction-update" framework of Kalman filtering to realize real-time robust estimation of the state of the unmanned aerial vehicle. The present application can realize high-precision real-time state estimation of unmanned aerial vehicles in an urban environment, effectively overcome estimation errors caused by nonlinear dynamics, model mismatch and non-Gaussian measurement disturbances, and provide technical support for autonomous navigation and precise operation of unmanned aerial vehicles in cities.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) state estimation technology, specifically a robust UAV state estimation method based on uncertainty-guided KalmanNet. Background Technology

[0002] Autonomous flight and operation of unmanned aerial vehicles (UAVs) in urban environments have become an important application area in fields such as aerial surveying and mapping, urban inspection, logistics delivery, and emergency rescue. High-precision state estimation is the core foundation for UAVs to achieve autonomous navigation, attitude control, and path planning. Its estimation accuracy directly determines the flight stability, positioning accuracy, and operational safety of the UAV. The urban environment contains many complex factors such as building obstruction and electromagnetic interference, which cause UAVs to face highly nonlinear kinematic dynamics, serious system model mismatch problems, and non-Gaussian measurement disturbances, posing a great challenge to UAV state estimation.

[0003] Currently, the mainstream methods for UAV state estimation are mainly divided into two categories: traditional model basis filtering methods and pure data-driven deep learning methods. Both types of methods have obvious technical defects and are difficult to meet the requirements of robust, high-precision, and real-time state estimation of UAVs in urban environments.

[0004] Traditional model basis filtering methods, represented by Kalman filtering and its nonlinear extensions (such as Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF), achieve recursive state estimation based on explicit system motion and measurement models. Under ideal conditions of model matching and Gaussian noise distribution, they can achieve optimal estimation results. However, in practical applications in urban environments, these methods have significant limitations: firstly, the high maneuverability of UAVs makes it difficult to establish accurate nonlinear motion models; building occlusion can lead to severe mismatch in the measurement model; and local linearization approximations can generate cumulative errors, resulting in a significant decrease in filtering performance. Secondly, these methods are highly dependent on the statistical characteristics of noise; non-Gaussian measurement disturbances can directly reduce estimation accuracy and even cause filter divergence.

[0005] Purely data-driven deep learning methods (such as Recurrent Neural Networks (RNNs) and Gated Recurrent Units (GRUs) achieve estimation by learning state evolution patterns from observed data, without relying on explicit system models, and have a certain degree of adaptability to nonlinear systems. However, these methods also have shortcomings: First, the network model is highly parameterized, requiring a large amount of training data and computational resources, making it difficult to deploy on the resource-constrained airborne embedded hardware of UAVs; second, the model lacks physical interpretability, and the state estimation process is a "black box" mapping, unable to incorporate prior physical knowledge of UAV motion; third, its generalization ability to unseen operating conditions and model-mismatched scenarios is weak, and it is prone to estimation bias under the complex disturbances of urban environments.

[0006] In recent years, the integration of the recursive efficiency of model-based methods and the adaptive capabilities of data-driven methods has become a research hotspot for UAV state estimation. KalmanNet, as a typical hybrid framework, solves the model mismatch problem to some extent by replacing traditional analytical computation with Kalman gain learning through neural networks. However, existing KalmanNets still have shortcomings: the scheduling of Kalman gain is implicitly generated by the network's hidden states, lacking physical interpretability; its robustness to strongly nonlinear and severely mismatched scenarios is insufficient, making it difficult to adapt to the complex working conditions of urban environments; and the network architecture has a large parameter scale, making it difficult to meet the real-time requirements of UAV airborne applications. Summary of the Invention

[0007] The technical solution adopted by this invention to solve its technical problem is to provide a robust state estimation method for UAVs in urban environments. By constructing an uncertainty-guided KalmanNet hybrid estimation model, it integrates the recursive framework of Kalman filtering and the adaptive gain learning of deep learning, and combines explicit uncertainty vectors to achieve interpretable tuning of Kalman gain. At the same time, it designs an ultra-lightweight network architecture to adapt to UAV onboard hardware, and finally achieves robust and high-precision state estimation of UAVs in urban environments.

[0008] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0009] A robust state estimation method for UAVs in urban environments includes the following steps:

[0010] 1) Construct an uncertainty-guided implicit KalmanNet mixture estimation model;

[0011] 2) The implicit KalmanNet mixture estimation model is trained using UAV flight data in an urban environment to obtain the optimal state estimation model;

[0012] 3) Collect sensor data in real time and generate uncertainty vectors. Output Kalman gain through model inference. Combined with the "prediction-update" framework of Kalman filtering, realize real-time robust estimation of UAV state.

[0013] The implicit KalmanNet mixture estimation model includes:

[0014] The system model building module is used to establish a nonlinear state evolution model and measurement model in the three-dimensional space of the UAV, which serves as the physical prior basis for state estimation.

[0015] The uncertainty vector generation module is used to construct uncertainty vectors, which serve as key input features for the gain learning module.

[0016] The ultra-lightweight KalmanNet gain learning module is used to construct an ultra-lightweight recursive network architecture and adaptively learn the Kalman gain using system state information, observation residuals, and uncertainty vectors.

[0017] The state recursive estimation module is used for a Kalman filter-based “prediction-update” recursive framework, which integrates a nonlinear state evolution model, a measurement model, and Kalman gain to perform real-time recursive estimation of the UAV state.

[0018] The uncertainty vector generation module performs the following steps:

[0019] State deviation calculation: Calculate the component-level state deviation between the current estimated state of the UAV and the nominal linearized point, forming a nine-dimensional deviation vector consistent with the dimension of the state vector;

[0020] Uncertainty sensitivity weight calculation: Define a nine-dimensional uncertainty sensitivity weight vector, each vector is calculated by the Euclidean norm of the corresponding column of the uncertainty matrix of the model set in the experiment;

[0021] Uncertainty vector construction: The absolute values ​​of the sensitivity weight vector and the component-level state deviation vectors are multiplied element by element to generate the final uncertainty vector.

[0022] The ultra-lightweight KalmanNet gain learning module performs the following steps:

[0023] Input feature extraction: Extract observation difference, innovation difference, state evolution difference, state update difference, and uncertainty vector generated by each difference to construct an input feature set that integrates physical prior and data features;

[0024] Constructing an ultra-lightweight recursive network: Construct an architecture consisting of an input fully connected layer, a cascaded GRU layer, and an output fully connected layer. Perform physical prior and data feature fusion on the input feature set to generate the optimal Kalman gain.

[0025] The ultralightweight recursive network includes:

[0026] The input fully connected layer is used to map high-dimensional input features into low-dimensional feature vectors, so that the input features are adapted to the input dimension of the cascaded GRU layer;

[0027] The cascaded GRU layer is used to track the unknown state noise covariance, prediction covariance and observation covariance by setting up 3 independent GRU sub-layers respectively. The sub-layers are cascaded according to the physical logic of Kalman filtering.

[0028] Output a fully connected layer to map the hidden states of the cascaded GRU layers into a Kalman gain matrix that matches the state dimension and the observation dimension.

[0029] The state recursive estimation module performs the following steps:

[0030] Prediction Phase: Using a nonlinear state evolution model, the prior state at the current time is predicted from the posterior state estimate of the previous time step. Simultaneously, the prior observation at the current time step is predicted based on a measurement model.

[0031] Update phase: Using Kalman gain and combining the difference between the current actual observation and the prior observation prediction, the prior state is corrected and updated to obtain the posterior state estimate at the current time.

[0032] Step 3) includes the following steps:

[0033] 3.1) Construct a labeled dataset containing real state sequences and sensor observation sequences, and divide it into training and testing sets;

[0034] 3.2) The mean squared error of the state estimation is used as the core loss function, and an L2 regularization term is added;

[0035] 3.3) A mini-batch training strategy using stochastic gradient descent is adopted, combined with the backpropagation algorithm to train the model, and a learning rate decay strategy is adopted to stop training when the validation set loss does not decrease for several consecutive rounds.

[0036] 3.4) During the training process, the uncertainty vector is used as a key input feature to perform interpretable tuning of the Kalman gain.

[0037] A robust state estimation system for unmanned aerial vehicles (UAVs) in urban environments includes:

[0038] The implicit KalmanNet mixture estimation model building module is used to build uncertainty-guided implicit KalmanNet mixture estimation models.

[0039] The model training module is used to train the implicit KalmanNet mixture estimation model using UAV flight data in an urban environment to obtain the optimal state estimation model.

[0040] The robust state estimation module is used to collect sensor data in real time and generate uncertainty vectors. It outputs Kalman gain through model inference and combines the "prediction-update" framework of Kalman filtering to achieve real-time robust estimation of the UAV state.

[0041] The present invention has the following beneficial effects and advantages:

[0042] This invention deeply integrates the recursive efficiency and physical interpretability of model-based Kalman filtering with the adaptive and nonlinear fitting capabilities of data-driven deep learning. By constructing an explicit uncertainty vector and an ultra-lightweight KalmanNet architecture, it effectively solves the problems of model mismatch, nonlinear dynamics, and non-Gaussian perturbation in UAV state estimation in urban environments. Compared with existing technologies, it has the following significant advantages:

[0043] 1) Improved estimation accuracy and robustness: The explicit uncertainty vector quantifies model mismatch and linearization error, enabling interpretable adaptive tuning of the Kalman gain, making the model highly robust to strong nonlinearity and severe model mismatch scenarios; Experimental results show that in complex urban environments, compared with traditional extended Kalman filter (EKF) and pure data-driven RNN, the mean square error (MSE) of state estimation is reduced by more than 60%, and the tracking error is significantly reduced.

[0044] 2) Enhance the physical interpretability of the model: The uncertainty vector is constructed based on the UAV state deviation and model sensitivity, directly reflecting the physical uncertainty of the model. The Kalman gain tuning process is guided by the uncertainty signal, breaking the "black box" defect of the traditional data-driven method. It integrates physical prior knowledge into the data-driven gain learning, making the state estimation process both data-adaptive and physically interpretable.

[0045] 3) Reduced computational cost and adaptable to airborne deployment: The ultra-lightweight recurrent network architecture controls the training parameters to within 3000. Compared with traditional KalmanNet and deep recurrent neural networks, the computational cost is reduced by more than 70%, and the single-step state estimation time is ≤0.05s. It can be directly deployed on the resource-constrained airborne embedded hardware of UAVs to meet the real-time state estimation requirements.

[0046] 4) Strong generalization ability and adaptability to multiple operating conditions: The model integrates the physical priors of UAV motion and the nonlinear evolution law learned from the data. It does not require retraining for different urban operating conditions and has good adaptability to operating conditions with different flight speeds, attitudes and building occlusion levels. The generalization error is less than 3%.

[0047] 5) Highly practical and easy to integrate: The state estimation process of this invention is compatible with the interface of the existing flight control system of UAVs and can be directly integrated into the autonomous navigation system of UAVs without large-scale modification of existing hardware; the estimation process is highly automated and requires no manual intervention, which can provide continuous, stable and high-precision state support for the autonomous flight of UAVs in urban environments, reduce flight risks and improve operational reliability. Attached Figure Description

[0048] Figure 1The overall architecture diagram of the UAV robust state estimation model based on UGL-KalmanNet in this invention. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0050] A robust state estimation method for UAVs in urban environments includes the following steps:

[0051] (1) Construction of a robust state estimation model for UAVs based on UGL-KalmanNet

[0052] The estimation model constructed in this invention mainly consists of four parts: a system model construction module, an uncertainty vector generation module, an ultra-lightweight KalmanNet gain learning module, and a state recursive estimation module. These modules work together to achieve robust estimation of the UAV state. The specific structure is as follows:

[0053] (1-1) System Model Construction Module: Establish a nonlinear state evolution model and measurement model for the UAV in three-dimensional space as the physical prior basis for state estimation. The specific steps are as follows:

[0054] 1) State definition: Define a nine-dimensional state vector for the UAV, where the motion state includes the UAV's position (x, y, z directions) and velocity (velocity components in the three directions) in three-dimensional space, and the attitude state includes three Euler angles (roll angle, pitch angle, and yaw angle) describing the UAV's orientation.

[0055] 2) Nonlinear state evolution model: Based on the kinematic laws of UAVs, a nonlinear state evolution relationship in continuous time is established, and process noise is introduced to characterize the inherent perturbation of the model; the continuous model is discretized by the first-order Euler integral to obtain a discrete state evolution relationship suitable for recursive estimation, and the influence of sampling interval on the discretized model and the calculation logic of discrete process noise are clarified.

[0056] 3) Measurement Model: Establish the nonlinear measurement relationship of the UAV's airborne sensors, describe the mapping law from the state vector to the observation value, introduce measurement noise to characterize the sensor observation error, and clarify the correspondence between the observation value and the state vector of different sensors.

[0057] (1-2) Uncertainty Vector Generation Module: Constructs explicit, physically interpretable uncertainty vectors to quantify the uncertainties caused by system model mismatch and linearization errors, serving as key input features for the gain learning module. The specific steps are as follows:

[0058] 1) State Deviation Calculation: Calculate the component-level state deviation between the current estimated state of the UAV and the nominal linearized point, forming a nine-dimensional deviation vector consistent with the dimension of the state vector;

[0059] 2) Uncertainty sensitivity weight calculation: Define a nine-dimensional uncertainty sensitivity weight vector. Each component is calculated using the Euclidean norm of the corresponding column of the model uncertainty matrix. This is used to quantify the contribution of the unit error of a single-state component to the overall model uncertainty.

[0060] 3) Uncertainty vector construction: The sensitivity weights are multiplied element by element by the absolute values ​​of the component-level state deviations to generate the final uncertainty vector. Each component in the vector intuitively represents the magnitude of the uncertainty contributed by the corresponding state component at the current moment, thus realizing the explicit quantification of model mismatch and linearization error.

[0061] (1-3) Ultra-lightweight KalmanNet gain learning module: An ultra-lightweight recurrent network architecture with fewer than 3000 training parameters is designed. It adaptively learns the Kalman gain using system state information, observation residuals, and uncertainty vectors as input, replacing the analytical gain calculation of traditional Kalman filtering. The specific structure is as follows:

[0062] 1) Input Feature Extraction: Extract multi-dimensional input features, including observation difference (difference between the current and previous observations), innovation difference (difference between the current observation and the prior observation prediction), state evolution difference (difference between the current posterior state and the previous posterior state), state update difference (difference between the current posterior state and the current prior state), and the uncertainty vector generated above, to form an input feature set that integrates physical priors and data features;

[0063] 2) Ultra-lightweight recursive network: It adopts an architecture of "input fully connected layer + cascaded GRU layer + output fully connected layer", where:

[0064] Input fully connected layer: maps high-dimensional input features to low-dimensional feature vectors, adapts to the input dimension of the GRU layer, and reduces redundant computation;

[0065] Cascaded GRU layers: Three independent GRU sub-layers are set up to track the unknown state noise covariance, prediction covariance and observation covariance respectively. The sub-layers are cascaded according to the physical logic of Kalman filtering, so that the network learning process conforms to the physical law of state estimation and significantly reduces redundant parameters.

[0066] Output fully connected layer: Maps the hidden states of the GRU layer to a Kalman gain matrix that matches the state dimension and observation dimension;

[0067] 4) Network constraints: Through architectural optimization design, the training parameters are strictly controlled within 3000 to achieve ultra-lightweight design and meet the limited computing resource constraints of UAV onboard embedded hardware.

[0068] (1-4) State Recursive Estimation Module: Based on the "prediction-update" recursive framework of Kalman filtering, the module integrates the above nonlinear system model and the learned Kalman gain to achieve real-time recursive estimation of the UAV state. The specific steps are as follows:

[0069] 1) Prediction phase: Using a nonlinear state evolution model, the prior state at the current moment is predicted from the posterior state estimate of the previous moment, and the prior observation at the current moment is predicted based on the measurement model.

[0070] 2) Update phase: Using the Kalman gain output by the gain learning module, combined with the observation innovation value (the difference between the current actual observation value and the prior observation prediction value), the prior state is corrected and updated to obtain the estimated posterior state value at the current time.

[0071] By recursively executing the "prediction-update" process, continuous real-time estimation of the UAV's state is achieved.

[0072] (2) UGL-KalmanNet model training and UAV state estimation process

[0073] (2-1) Dataset Construction:

[0074] 1) Data Acquisition: Collect flight trajectory data of UAVs in urban environments, including real-time data of UAVs (acquired by a high-precision motion capture system) and observation data from airborne sensors (including observations of position, speed, attitude, etc.), covering different flight speeds, attitudes, and building obstruction conditions to ensure the comprehensiveness of the dataset;

[0075] 2) Data preprocessing: Denoise, interpolate, and time-align the collected data to eliminate sensor noise and time synchronization errors; construct a labeled dataset containing state sequences and observation sequences, where the state sequences are the labels and the observation sequences are the inputs;

[0076] 3) Dataset partitioning: The preprocessed dataset is divided into a training set and a test set in an 8:2 ratio. The training set is used for model parameter learning, and the test set is used for model performance verification.

[0077] (2-2) Model training:

[0078] 1) Loss function definition: The mean squared error of the state estimation is used as the core loss function to quantify the deviation between the estimated state and the true state; at the same time, an L2 regularization term is added to suppress model overfitting and clarify the regulatory effect of the regularization coefficient on the model's generalization ability.

[0079] 2) Optimization of Algorithm and Training Strategy: A mini-batch training strategy using stochastic gradient descent is adopted, combined with the backpropagation algorithm to train the recurrent network; the initial learning rate is set to 0.001, and a learning rate decay strategy is adopted (the learning rate decays to 1 / 10 of the original after every 50 training rounds); the training rounds are set to 200 rounds, and training is stopped when the validation set loss does not decrease for 10 consecutive rounds to save the optimal model parameters and avoid ineffective training and overfitting;

[0080] 3) Uncertainty-guided gain tuning: During training, the uncertainty vector is used as a key input feature to achieve interpretable tuning of the Kalman gain. When the uncertainty vector value is high (indicating model mismatch and large linearization error), the network outputs a smaller Kalman gain to reduce the impact of observations on state updates. When the uncertainty vector value is low (indicating high model matching and small linearization error), the network outputs a larger Kalman gain to fully utilize observations to optimize state estimation.

[0081] (2-3) Real-time state estimation of UAVs in urban environments: The trained UGL-KalmanNet model is deployed to the UAV's onboard embedded hardware (such as NVIDIA Jetson Nano) to achieve real-time robust estimation of the UAV's state in urban environments. The specific steps are as follows:

[0082] 1) Airborne initialization: After the UAV is powered on, the state estimation module is started to complete the initialization of the airborne sensors (GPS, IMU, visual sensor), data preprocessing program and optimal UGL-KalmanNet model. The sensors collect observation data at the set sampling frequency (≥10Hz).

[0083] 2) Real-time data processing: Preprocess the real-time observation data collected by the sensor by denoising, time alignment, etc., and extract the observation features; at the same time, calculate the deviation between the current state and the nominal point to generate a real-time uncertainty vector;

[0084] 3) Model inference and state estimation: The preprocessed observation features and real-time uncertainty vector are input into the UGL-KalmanNet model, and the model quickly outputs the Kalman gain; based on the Kalman filter-based "prediction-update" framework, the real-time state estimation of the UAV's position, velocity, and attitude is realized, with a single-step estimation time ≤0.05s, meeting the real-time requirements;

[0085] 4) Status output and feedback: The estimated UAV status is output to the UAV flight control system in real time, providing high-precision status feedback for autonomous navigation, attitude control and path planning; if the estimation error exceeds the preset threshold, an alarm signal is triggered, prompting the flight control system to take fault-tolerant strategies to ensure flight safety.

[0086] Example

[0087] like Figure 1 The diagram shown is the overall architecture of the robust state estimation model for UAVs based on UGL-KalmanNet according to this invention. This invention provides a robust state estimation method for UAVs in urban environments, characterized by the following steps:

[0088] Step (1): Unmanned aerial vehicle system model construction and data acquisition, outputting labeled state-observation datasets.

[0089] 1) Nonlinear system model construction: Based on the kinematic laws of the UAV, a nine-dimensional nonlinear state evolution model and measurement model are established. The state vector is defined to include key states such as position, velocity, and Euler angles. The core parameters of the model (such as rotation matrix, transformation matrix, and sampling interval of 0.01s) are determined.

[0090] 2) Airborne sensor deployment: A multi-source observation system consisting of GPS, IMU, and visual sensors is deployed on the UAV, and a high-precision motion capture system is built on the UAV flight test platform to obtain real-time data of the UAV.

[0091] 3) Urban environment data collection: Conduct UAV flight tests in urban street environments to collect airborne sensor observation data and motion capture system real-state data under different flight conditions (speed: 1-5m / s, attitude: roll / pitch angle 0-30°, occlusion degree: no occlusion / partial occlusion / complete occlusion). Collect 1000 sets of trajectory data for each flight condition.

[0092] 4) Dataset preprocessing and partitioning: The collected data is denoised (using median filtering), interpolated and time aligned to eliminate sensor noise and time synchronization errors; a labeled dataset of state sequence-observation sequence is constructed and divided into training set and test set in an 8:2 ratio for model training and performance verification.

[0093] Step (2): UGL-KalmanNet model construction and training, outputting the optimal state estimation model.

[0094] The UGL-KalmanNet model is constructed and trained based on the preprocessed dataset. The specific process is as follows:

[0095] 1) Model structure construction: According to the model architecture of this invention, the system model construction module, the uncertainty vector generation module, the ultra-lightweight KalmanNet gain learning module and the state recursive estimation module are built in sequence. The training parameters of the ultra-lightweight KalmanNet are controlled at around 2800 to meet the resource constraints of the airborne hardware.

[0096] 2) Uncertainty vector generation: Calculate the state bias and uncertainty sensitivity weights, construct an explicit uncertainty vector, and use it as the key input feature of the gain learning module;

[0097] 3) Model Training: A training environment was built based on the PyTorch deep learning framework. The mean squared error loss function was combined with L2 regularization. The model was trained using mini-batch stochastic gradient descent (SGD) and backpropagation in time (BPTT) algorithms. The initial learning rate was set to 0.001, which decayed to 1 / 10 every 50 rounds. The training lasted for 200 rounds. Training was stopped when the validation set loss did not decrease for 10 consecutive rounds, and the optimal model parameters were saved.

[0098] 4) Model performance validation: The trained model is validated using a test set. The state estimation accuracy (RMSE / MSE), real-time performance (single-step estimation time), and robustness of the model under different levels of model mismatch are evaluated to ensure that the model meets the requirements of UAV state estimation in urban environments.

[0099] Step (3): Real-time state estimation of UAV in urban environment, outputting high-precision state estimation results.

[0100] The trained UGL-KalmanNet model was deployed to the embedded hardware (NVIDIA Jetson Nano) on a drone to achieve real-time robust estimation of the drone's state in an urban environment. The specific steps are as follows:

[0101] 1) Airborne initialization: After the UAV is powered on, the state estimation module is started to complete the initialization of multi-source sensors, data preprocessing program and optimal UGL-KalmanNet model. The sensors collect observation data at a sampling frequency of 10Hz.

[0102] 2) Real-time data processing and uncertainty vector generation: Denoise and time-aligned preprocessing is performed on the real-time observation data collected by the sensor to extract observation features; at the same time, the deviation between the current estimated state and the nominal linearization point is calculated to generate a real-time uncertainty vector;

[0103] 3) Model inference and recursive state estimation: The observed features and uncertainty vector are input into the UGL-KalmanNet model, and the model quickly outputs the Kalman gain; the "prediction-update" recursive framework based on Kalman filtering realizes the real-time state estimation of UAV position, velocity and attitude, and the single-step estimation time is controlled within 0.05s.

[0104] 4) Status output and fault-tolerant feedback: The estimated UAV status is output to the UAV flight control system in real time to provide feedback for autonomous navigation and attitude control; if the status estimation error exceeds the preset threshold (RMSE>0.5m), an alarm signal is triggered, and the flight control system automatically adopts strategies such as reducing flight speed and starting fault-tolerant navigation to ensure the flight safety of the UAV.

[0105] Step (4): Online validation and optimization of model performance to ensure consistently stable estimation results.

[0106] During actual drone flight, the model's state estimation performance was validated online and dynamically optimized.

[0107] 1) Online performance evaluation: Real-time calculation of the deviation between the estimated state and the sensor fusion observations to evaluate the estimation accuracy and robustness of the model under current urban conditions;

[0108] 2) Incremental learning optimization: If the drone enters a new urban environment (such as a dense building complex or an area with strong electromagnetic interference), it collects a small amount of new trajectory data, performs incremental learning on the model, updates the network parameters, and improves the model's adaptability to the new environment.

[0109] 3) Hardware adaptation optimization: Based on the computing resource status of the airborne hardware, the inference accuracy of the model is dynamically adjusted. While ensuring the estimation accuracy, the real-time performance is further improved to ensure the stable operation of the model on the airborne hardware.

[0110] This invention, through the organic combination of the above steps, forms a complete robust state estimation scheme for UAVs in urban environments. By leveraging the synergistic effect of explicit uncertainty vectors and ultra-lightweight UGL-KalmanNet, it effectively solves the problems of poor robustness of traditional model-based methods, low interpretability of pure data-driven methods, and high computational cost. It has the advantages of clear principles, high accuracy, strong robustness, low computational cost, and easy integration, and can be widely applied to the autonomous navigation, precision operation, and safe flight of various UAVs in urban environments.

[0111] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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.

[0112] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A robust state estimation method for unmanned aerial vehicles (UAVs) in urban environments, characterized in that, Includes the following steps: 1) Construct an uncertainty-guided implicit KalmanNet mixture estimation model; 2) The implicit KalmanNet mixture estimation model is trained using UAV flight data in an urban environment to obtain the optimal state estimation model; 3) Collect sensor data in real time and generate uncertainty vectors. Output Kalman gain through model inference. Combined with the "prediction-update" framework of Kalman filtering, realize real-time robust estimation of UAV state.

2. The robust state estimation method for UAVs in urban environments according to claim 1, characterized in that, The implicit KalmanNet mixture estimation model includes: The system model building module is used to establish a nonlinear state evolution model and measurement model in the three-dimensional space of the UAV, which serves as the physical prior basis for state estimation. The uncertainty vector generation module is used to construct uncertainty vectors, which serve as key input features for the gain learning module. The ultra-lightweight KalmanNet gain learning module is used to construct an ultra-lightweight recursive network architecture and adaptively learn the Kalman gain using system state information, observation residuals, and uncertainty vectors. The state recursive estimation module is used for real-time recursive estimation of UAV state in a Kalman filter-based "prediction-update" recursive framework that integrates nonlinear state evolution model, measurement model and Kalman gain.

3. The robust state estimation method for UAVs in urban environments according to claim 2, characterized in that, The uncertainty vector generation module performs the following steps: State deviation calculation: Calculate the component-level state deviation between the current estimated state of the UAV and the nominal linearized point, forming a nine-dimensional deviation vector consistent with the dimension of the state vector; Uncertainty sensitivity weight calculation: Define a nine-dimensional uncertainty sensitivity weight vector, each vector is calculated by the Euclidean norm of the corresponding column of the uncertainty matrix of the model set in the experiment; Uncertainty vector construction: The absolute values ​​of the sensitivity weight vector and the component-level state deviation vectors are multiplied element by element to generate the final uncertainty vector.

4. A robust state estimation method for UAVs in urban environments according to claim 2, characterized in that, The ultra-lightweight KalmanNet gain learning module performs the following steps: Input feature extraction: Extract observation difference, innovation difference, state evolution difference, state update difference, and uncertainty vector generated by each difference to construct an input feature set that integrates physical prior and data features; Constructing an ultra-lightweight recursive network: Construct an architecture consisting of an input fully connected layer, a cascaded GRU layer, and an output fully connected layer to fuse physical priors and data features of the input feature set and generate the optimal Kalman gain.

5. A robust state estimation method for UAVs in urban environments according to claim 4, characterized in that, The ultralightweight recursive network includes: The fully connected input layer is used to map high-dimensional input features into low-dimensional feature vectors, so that the input features are adapted to the input dimension of the cascaded GRU layer. The cascaded GRU layer is used to track the unknown state noise covariance, prediction covariance and observation covariance by setting up 3 independent GRU sub-layers respectively. The sub-layers are cascaded according to the physical logic of Kalman filtering. Output a fully connected layer to map the hidden states of the cascaded GRU layers into a Kalman gain matrix that matches the state dimension and the observation dimension.

6. A robust state estimation method for UAVs in urban environments according to claim 2, characterized in that, The state recursive estimation module performs the following steps: Prediction Phase: Using a nonlinear state evolution model, the prior state at the current time is predicted from the posterior state estimate of the previous time step. Simultaneously, the prior observation at the current time step is predicted based on a measurement model. Update phase: Using Kalman gain and combining the difference between the current actual observation and the prior observation prediction, the prior state is corrected and updated to obtain the posterior state estimate at the current time.

7. A robust state estimation method for UAVs in urban environments according to claim 1, characterized in that, Step 3) includes the following steps: 3.1) Construct a labeled dataset containing real state sequences and sensor observation sequences, and divide it into training and testing sets; 3.2) The mean squared error of the state estimation is used as the core loss function, and an L2 regularization term is added; 3.3) A mini-batch training strategy using stochastic gradient descent is adopted, combined with the backpropagation algorithm to train the model, and a learning rate decay strategy is adopted to stop training when the validation set loss does not decrease for several consecutive rounds. 3.4) During the training process, the uncertainty vector is used as a key input feature to perform interpretable tuning of the Kalman gain.

8. A robust state estimation system for unmanned aerial vehicles (UAVs) in urban environments, characterized in that, include: The implicit KalmanNet mixture estimation model building module is used to build uncertainty-guided implicit KalmanNet mixture estimation models. The model training module is used to train the implicit KalmanNet mixture estimation model using UAV flight data in an urban environment to obtain the optimal state estimation model. The robust state estimation module is used to collect sensor data in real time and generate uncertainty vectors. It outputs Kalman gain through model inference and combines the "prediction-update" framework of Kalman filtering to achieve real-time robust estimation of the UAV state.

Images