Unmanned aerial vehicle robust pose estimation method and system for complex mountain forest environment

Abstract

The application discloses a kind of unmanned aerial vehicle robust attitude estimation method and system for complex mountain forest environment, comprising: the multi-modal observation data of inertial measurement unit IMU and global satellite navigation system GPS are preprocessed, to construct unified observation sequence;Based on multiscale time series feature extraction generates channel level noise mask, and the observation noise covariance matrix is adaptively adjusted, to suppress multiscale non-stationary noise;The double-branch closed-loop architecture of inertial inference branch and filtering estimation branch is constructed, realizes uncertainty information interaction and noise parameter feedback;Based on physical uncertainty index and time series feature, fusion weight is dynamically generated by gate fusion network, and under the constraint of SO (3) manifold, attitude fusion is completed, and stable continuous attitude estimation result is output.The application can effectively suppress heterogeneous noise in complex environment, maintain estimation stability under the condition of intermittent absence of GPS, improve the robustness and consistency of unmanned aerial vehicle attitude estimation.

Description

Technical Field

[0001] This invention belongs to the field of UAV autonomous navigation and state estimation technology, and relates to a robust attitude estimation method and system for UAVs in complex mountainous and forest environments. Background Technology

[0002] In recent years, unmanned aerial vehicles (UAVs) have been widely used in forest fire prevention patrols, mountain surveying, power line inspections, natural resource surveys, and emergency rescue due to their advantages such as high mobility, flexible deployment, and low operating costs. These applications often require UAVs to fly at low altitudes for extended periods or autonomously in complex mountainous and forested environments. Flight control, navigation, and mission execution all heavily rely on the accuracy and continuity of attitude estimation results. Attitude estimation typically involves real-time acquisition of the UAV's roll angle, pitch angle, yaw angle, and state variables such as position and velocity. Its accuracy and stability directly affect the UAV's flight safety and operational quality.

[0003] With the widespread application of low-altitude economy and UAVs in complex mountainous and forested environments, attitude estimation technology faces severe challenges. Current mainstream solutions are based on the combination of inertial measurement units (IMUs) and global navigation satellite systems (GPS). While IMUs have the advantage of short-term dynamic response, they are prone to divergence due to zero-bias drift and accumulated errors when external correction is lacking. GPS is frequently obstructed in mountainous and forested environments, and signal blockage, multipath effects, and interference often lead to accuracy degradation or even intermittent interruptions. Traditional IMU / GPS combinations are difficult to maintain stable and reliable attitude output.

[0004] The interference in complex mountainous forest environments is mainly reflected in three aspects: First, the IMU is affected by multiple factors such as vibration, gusts and temperature changes. The noise of each channel has heterogeneous and non-stationary characteristics, including various forms such as spike disturbances, thick tail distribution and long-term drift. Traditional filtering methods use the assumption of fixed covariance, which cannot effectively suppress abnormal noise. If the polluted channel is given a high confidence level, the state estimation will be severely polluted. Second, GPS often experiences short-term lockout or accuracy degradation in mountainous forest areas. At this time, the system degenerates into a pure inertial mode. Due to the lack of external constraints, the heading angle and zero bias accumulate rapidly, resulting in attitude distortion. Third, existing fusion strategies mostly rely on static weights or hard switching mechanisms, which lack the ability to model the evolution of uncertainty. The weights are prone to jitter, and the output is not smooth. Moreover, the attitude is described by quaternions or rotation matrices, located on the SO(3) manifold. Direct linear weighting destroys the unit norm constraint, and the output loses physical validity. Current methods can be categorized into three types: model-driven methods such as extended Kalman filtering, unscented Kalman filtering, and particle filtering, which, while having clear physical meaning, rely on accurate models and fixed noise assumptions, making them difficult to handle strong non-Gaussian noise and missing observations; data-driven methods such as long short-term memory networks, convolutional neural networks, and Transformers, which can fit complex nonlinearities but lack physical constraints and have insufficient generalization; and hybrid methods that combine models and learning still have significant shortcomings in terms of noise modeling granularity, bi-branch collaboration, temporal continuity, and attitude manifold constraints.

[0005] Therefore, current technologies lack a robust attitude estimation method for UAVs that can simultaneously adapt to multi-scale heterogeneous noise, intermittent GPS loss, and attitude manifold physical constraints in complex mountainous and forest environments. There is an urgent need to propose a new technical solution that can uniformly process multimodal observations, dynamically perceive the reliability of noise across different channels and time scales, maintain continuous and stable state estimation even under GPS signal degradation or interruption, and ensure the smoothness and physical consistency of attitude output during multi-source fusion, thereby meeting the long-term autonomous operation requirements of UAVs in complex unstructured environments. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in UAV attitude estimation in complex mountainous and forest environments, such as coarse noise modeling, estimation degradation due to intermittent GPS signal loss, discontinuous multi-branch fusion process, and insufficient physical consistency of attitude representation. This invention proposes a robust UAV attitude estimation method and system for complex mountainous and forest environments. Based on unified multimodal observation preprocessing, this method introduces a channel-level multi-scale noise adaptive suppression mechanism, an interactive bi-branch closed-loop attitude inference architecture, and an uncertainty-aware temporal gating fusion mechanism. This enables deep collaboration between model-driven estimation and data-driven inference, resulting in continuous, stable, robust, and physically consistent attitude estimation results in complex environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention: a robust attitude estimation method for unmanned aerial vehicles (UAVs) in complex mountainous forest environments, comprising the following steps: Step 1: Perform unified time reference synchronization and coordinate system transformation on the multimodal observation data output by the inertial measurement unit (IMU) and the global satellite navigation system (GPS) on the UAV, and normalize the unified observation vectors and perform outlier detection and smoothing correction to obtain the preprocessed discrete observation sequence. Step 2: Extract multi-scale temporal features from the preprocessed discrete observation sequence using a temporal convolutional network (TCN) based on dilated convolution, and generate a noise reliability mask online for each observation channel based on the multi-scale temporal features. Then, use the mask to adaptively diagonally scale the basic observation noise covariance matrix to obtain a dynamically adjusted observation noise covariance matrix. Step 3: Construct an interactive dual-branch closed-loop attitude inference architecture, which includes a data-driven inertial inference branch and a model-driven filter estimation branch; wherein: The model-driven filter estimation branch uses the dynamically adjusted observation noise covariance matrix to update the filter and outputs an explicit uncertainty index. The data-driven inertial inference branch performs state inference based on IMU timing input and receives the explicit uncertainty index as additional input to sense the reliability level of the filtering branch, while outputting the correction amount of process noise and / or observation noise to the filtering branch. Step 4: Through the uncertainty-aware temporal gating fusion mechanism, the output of the data-driven inertial inference branch and the output of the model-driven filtering estimation branch are continuously fused in the time dimension to obtain the final attitude estimation result.

[0008] Specifically, the unified time reference synchronization in step 1 includes: aligning the three-axis angular velocity and three-axis specific force output by the IMU with the three-dimensional position and three-dimensional velocity output by the GPS based on the unified time reference; the coordinate system transformation includes transforming the observations in the IMU carrier coordinate system to the Northeast Sky ENU geographic coordinate system.

[0009] Specifically, the normalization process is performed in a standardized form: ; or minimum-maximum normalized form: ; in, and These are the mean and standard deviation of the corresponding components on the training set or calibration set, respectively; the outlier detection and smoothing correction include: identifying abrupt change points using a sliding window mean filter, the three-standard-deviation principle, or a threshold discrimination method based on local consistency, and processing outliers by using neighboring sample interpolation, local smoothing replacement, or marking them as low-confidence observations.

[0010] Specifically, the TCN is composed of multiple residual dilated convolutions, with different dilation rates corresponding to different time scales. The dilation rate of each layer increases exponentially. The local scale is used to capture spikes and abrupt changes, the mesoscale is used to model short- and medium-term correlated noise, and the long scale is used to perceive slow drift and systematic shifts. The noise reliability mask is generated by fully connected mapping and Sigmoid activation through the temporal features extracted by the TCN. The mask value is in the range of (0,1) and represents the unreliability of the corresponding observation channel.

[0011] Specifically, the adaptive diagonal scaling includes: constructing a diagonal scaling matrix: ;in, For the first Each observation channel at time The mask value, The scaling factor is used to define the adaptive observation noise covariance matrix as follows: Or, under the condition of diagonal approximation, it can be written as: ;in, Based on the observation noise covariance matrix, It is its diagonal element; when the mask value of a certain channel increases, the corresponding covariance is amplified, thereby reducing the contribution of the channel observation to the filter update.

[0012] Specifically, the model-driven filtering estimation branch employs the Unscented Kalman Filter (UKF), which uses the dynamically adjusted observation noise covariance matrix to calculate the observation covariance and Kalman gain, and performs state updates; the explicit uncertainty index includes at least one of the normalized squared innovation, the state covariance trace, and the diagonal elements of the observation residuals.

[0013] Specifically, the data-driven branch employs a recurrent neural network, specifically a Long Short-Term Memory (LSTM) network, to output a relative state estimate based on the IMU time-series input. The data-driven branch receives the explicit uncertainty index and uses it, along with the IMU time-series features, as input to the recurrent neural network to enhance the perception of changes in system reliability. The corrections output by the data-driven branch to the filtering branch include process noise covariance corrections and observation noise covariance corrections, used to adjust the noise setting of the filtering branch in real time.

[0014] Furthermore, the uncertainty-aware temporal gating fusion mechanism includes: constructing a fusion feature vector, wherein the fusion feature vector contains the normalized squared innovation, state covariance trace, GPS signal quality index, and hidden state features of the data-driven branch output; and inputting the fusion feature vector into a gated recurrent unit (GRU) to obtain smoothly varying fusion weights. The fusion weights are used to perform a linear convex combination of the position and velocity estimates of the two branches, and to perform a weighted normalization fusion of the attitude quaternions of the two branches after sign alignment.

[0015] Furthermore, it also includes a self-supervised composite loss function for end-to-end training of the noise reliability mask generation network; the composite loss function includes residual consistency loss, dynamic consistency loss and mask prior loss, wherein the residual consistency loss encourages the observed residuals to match the estimated noise level after adaptive covariance adjustment, the dynamic consistency loss constrains the estimated state to satisfy the rigid body motion law, and the mask prior loss encourages mask sparsity and temporal smoothness.

[0016] A second aspect of the present invention: a robust attitude estimation system for unmanned aerial vehicles (UAVs) in complex mountainous and forested environments, comprising: Data preprocessing module: performs unified time reference synchronization and coordinate system transformation on the multimodal observation data output by the inertial measurement unit (IMU) and the global satellite navigation system (GPS) on the UAV, and performs normalization processing, outlier detection and smoothing correction on the unified observation vector to obtain the preprocessed discrete observation sequence; Noise adaptive suppression module: It uses a time convolutional network (TCN) based on dilated convolution to extract multi-scale temporal features from the preprocessed discrete observation sequence, and generates a noise reliability mask online for each observation channel based on the multi-scale temporal features. Then, it uses the mask to adaptively diagonally scale the basic observation noise covariance matrix to obtain a dynamically adjusted observation noise covariance matrix. Dual-branch attitude inference module: Constructs an interactive dual-branch closed-loop attitude inference architecture, which includes a data-driven inertial inference branch and a model-driven filter estimation branch; wherein: The model-driven filter estimation branch uses the dynamically adjusted observation noise covariance matrix to update the filter and outputs an explicit uncertainty index. The data-driven inertial inference branch performs state inference based on IMU timing input and receives the explicit uncertainty index as additional input to sense the reliability level of the filtering branch, while outputting the correction amount of process noise and / or observation noise to the filtering branch. Temporal Gated Fusion Module: Through an uncertainty-aware temporal gated fusion mechanism, the output of the data-driven inertial inference branch and the output of the model-driven filtering estimation branch are continuously fused in the time dimension to obtain the final attitude estimation result.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects:

[0018] First, this invention improves the traditional fixed observation covariance into a learnable covariance model that dynamically changes with time and channels through a channel-level multi-scale noise adaptive suppression mechanism. This model can more precisely suppress instantaneous spike noise, thick-tailed perturbation noise, and long-term drift noise, reducing the pollution of overall state estimation by abnormal observations while retaining effective dynamic information. Second, this invention deeply couples model-driven filtering with data-driven inertial inference through an interactive dual-branch closed-loop architecture. This not only achieves high-precision absolute estimation when GPS is normal but also enhances the continuous recursion capability when GPS is missing, avoiding state breaks caused by traditional hard switching. Third, this invention uses a gated fusion mechanism jointly driven by explicit uncertainty and implicit temporal features, enabling the fusion weights to change smoothly with the evolution of uncertainty, significantly reducing the flickering and jitter of the estimation output when the observation boundary conditions change. Finally, this invention fuses attitude quaternions under SO(3) manifold constraints, ensuring that the attitude representation always satisfies the unit norm constraint and rotational physical validity, thus it can be directly applied to UAV real-time control and navigation systems.

[0019] In summary, this invention provides a robust attitude estimation method and system for UAVs in complex mountainous and forest environments. Through unified preprocessing of multimodal observations, adaptive suppression of channel-level multi-scale noise, interactive bi-branch closed-loop inference, and collaborative design of uncertainty-aware temporal gating fusion, it systematically solves the key problems of existing technologies in scenarios with strong noise, intermittent GPS absence, and discontinuous multi-source fusion. It has strong theoretical innovation value and engineering application prospects. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall attitude estimation process for the UAV according to the present invention. Figure 2 A diagram illustrating the noise adaptive masking mechanism framework of this invention; Figure 3 A diagram illustrating the core mechanism of multi-scale dilated convolution in this invention; Figure 4 The dual-branch interaction diagram of this invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to specific embodiments. This embodiment describes the implementation form, parameter settings and calculation process of key modules without departing from the technical solution of the present invention, but those skilled in the art can make equivalent substitutions or adjustments based on this.

[0022] The purpose of this invention is to address the problems in UAV attitude estimation in complex mountainous and forested environments, such as coarse noise modeling, estimation degradation due to intermittent GPS signal loss, discontinuous multi-branch fusion process, and insufficient physical consistency of attitude representation. This invention proposes a robust UAV attitude estimation method and system for complex mountainous and forested environments. Based on unified multimodal observation preprocessing, this method introduces a channel-level multi-scale noise adaptive suppression mechanism, an interactive dual-branch closed-loop attitude inference architecture, and an uncertainty-aware temporal gating fusion mechanism. This enables deep collaboration between model-driven estimation and data-driven inference, resulting in continuous, stable, robust, and physically consistent attitude estimation results in complex environments. To achieve the above objectives, this invention adopts the following technical solution, the flowchart and structure of which are shown below. Figure 1 .

[0023] This invention first performs unified preprocessing on the multimodal observation data output by the Inertial Measurement Unit (IMU) onboard the UAV and the Global Navigation Satellite System (GPS). Considering that the IMU and GPS typically have different sampling frequencies and time bases in practical systems—IMU sampling frequencies generally range from 200Hz to 1000Hz, while GPS sampling frequencies generally range from 1Hz to 20Hz—directly concatenating or feeding these two types of data into the estimator will introduce additional errors due to time offset and sampling inconsistencies. Therefore, this invention uses a unified time base for synchronization processing, using the GPS time or system control time as the reference time, and performs linear interpolation on the high-speed IMU data to align them on the same time axis, forming a consistent discrete observation sequence. Specifically:

[0024] In this embodiment, the UAV platform adopts a quadcopter structure and is equipped with an inertial measurement unit (IMU) and a global navigation satellite system (GNSS) receiver. The IMU outputs three-axis angular velocity and three-axis specific force data, with a sampling frequency set to [value missing]. The global navigation satellite system outputs three-dimensional position and three-dimensional velocity information, with a sampling frequency set to [value missing]. The system's computing unit employs an embedded real-time processor with an ARM Cortex-A53 or similar chip architecture, running a Linux real-time operating system to execute real-time attitude estimation algorithms.

[0025] In the data processing phase, the first step is to synchronize the time of the multimodal observations. Let the original sampling time of the IMU be... GPS sampling time is Using GPS time as a reference time axis, first-order linear interpolation is performed on the IMU data for resampling, ensuring that it is accurate in each... Corresponding observations can be obtained at any time, thus constructing a unified observation sequence. At the same time, build This is only the local time step index within the sliding window of the multi-scale TCN network, and the corresponding observation vector is... The vector dimensions and physical meanings are completely identical; among them, This represents the interpolated triaxial angular velocity observations. This represents the interpolated triaxial specific force observation. and The initial GPS observations were used. Subsequently, a coordinate system transformation was performed on the IMU observations. Let the body coordinate system be... The navigation coordinate system is Using the attitude quaternion quaternion obtained from the previous moment Construct rotation matrix Convert the ratio into a navigation system expression. ;in, This corresponds to the attitude quaternion from the previous time step. This represents the rotation matrix generated at the corresponding time. The UAV attitude changes only slightly within adjacent sampling intervals, and using historical attitude data for coordinate transformation will not introduce significant errors. This transformation unifies the IMU and GPS observation reference coordinate systems, eliminating error propagation caused by inconsistencies in coordinate references.

[0026] To mitigate the impact of differences in the dimensions of different physical quantities on subsequent network learning and filter updates, all observations are standardized to meet a uniform numerical scale. ;in and This represents the statistical mean and standard deviation of the component on the training or calibration set. Alternatively, a min-max normalization form can be used depending on the specific scenario. ; Normalized observation data is mapped to a unified numerical range, which helps improve the stability of subsequent deep network training and makes different observation sources more balanced in numerical scale. Meanwhile, to avoid outliers exceeding the physical range directly contaminating the estimation process, this invention implements outlier detection and smoothing correction on the input sequence. Abrupt changes can be identified using a sliding window midpoint filter, the three-standard-deviation principle, or a threshold discrimination method based on local consistency. If a certain observation component... satisfy ; Then it is judged as an outlier, where and They represent respectively with With center length as The mean and standard deviation of the sliding window. This is the threshold coefficient. Outliers can be handled by methods such as neighbor interpolation, local smoothing replacement, or marking them as low-confidence observations.

[0027] After completing the multimodal observation preprocessing, such as Figure 3 As shown, this invention constructs a channel-level multi-scale noise adaptive suppression mechanism to address the issues of heterogeneous observation noise channels, temporal non-stationarity, and heavy-tailed distribution in complex mountainous and forested environments. The specific process and structure are detailed in [link to documentation]. Figure 2 For ease of description, this invention still uses a discrete-time nonlinear state-space model as its basic framework. Let the system state be: ; Its prediction model is: ; The observation model is: ; Unlike existing technologies, Treating it as a fixed matrix, this invention introduces a noise reliability mask that corresponds one-to-one with the observation channel. The data is generated online by a multi-scale temporal feature extraction network. Let the dimension of the observation vector be... ,but ; in, Indicates the first Each observation channel at time The unreliability level or suppression weight is determined. Since the anomalies in observation noise are not independent and identically distributed, and often have a distinct temporal structure, such as instantaneous spikes, short-to-medium-term correlated perturbations, long-term drift, and systematic shifts, a sliding overlapping window mechanism is used to construct a time-series observation window with a length of [missing information]. time window The data is then input into a multi-scale temporal convolutional network. This network employs a three-layer dilated convolutional structure, with dilation rates set to [values ​​to be specified in the original text]. The kernel size is 3, and the output channel dimensions of the three dilated convolution layers are set to 64, 128, and 64 respectively. The final output dimension is uniformly compressed to a 32-dimensional temporal feature vector for subsequent noise mask mapping.

[0028] The dataset consists of a hybrid of real-world mountain forest flight data and simulated disturbance data: the real-world dataset was collected using a quadcopter drone equipped with the IMU / GPS combination device of this invention, under conditions of mountainous forest obstruction, multiple paths, and gusts of wind disturbance, using raw observation time series; the simulated dataset is expanded based on the drone's rigid body motion model superimposed with three types of typical noise (instantaneous spike noise, intermediate time-correlated color noise, and long-term temperature drift), with the training set, validation set, and test set divided in a 7:2:1 ratio. This scheme does not use artificial ground truth labels, but adopts the self-supervised composite loss function mentioned above to achieve unsupervised training, eliminating the need for attitude and noise ground truth labeling.

[0029] The optimizer chosen is the AdamW optimizer, with a base learning rate set to... The cosine annealing learning rate decay strategy is adopted; batch size = 64, total training iterations = 200; weight decay coefficient is set as follows. The gradient clipping threshold is set to 1.0 to prevent gradient explosion; the hardware training platform uses GPU offline training, and after training is completed, the network weights are exported and deployed to the onboard embedded processor to achieve real-time forward inference.

[0030] A temporal convolutional network (TCN) based on dilated convolutions is used as a unified encoder to obtain temporal feature representations. ;in (⋅) is composed of multiple residual dilated convolutions, with different dilation rates corresponding to different time scales. The dilation rate of each layer increases exponentially, thus covering multiple time scales in a single network: local scale (small dilation rate) is used to capture spikes, abrupt changes, and observational discontinuities; medium scale (medium dilation rate) is used to model short- and medium-term correlated noise and modal consistency changes; long scale (large dilation rate) is used to perceive slow drift, low-frequency trends, and systematic shifts caused by occlusion. These are the encoded temporal features.

[0031] Then, a channel-level noise mask is generated using linear mapping and the Sigmoid activation function. ;in, and For learnable parameters, This represents the element-wise Sigmoid function. Since TCN uses multiple dilated convolutions, its... The convolution operation of a layer in the time dimension can be written as: ; in, The kernel length is 1. For the first The void ratio of the layer. This is determined by setting... It grows exponentially, for example It can take into account short-term, medium-term, and long-term noise patterns without significantly increasing the number of parameters, thereby capturing instantaneous spikes, gust-induced thick-tailed disturbances, and long-term temperature drift changes, such as... Figure 3 As shown, different noises are addressed by TCNs of different dimensions.

[0032] In this invention, the "soft suppression" does not involve hard thresholding of the observed data, but rather reduces the impact of outlier observations on state estimation through a continuously adjustable method. For ease of understanding, this soft suppression process can be equivalently represented as a weighted scaling of the observation residuals: ; in, To observe the residuals, This is the diagonal weight matrix generated by the noise mask, and the values ​​of its diagonal elements range from... This is used to characterize the reliability of each observation channel at the current moment. When the noise of a certain channel is large, its corresponding weight is reduced, thereby reducing the contribution of the channel's residual to the state update and achieving flexible suppression of abnormal observations.

[0033] In a specific implementation, the present invention preferably achieves the above equivalent effect by adaptively scaling the observation noise covariance matrix, without directly modifying the original observation data, thereby maintaining the stability and mathematical consistency of the filtering structure.

[0034] After obtaining the noise mask, this invention performs a differentiable diagonal recalibration of the basic observation noise covariance matrix. Let the basic covariance be... Construct a diagonal scaling matrix: ; in, The adaptive observation noise covariance is defined as the scaling factor: Or, under the condition of diagonal approximation, it can be written as: ;in, This represents the diagonal element of the fundamental noise covariance. This construction method has two advantages. First, when the mask value of a channel is large, it indicates that the channel is currently unreliable, and the corresponding covariance is amplified. The filter will automatically reduce its confidence in the observations of that channel, achieving flexible suppression rather than simple elimination. Second, because... Let be a diagonal positive matrix, if If the matrix is ​​a symmetric positive definite matrix, then It remains symmetric and positive definite, thus satisfying the mathematical requirements of Kalman filter algorithms for noise covariance.

[0035] In the filtering update process, this invention preferably uses unscented Kalman filtering as the model-driving branch. Its core lies in propagating the mean and covariance under the approximate nonlinear transformation through sigma point propagation. Let the mean of the predicted state be... The predicted covariance is In the update phase, the predicted observation mean is calculated first. With observation covariance ,in: ; ; ; ; here, Indicates the predicted observation covariance. This represents the cross-covariance between the state and the observation. Indicates Kalman gain. Because In this invention, the filter is dynamically adjusted based on channel reliability, so the filter can automatically adjust the update amplitude when faced with different types of noise.

[0036] To ensure the noise-adaptive mask possesses physical plausibility rather than solely relying on supervised label fitting, this invention further constructs a self-supervised composite loss function for end-to-end training. This loss function comprehensively considers observation residual consistency, rigid body dynamics consistency, and mask prior constraints, and can be expressed as: ;

[0037] in, For residual consistency loss, and The balance coefficient is used. The residual consistency loss can be defined as: ;

[0038] This encourages matching the observed residuals to the estimated noise level after adaptive covariance adjustment. (Dynamic consistency loss) This is used to constrain the estimation state to satisfy the laws of rigid body motion. For example, in a simplified form, it can be written as: ;

[0039] in, Indicates the sampling time interval. This represents the gravity term. This term ensures that the state evolution still conforms to basic kinematic laws after noise suppression. Mask prior loss. To encourage mask sparsity and temporal smoothing, it can be written as ;in, To smooth out the weights, the first term encourages activation of suppression only on necessary channels, avoiding over-amplification of all channels simultaneously; the second term suppresses severe mask jitter, preventing noise weights from flickering frequently over time.

[0040] After implementing noise-robust preprocessing, this invention further proposes an interactive dual-branch closed-loop attitude inference architecture for intermittent GPS outages. This architecture consists of a data-driven inertial inference branch and a model-driven filtering estimation branch. The former provides continuous inertial estimation capability when GPS is interrupted, while the latter provides physically constrained absolute state correction when GPS is available. Specifically, the data-driven branch employs a recurrent neural network, preferably a long short-term memory network, which outputs state increments or relative state estimates based on IMU timing inputs. Let its input window be... ,but ;in, Indicates a hidden state. This represents the inertial inference result of branch A. The model-driven branch performs prediction and updates via UKF, and outputs... and covariance .

[0041] This invention differs from existing technologies that simply splice or weight the two branches at their ends; instead, it constructs a bidirectional closed-loop information interaction mechanism. See the specific structure below. Figure 4 On the one hand, the explicit uncertainties generated by the model-driven branch are injected into the data-driven branch, allowing it to perceive the current reliability level of the filtering branch during time series modeling. Let the filtering branch at time... The innovation generated is ; Innovation covariance is Then the normalized squared information can be expressed as: ; Furthermore, state uncertainty can also be represented by the covariance trace. ;

[0042] The aforementioned quantities, together with explicit physical indices such as the diagonal elements of the observed residuals, constitute an additional feature vector. ; The data is then input into the LSTM along with the IMU time series features to enhance its ability to detect changes in system reliability. On the other hand, the data-driven branch outputs corrections for process and observation noise based on historical motion patterns and the current time series context, written as... ; in, This represents the parameterized mapping. It is then applied to the noise model of the filtering branch: ; in, This represents the observed noise covariance after noise masking. Thus, in high-dynamic maneuvering, gust interference, or GPS degradation scenarios, the filtering branch can adjust the noise setting in real time based on the context information provided by the learning branch, avoiding gain mismatch and state divergence caused by fixed noise assumptions.

[0043] When GPS is readily available, the model-driven branch, having obtained absolute position and velocity observations, typically achieves higher long-term accuracy and stronger physical consistency in its estimations. When GPS signals are intermittently interrupted or significantly degraded in quality, the data-driven branch, relying on the learning capability of inertial time-series patterns, can maintain a relatively smooth relative state recursion for a short period. This invention does not employ a simple hard-switching approach, but rather uses an uncertainty-aware temporal gating fusion mechanism to continuously adjust the relative weights of the two branch outputs in the time dimension.

[0044] Specifically, this invention constructs a fused feature vector, combining the explicit physical uncertainty index output by the filtering branch with the hidden temporal features of the learning branch to form... ;in, The GPS signal quality index can be represented by the signal-to-noise ratio, the number of satellites, the DOP value, or a combination thereof. This feature vector is input into a gated recurrent unit to obtain smoothly varying fusion weights. : ; ;in, This indicates the hidden state of the GRU. Indicates time The weight of branch A, This corresponds to the weight of branch B. Since GRU has memory capabilities in the time dimension, therefore... It is no longer a simple function of instantaneous indicators, but rather takes into account the evolutionary history of uncertainty. In this way, as the GPS signal gradually fades from usable, the weights can smoothly shift from the filtering branch to the inertial branch; when the GPS signal recovers, it can also gradually return to the filtering branch as the main component, thereby avoiding the estimation jumps and control jitters caused by traditional hard handover.

[0045] In the fused output layer, this invention uses a linear convex combination method to fuse position and velocity, that is... ; ;in, The positions and velocities of the output from branch A are estimated respectively. The position and velocity estimates for branch B are given respectively. For the attitude variables, since the quaternions lie on the unit quaternion manifold corresponding to the three-dimensional rotating Lie group SO(3), their geometry cannot be directly ignored for ordinary Euclidean weighting. Let the two branch output quaternions be respectively... and First, calculate their dot product: ;like This indicates that the two quaternions are opposite representations in opposite half-spaces of the parameter space. Although they correspond to the same or approximately rotated values, direct interpolation will vary along a long path. Therefore, we invert one of the quaternions, denoted as . Then perform a weighted summation. Then perform L2 normalization. This ensures that the final quaternion satisfies the unit norm constraint. This ensures that the attitude output always remains within the valid SO(3) representation space. Through this symbol alignment and normalization process, the present invention effectively avoids angle jumps and physical invalidity problems during attitude interpolation.

[0046] In the subsequent output stage, this invention can further perform amplitude limiting, low-pass smoothing, and consistency verification on the attitude estimation results. For example, the roll angle and pitch angle can be limited to the controllable range of the aircraft, and a first-order filter can be applied to the output signal. ;in, Indicates the current fusion output. This indicates the output after smoothing. This represents the filter coefficient. This processing can further suppress high-frequency jitter, making the final output more suitable as a closed-loop feedback quantity for flight control.

[0047] In actual testing, flight verification was conducted in a complex mountainous and forested environment, including conditions such as tree obstruction, gusts of wind, and intermittent GPS signal loss. Results showed that this implementation method maintained stable attitude estimation even when GPS signal loss exceeded half the time, with no significant divergence or jumps. The attitude error was significantly reduced compared to traditional methods, verifying the effectiveness of the proposed method. In this implementation, the fused attitude angles, quaternions, position, and velocity state variables can be sent to the flight control system in real time and used as inputs to the attitude control loop and navigation control loop. These inputs are then used by the PID controller or other controllers to calculate control quantities to control the UAV motor speed and flight attitude, achieving stable hovering, trajectory tracking, and disturbance-resistant flight.

[0048] This invention also provides a robust attitude estimation system for unmanned aerial vehicles (UAVs) in complex mountainous and forested environments, comprising: Data preprocessing module: performs unified time reference synchronization and coordinate system transformation on the multimodal observation data output by the inertial measurement unit (IMU) and the global satellite navigation system (GPS) on the UAV, and performs normalization processing, outlier detection and smoothing correction on the unified observation vector to obtain the preprocessed discrete observation sequence; Noise adaptive suppression module: It uses a time convolutional network (TCN) based on dilated convolution to extract multi-scale temporal features from the preprocessed discrete observation sequence, and generates a noise reliability mask online for each observation channel based on the multi-scale temporal features. Then, it uses the mask to adaptively diagonally scale the basic observation noise covariance matrix to obtain a dynamically adjusted observation noise covariance matrix. Dual-branch attitude inference module: Constructs an interactive dual-branch closed-loop attitude inference architecture, which includes a data-driven inertial inference branch and a model-driven filter estimation branch; wherein: The model-driven filter estimation branch uses the dynamically adjusted observation noise covariance matrix to update the filter and outputs an explicit uncertainty index. The data-driven inertial inference branch performs state inference based on IMU timing input and receives the explicit uncertainty index as additional input to sense the reliability level of the filtering branch, while outputting the correction amount of process noise and / or observation noise to the filtering branch. Temporal Gated Fusion Module: Through an uncertainty-aware temporal gated fusion mechanism, the output of the data-driven branch and the output of the model-driven branch are continuously fused in the time dimension to obtain the final attitude estimation result.

[0049] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0050] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A robust attitude estimation method for unmanned aerial vehicles (UAVs) in complex mountainous forest environments, characterized in that, Includes the following steps: Step 1: Perform unified time reference synchronization and coordinate system transformation on the multimodal observation data output by the inertial measurement unit (IMU) and the global satellite navigation system (GPS) on the UAV, and normalize the unified observation vectors and perform outlier detection and smoothing correction to obtain the preprocessed discrete observation sequence. Step 2: Extract multi-scale temporal features from the preprocessed discrete observation sequence using a temporal convolutional network (TCN) based on dilated convolution, and generate a noise reliability mask online for each observation channel based on the multi-scale temporal features. Then, use the mask to adaptively diagonally scale the basic observation noise covariance matrix to obtain a dynamically adjusted observation noise covariance matrix. The adaptive diagonal scaling includes: constructing a diagonal scaling matrix: ;in, For the first Each observation channel at time The mask value, The scaling factor is used to define the adaptive observation noise covariance matrix as follows: Or, under the condition of diagonal approximation, it can be written as: ;in, Based on the observation noise covariance matrix, It is its diagonal element; when the mask value of a certain channel increases, the corresponding covariance is amplified, thereby reducing the contribution of channel observations to the filter update; Step 3: Construct an interactive dual-branch closed-loop attitude inference architecture, which includes a data-driven inertial inference branch and a model-driven filter estimation branch; wherein: The model-driven filtering estimation branch uses the dynamically adjusted observation noise covariance matrix to perform filtering updates and outputs an explicit uncertainty index. Specifically, the model-driven filtering estimation branch employs an unscented Kalman filter (UKF), which uses the dynamically adjusted observation noise covariance matrix to calculate the observation covariance and Kalman gain, and performs state updates. The explicit uncertainty index includes at least one of the following: normalized innovation squared, state covariance trace, and diagonal elements of the observation residual. The data-driven inertial inference branch performs state inference based on IMU time-series input and receives the explicit uncertainty index as additional input to perceive the reliability level of the filtering branch. Simultaneously, it outputs corrections for process noise and / or observation noise to the filtering branch. Specifically, the data-driven branch employs a recurrent neural network, specifically a Long Short-Term Memory (LSTM) network, to output a relative state estimate based on the IMU time-series input. The data-driven branch receives the explicit uncertainty index and uses it, along with the IMU time-series features, as input to the recurrent neural network to enhance the perception of changes in system reliability. The corrections output by the data-driven branch to the filtering branch include process noise covariance corrections and observation noise covariance corrections, used to adjust the noise setting of the filtering branch in real time. Step 4: Through the uncertainty-aware temporal gating fusion mechanism, the output of the data-driven inertial inference branch and the output of the model-driven filtering estimation branch are continuously fused in the time dimension to obtain the final attitude estimation result.

2. The method according to claim 1, characterized in that, The unified time reference synchronization in step 1 includes: aligning the three-axis angular velocity and three-axis specific force output by the IMU with the three-dimensional position and three-dimensional velocity output by the GPS based on the unified time reference; the coordinate system transformation includes transforming the observations in the IMU carrier coordinate system to the Northeast Sky ENU geographic coordinate system.

3. The method according to claim 1, characterized in that, The normalization process is performed in a standardized form: ; or minimum-maximum normalized form: ; in, and These are the mean and standard deviation of the corresponding components on the training set or calibration set, respectively; the outlier detection and smoothing correction include: identifying abrupt change points using a sliding window mean filter, the three-standard-deviation principle, or a threshold discrimination method based on local consistency, and processing outliers by using neighboring sample interpolation, local smoothing replacement, or marking them as low-confidence observations.

4. The method according to claim 1, characterized in that, The TCN is composed of multiple residual dilated convolutions, with different dilation rates corresponding to different time scales. The dilation rate of each layer increases exponentially. The local scale is used to capture spikes and abrupt changes, the medium scale is used to model short- and medium-term correlated noise, and the long scale is used to sense slow drift and systematic shifts. The noise reliability mask is generated by fully connected mapping and Sigmoid activation through the temporal features extracted by the TCN. The mask value is in the range of (0,1) and represents the unreliability of the corresponding observation channel.

5. The method according to claim 1, characterized in that, The uncertainty-aware temporal gating fusion mechanism includes: constructing a fusion feature vector, which contains the normalized squared innovation, state covariance trace, GPS signal quality index, and hidden state features of the data-driven branch output; and inputting the fusion feature vector into a gated recurrent unit (GRU) to obtain smoothly varying fusion weights. The fusion weights are used to perform a linear convex combination of the position and velocity estimates of the two branches, and to perform a weighted normalization fusion of the attitude quaternions of the two branches after sign alignment.

6. The method according to claim 1, characterized in that, It also includes a self-supervised composite loss function for end-to-end training of the generation network of the noise reliability mask; the composite loss function includes residual consistency loss, dynamic consistency loss and mask prior loss, wherein the residual consistency loss encourages the observed residuals after adaptive covariance adjustment to match the estimated noise level, the dynamic consistency loss constrains the estimated state to satisfy the rigid body motion law, and the mask prior loss encourages mask sparsity and temporal smoothness.

7. A system for implementing the robust attitude estimation method for unmanned aerial vehicles (UAVs) in complex mountainous forest environments as described in any one of claims 1-6, characterized in that, include: Data preprocessing module: performs unified time reference synchronization and coordinate system transformation on the multimodal observation data output by the inertial measurement unit (IMU) and the global satellite navigation system (GPS) on the UAV, and performs normalization processing, outlier detection and smoothing correction on the unified observation vector to obtain the preprocessed discrete observation sequence; Noise adaptive suppression module: It uses a time convolutional network (TCN) based on dilated convolution to extract multi-scale temporal features from the preprocessed discrete observation sequence, and generates a noise reliability mask online for each observation channel based on the multi-scale temporal features. Then, it uses the mask to adaptively diagonally scale the basic observation noise covariance matrix to obtain a dynamically adjusted observation noise covariance matrix. Dual-branch attitude inference module: Constructs an interactive dual-branch closed-loop attitude inference architecture, which includes a data-driven inertial inference branch and a model-driven filter estimation branch; wherein: The model-driven filter estimation branch uses the dynamically adjusted observation noise covariance matrix to update the filter and outputs an explicit uncertainty index. The data-driven inertial inference branch performs state inference based on IMU timing input and receives the explicit uncertainty index as additional input to sense the reliability level of the filtering branch, while outputting the correction amount of process noise and / or observation noise to the filtering branch. Temporal Gated Fusion Module: Through an uncertainty-aware temporal gated fusion mechanism, the output of the data-driven inertial inference branch and the output of the model-driven filtering estimation branch are continuously fused in the time dimension to obtain the final attitude estimation result.

Images