TCN-LSTM series network multi-source state real-time sensing method and system for unmanned aerial vehicle control

Abstract

The invention provides an unmanned aerial vehicle control-oriented TCN-LSTM series network multi-source state real-time sensing method and system. The method comprises the following steps of: constructing a deep neural network model formed by connecting a TCN and an LSTM in series; the speed, attitude and position information of the unmanned aerial vehicle is acquired through cooperation of the RTK-GPS, the INS and the IMU, and time alignment is carried out on IMU data and RTK-GPS / INS data by using an interpolation method; the collected data is divided into a training set, a verification set and a test set, and a composite loss function fusing RMSE, MAE, quaternion angle errors and offset errors is adopted to train and evaluate the model; serializing sensor data based on a sliding time window strategy to form an input feature vector, and standardizing IMU data according to channels; and inputting the standardized sensor data into the TCN-LSTM model, and outputting a state estimation result for navigation, path planning or closed-loop control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle flight control and state perception, in particular to a TCN-LSTM series network multi-source state real-time perception method and system for unmanned aerial vehicle control, and belongs to the cross-technology of intelligent aircraft, attitude and heading estimation, and artificial intelligence in aerospace. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle autonomous flight technology, achieving high-precision, real-time, and robust state estimation in complex dynamic environments (such as urban low altitude, electromagnetic interference area) has become the core foundation of flight safety and closed-loop control. Current technical solutions mainly fall into two categories: model-driven methods based on traditional filtering and time-series modeling methods based on deep learning. However, both have significant limitations.

[0003] I. Model dependency and environmental adaptability defects of traditional filtering methods:

[0004] Existing industrial-grade unmanned aerial vehicle systems generally use Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), or Bayesian fusion framework to estimate the attitude, velocity, and position of the unmanned aerial vehicle by fusing IMU, GPS, and other sensor data. Although such methods perform stably in ideal environments, they expose inherent defects in complex scenarios: strong model dependency, the filtering algorithm needs to predefine accurate system dynamics equations and noise distribution models (such as kinematic equations, IMU error models). For example: the VINS-Mono solution relies on visual-inertial tightly coupled optimization, but requires manual calibration of camera-IMU transformation parameters. Model errors lead to attitude drift in dynamic environments; Pan Jian verified the effectiveness of Kalman filtering in rotorcraft state estimation, but emphasized that "model errors will significantly reduce filtering accuracy", especially in unmodeled dynamics such as sudden wind disturbances; Wang et al.'s EKF framework needs to predefine the state transition matrix, and non-Gaussian noise (such as GPS multipath effects) will disrupt filtering stability. The ability to process multi-source asynchronous data is insufficient, and in the face of IMU (high frequency), GPS (low frequency), and other asynchronous sensors: Zhou et al. used double Kalman filters (vision and IMU) to compensate for data delays, but increased the computational load by 15%; Conte et al.'s Bayesian framework needs to introduce pseudo-measurements to fill in data gaps, leading to cumulative state estimation bias. Therefore, traditional methods are heavily dependent on prior models, and in dynamic environments (such as sudden wind speed changes), non-Gaussian noise (sensor interference), and multi-source asynchronous data scenarios, the estimation accuracy and robustness decrease dramatically, and additional recovery mechanisms are needed when the complete state is lost.

[0005] II. Time-series modeling defects and one-sided state estimation of deep learning methods:

[0006] In recent years, sequence models such as RNN / LSTM are introduced into state estimation tasks, which partially alleviates the model dependence problem, but there are still unresolved problems: imbalance in modeling long-term and short-term dependencies, LSTM is prone to gradient vanishing when capturing long-term state dependencies, and it is difficult to balance high-frequency local changes. For example: Liang Shaojun et al. pointed out that "pure LSTM prediction has a lag error of 12% in high-speed maneuvering scenarios"; Benziane verified that the estimation error of LSTM for attitude angular velocity increases exponentially with the sequence length. Although TCN optimizes local feature extraction through dilated convolution, it has not been maturely applied in the laser radar scheme of unmanned aerial vehicle full state estimation. The state estimation target is one-sided, and the existing method focuses on optimizing position error, ignoring the joint modeling of attitude and system bias. Ding Yinying et al. visual landing scheme only minimizes position deviation, and attitude estimation relies on post-processing of extended Kalman filter; Emphasizing that "attitude estimation requires independent wind speed and torque mode", end-to-end learning is not realized; the fault diagnosis research of Reg et al. confirmed that ignoring the IMU zero bias error will increase the attitude estimation drift rate by 20%.

[0007] In summary, the current technology has poor model robustness of filtering method, and the time series modeling capability of learning method is insufficient; asynchronous sensor fusion calculation is complex, which is difficult to meet the real-time demand of embedded platform; the optimization of position, attitude and bias error is separated, which cannot support high stability flight control. SUMMARY

[0008] The purpose of the present application is to provide a TCN-LSTM serial network multi-source state real-time perception method for unmanned aerial vehicle control, which replaces the traditional filter with an end-to-end deep learning architecture, uses TCN-LSTM serial network to collaboratively model local and long-term time series features, designs a composite loss function that integrates position / attitude / bias error, and constructs a lightweight real-time processing framework, ultimately realizing high-precision, low-latency, full-dimensional estimation of unmanned aerial vehicle state in complex environments.

[0009] The purpose of the present application is achieved by the following technical solutions:

[0010] A TCN-LSTM serial network multi-source state real-time perception method for unmanned aerial vehicle control, comprising the following steps:

[0011] (1) Construct a deep neural network model composed of a time series convolution network (TCN) and a double-layer long short-term memory network (LSTM) in series;

[0012] (2) Collect the speed, attitude and position information of the unmanned aerial vehicle through RTK-GPS, high-precision inertial navigation system (INS) and inertial measurement unit (IMU) in collaboration, and use interpolation method to time-align the IMU data and RTK-GPS / INS data;

[0013] (3) The collected data is divided into a training set, a validation set and a test set, and a composite loss function fusing RMSE, MAE, quaternion angle error and bias error is used to train and evaluate the model;

[0014] (4) The sensor data is sequenced based on a sliding time window strategy to form an input feature vector with a length of 10 time steps, and the IMU data is processed according to a channel standardization;

[0015] (5) The real-time collected standardized sensor data is input into the trained TCN-LSTM model to output the state estimation results of the current position, speed and attitude of the unmanned aerial vehicle, which are used for navigation, path planning or closed-loop control.

[0016] As a more optimal technical solution of the application, the TCN-LSTM series structure in step (1) comprises:

[0017] The TCN part comprises three layers of dilated causal convolution layers, the convolution kernel width is 3, the channel number is 64, 32 and 16 in turn, and the dilated coefficient is 1, 2 and 4 in turn;

[0018] Each layer of TCN comprises residual connection, weight normalization, ReLU activation function and dropout mechanism, and the dropout rate is 0.2;

[0019] The feature sequence output by the TCN is input into a double-layer LSTM, the number of hidden units of each layer of the LSTM is 128 and 64 respectively, and the time dependence relationship is modeled using tanh and sigmoid activation functions;

[0020] Finally, the state estimation results are generated through a full connection layer.

[0021] As a more optimal technical solution of the application, the composite loss function in step (3) is defined as:

[0022]

[0023]

[0024] F=λ1A1+λ2A2+λ3A3+λ4A4

[0025] In the formula: λ1=1.0, λ2=1.0, λ3=5.0, λ4=0.1; is the estimated accelerometer bias of the model; is the real accelerometer bias; is the estimated gyroscope bias of the model; is the real gyroscope bias; is the estimated attitude quaternion; is the real attitude quaternion; a velocity vector of the estimated UAV at time t; a real velocity of the UAV; a position vector of the estimated UAV at time t for model evaluation; a real position corresponding to time t; N is the number of time steps.

[0026] As a more optimal technical solution of the application, the linear interpolation method is used for time alignment in step (2) to align the sampling time of GPS and INS data to the IMU timestamp.

[0027] As a more optimal technical solution of the application, the step length of the sliding time window in step (4) is 10 seconds, and each window contains a group of UAV state information sequences.

[0028] As a more optimal technical solution of the application, the real-time state estimation of step (5) meets the deployment requirements of the edge computing platform, and the single inference time is less than 5ms, supporting a control frequency of more than 100Hz.

[0029] As a more optimal technical solution of the application, the quaternion angular error is defined as:

[0030]

[0031] wherein: θ t is the estimated attitude rotated to the real attitude the minimum rotation angle required around a rotation axis.

[0032] As a more optimal technical solution of the application, the proportion of the training set in step (3) is not less than 70%, and the validation set is used for hyperparameter tuning and early stopping strategy.

[0033] As a more optimal technical solution of the application, the state estimation result in step (5) includes: three-dimensional position coordinates, three-axis velocity vector, pitch angle / roll angle / yaw angle attitude angle.

[0034] The application also aims to provide a UAV control system that uses the state estimation result output by any one of the methods of claims 1-9 to generate navigation instructions or closed-loop control signals in real time.

[0035] Beneficial effects

[0036] The application realizes the following breakthrough improvements in the field of UAV state estimation through the improvement of TCN-LSTM series architecture, multi-source asynchronous sensor cooperation, composite loss function and edge lightweight deployment, and can be widely applied to various UAV systems, especially suitable for urban low-altitude, complex environment, multi-source navigation fusion and other scenarios that require high-precision state perception.

[0037] The TCN-LSTM series architecture is adopted, the local time sequence feature and the long-term dependence modeling are considered, and the time perception ability of state estimation is improved. The TCN layer extracts high-frequency local features (such as sudden disturbance response) through dilated causal convolution (dilation coefficients 1 / 2 / 4), and the LSTM layer models long-term state evolution (such as continuous attitude drift), solves the problems of gradient disappearance and information loss of pure LSTM. Residual connection + weight normalization ensures the stability of deep network training, and avoids the degradation risk of traditional RNN.

[0038] The application adopts multi-source asynchronous sensor cooperation to enhance the adaptability of the model to complex data distribution and noise interference; RTK-GPS provides global positioning, INS outputs high-precision attitude, and IMU captures high-frequency motion, covering all state dimensions; the asynchronous data is aligned to a unified timestamp, and the dimension difference and interpolation error of the sensor are standardized and compensated.

[0039] The application adopts a composite loss function to construct a training target, and comprehensively optimizes the state estimation results such as position, attitude and bias; the quaternion angle error directly quantifies the rotation deviation, solves the suboptimal problem of attitude estimation; models the system error such as IMU zero offset, and suppresses long-term drift.

[0040] The application adopts lightweight deployment in an edge computing platform (such as an embedded device carried by a UAV) to realize real-time estimation and high-robustness feedback. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 It is a schematic diagram of a UAV state real-time estimation method based on a multi-sensor and TCN-LSTM series hybrid network.

[0042] Figure 2 It is a TCN-LSTM series hybrid network structure diagram. DETAILED DESCRIPTION

[0043] The application will be described in detail below in combination with specific embodiments.

[0044] As shown in Figure 1 , the application provides a TCN-LSTM series network multi-source state real-time perception method for UAV control, comprising the following steps:

[0045] (1) constructing a deep neural network model composed of TCN and LSTM in series;

[0046] (2) collecting the speed, attitude and position information of the UAV through RTK-GPS, INS and IMU cooperation, and aligning the IMU data and RTK-GPS / INS data in time by using interpolation method;

[0047] (3) The collected data is divided into a training set, a validation set and a test set, and a composite loss function combining RMSE, MAE, quaternion angle error and bias error is used to train and evaluate the model;

[0048] (4) The sensor data is sequenced based on a sliding time window strategy to form an input feature vector with a length of 10 time steps, and the IMU data is processed according to the channel standardization;

[0049] (5) The real-time collected standardized sensor data is input into the trained TCN-LSTM model, and the state estimation results of the current position, velocity and attitude of the unmanned aerial vehicle are output, which are used for navigation, path planning or closed-loop control.

[0050] The specific implementation of the above method is as follows:

[0051] 1. A deep neural network for real-time unmanned aerial vehicle state estimation is constructed, and the overall structure is composed of TCN and double-layer LSTM in series.

[0052] The TCN part contains three layers of dilated causal convolution, each layer of convolution kernel width is 3, the number of channels is 64, 32 and 16 respectively, and the dilated coefficient is 1, 2 and 4 respectively. Each layer of TCN contains residual connection, weight normalization, ReLU activation function and dropout (rate of 0.2), and the time sequence dimension is kept consistent by Zero Padding. The input of TCN is a state history sequence with a length of 10 time steps. The feature sequence output by TCN is used as the input of double-layer LSTM, and the number of hidden units of LSTM is 128 and 64 respectively, and the long-term time dependence relationship is modeled using tanh and sigmoid activation functions. The final output generates the state estimation results of the unmanned aerial vehicle through a fully connected layer. The loss function used is a composite loss function based on RMSE, MAE, quaternion angle error and bias estimation error.

[0053] The calculation process of the TCN-LSTM cascade network:

[0054] TCN module calculation process:

[0055] Each layer of TCN adopts dilated causal convolution structure. For the l-th layer of TCN, the output is:

[0056]

[0057] In the formula: d (l) is the dilated coefficient of the l-th layer, which is 1, 2 and 4 in turn; Conv1D represents 1 convolution operation, the convolution kernel width is 3, and the output channel is 64, 32 and 16 in turn; Norm represents weight normalization; Dropout rate is 0.2, residual connection keeps information transmission, and zero padding is used to keep the time sequence dimension unchanged.

[0058] TCN final output:

[0059] H TCN ∈R 10×16 ;X LSTM =H TCN ∈R 10×16 ;x t ∈X LSTM

[0060] The time step T of the LSTM module is consistent with the TCN module, so the input feature vector of the LSTM is the output feature vector of the TCN.

[0061] LSTM module calculation process:

[0062] Forget gate:

[0063] f t =σ(W fh h t-1 +W fx x t +b f );

[0064] Input gate:

[0065] i t =σ(W ih h t-1 +W ix x t +b i );

[0066] Candidate memory update:

[0067]

[0068] Memory cell update:

[0069]

[0070] Output gate:

[0071] o t =σ(W oh h t-1 +W ox x t +b o );

[0072] Hidden state output:

[0073]

[0074] In the formula, f t represents the forget gate output; i t represents the input gate output; is the current input unit state; x t is the input data at time t in the time series; o t represents the output gate output; h t-1 is the output result at time t-1; W fh is the output weight of the forget gate to the unit; W fx is the input weight of the forget gate to the unit; W oh is the weight of the unit to the input gate; W ox represents the weight of the current output gate to the unit output; b f represents the bias of the forget gate; b i represents the bias of the input gate; b o represents the bias of the output gate; σ() is a sigmod function; c t represents the cell memory of the unit state.

[0075] The position error is calculated based on RRMSE (Root Mean Square Error):

[0076]

[0077] In the formula: is the position vector of the estimated unmanned aerial vehicle at time t; is the true position corresponding to time t; N is the number of time steps.

[0078] The speed error is calculated based on MAE (Mean Absolute Error):

[0079]

[0080] In the formula: is the speed vector of the estimated unmanned aerial vehicle at time t; is the true speed of the unmanned aerial vehicle.

[0081] The attitude error is calculated based on the angle difference of quaternions:

[0082]

[0083] In the formula: is the estimated attitude quaternion; is the true attitude quaternion; θ t is the minimum rotation angle required to rotate from the estimated attitude to the true attitude around a certain rotation axis.

[0084] The accelerometer bias error and the gyroscope bias error are calculated based on the L2 norm:

[0085]

[0086] In the formula: Errora Error for accelerometer bias estimation error w Error for gyroscope bias estimation error Model estimated accelerometer bias True accelerometer bias Model estimated gyroscope bias True gyroscope bias

[0087] Construct the composite loss function:

[0088]

[0089] F = λ1A1 + λ2A2 + λ3A3 + λ4A4

[0090] Where: λ1 = 1.0, λ2 = 1.0, λ3 = 5.0, λ4 = 0.1, which can be adjusted according to the verification results.

[0091] 2. Model training

[0092] Data set collection: Use RTK-GPS (Real-Time Kinematic Positioning Technology) and high-precision INS (Inertial Navigation System) equipment and IMU to detect the same space-time unmanned aerial vehicle target, and obtain the speed, acceleration, angular acceleration, attitude information and position information of the unmanned aerial vehicle.

[0093] Data processing: Preprocess the data obtained by using RTK-GPS (Real-Time Kinematic Positioning Technology) and high-precision INS (Inertial Navigation System) equipment, including data cleaning, invalid data filtering, data format conversion and other operations.

[0094] Data set division: divide the data into training set, validation set and test set, where the training set accounts for not less than 70%, the first part is the training set for model parameter optimization, the second part is the validation set for hyperparameter tuning and early stopping, and the third part is the test set for evaluating the final performance of the model.

[0095] Serialized data

[0096] (1) Time alignment of multi-sensor data: Align RTK-GPS data and INS data to IMU timestamp through linear interpolation to generate data stream under unified time reference.

[0097] (2) Splice acceleration, angular velocity, position information according to time sequence to form multi-dimensional feature vector:

[0098] F t = [a x , a y , a z , ωx ,ω y ,ω z ,x,y,z]

[0099] (3) Define the window length T and step size S, and construct the time series sample matrix with the continuous feature vectors.

[0100] X i = [F iS , F iS+1 ,..., F iS+T-1 ] T

[0101] Input the time series sample matrix X i into the network structure of three-layer TCN and double-layer LSTM cascade connection for training, and complete the training and evaluation of the model based on the four indicators of RMSE, MAE, quaternion angle error, and bias estimation error in step 1.

[0102] 3. Model application

[0103] Use IMU, GPS, and barometer to detect the same space-time unmanned aerial vehicle, and obtain the acceleration, angular velocity, and position information of the unmanned aerial vehicle. Use linear interpolation to align the unmanned aerial vehicle position data obtained by GPS and barometer to the IMU timestamp, and realize multi-sensor data synchronization:

[0104] IMU timestamp sequence:

[0105] T IMU = {t0, t1,..., t n}

[0106] GPS timestamp sequence:

[0107] T GPS = {t'0, t1',..., t' m}

[0108] Barometer timestamp sequence:

[0109] T baro = {t'0', t1”,..., t' k '}

[0110] Align the GPS and barometer data to each timestamp t i of the IMU using linear interpolation:

[0111] (1) Find two adjacent low-frequency sampling points surrounding t i :

[0112] For GPS:

[0113] t' j ≤ ti <t' j+1 j ,t' j+1 ∈T GPS

[0114] For barometer:

[0115] t″ k ≤t i <t″ k+1 , where t″ k ,t″ k+1 ∈T baro

[0116] (2) Calculate interpolation weight:

[0117] GPS interpolation weight:

[0118]

[0119] Barometer interpolation weight:

[0120]

[0121] (3) Linear interpolation calculation:

[0122] GPS interpolation value:

[0123] y GPS (t i )=(1-α GPS )·y GPS (t' j )+α GPS ·y GPS (t' j+1 )

[0124] Barometer interpolation value:

[0125] y baro (t i )=(1-α baro )·y baro (t″ k )+α baro ·y baro (t″ k+1 )

[0126] In the formula: y GPS (t i ) is the interpolation result of GPS at target time t i ; y baro (t i ) is the interpolation result of barometer at target time t i .

[0127] ​The UAV state information is serialized to form a sequence dataset, a sliding window length is set, and each time window sequence contains a set of target attribute information. The construction of the sliding window is completed, and all data after multi-sensor fusion is standardized according to the channel standard:

[0128] (1) Channel-specific IMU data normalization:

[0129] For each channel c∈{a x ,a y ,a z ,ω x ,ω y ,ω z ,p x ,p y ,p z}, the mean μ c and standard deviation σ c are calculated independently:

[0130]

[0131] where x is the measurement at the i-th timestamp, channel c. N is the number of timestamps over the entire training set.

[0132] Each channel is independently standardized:

[0133]

[0134] The normalized input matrix X is obtained:

[0135] The sensors carried by the UAV continuously monitor acceleration, angular velocity and position information during flight, and input these sensor data into the TCN-LSTM state estimation model in real time. If the state estimation result shows that the flight trajectory deviates or the attitude is abnormal, the control system will dynamically adjust the flight parameters such as thrust, attitude angle or yaw rate based on the estimated value, to realize real-time correction of the flight path and attitude. The state estimation and feedback control constitute a closed-loop mechanism, enabling the UAV to adapt and stabilize flight in complex environments.

[0136] The method makes the state estimation lag error of the unmanned aerial vehicle in a high-speed maneuvering scene decrease from 12% of pure LSTM to 3%, and the time series feature extraction efficiency increase by 40%. In a city low-altitude gust disturbance environment, the attitude angle prediction variance of the unmanned aerial vehicle is reduced by 35%, meeting the high dynamic control demand. In a GPS denial / EMI scene, the position estimation error fluctuation range of the unmanned aerial vehicle is reduced by 40% (compared with the model of the middle EKF); the multi-source noise (non-Gaussian distribution) adaptability is improved, and the positioning error is reduced by 30%-50% (actual measurement data). In a complex electromagnetic environment, the positioning accuracy is better than 10 m only by relying on IMU / barometer data. The attitude angle error is ≤1.5°, and the accuracy is improved by more than 30%; the position estimation RMSE is 0.48 m (0.82 m under traditional EKF gust disturbance), and the accuracy is improved by 41.5%. In agricultural plant protection precision spraying, the attitude stability of the unmanned aerial vehicle is improved by 25%, and the liquid coverage uniformity reaches 98%. In 200Hz closed-loop control of railway inspection, the trajectory tracking error of the unmanned aerial vehicle is reduced by 60%.

[0137] The method provided by the application can be applied to logistics, reconnaissance and disaster rescue fields on a large scale in power inspection, agricultural plant protection, railway inspection and other scenes, and promote high-precision control of unmanned aerial vehicle autonomous systems.

[0138] Although the embodiments of the application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the application.

Claims

1. A TCN-LSTM series network multi-source state real-time perception method for unmanned aerial vehicle control, characterized in that, The method comprises the following steps: (1) constructing a deep neural network model composed of TCN and LSTM in series; (2) collecting the speed, attitude and position information of the unmanned aerial vehicle through RTK-GPS, INS and IMU in cooperation, and aligning the IMU data and the RTK-GPS / INS data in time by using the interpolation method; (3) dividing the collected data into a training set, a validation set and a test set, and training and evaluating the model by using a composite loss function that fuses RMSE, MAE, quaternion angle error and bias error; (4) based on the sliding time window strategy, the sensor data sequence is formed into an input feature vector, and the IMU data is processed according to the channel standardization; (5) inputting the real-time collected standardized sensor data into the trained TCN-LSTM model to output the state estimation results of the current position, speed and attitude of the unmanned aerial vehicle for navigation, path planning or closed-loop control.

2. The method of claim 1, wherein, The TCN-LSTM series structure in step (1) comprises: The TCN part contains three layers of dilated causal convolution layers, the convolution kernel width is 3, the channel number is 64, 32 and 16 in turn, and the dilation coefficient is 1, 2 and 4 in turn; Each layer of TCN contains residual connection, weight normalization, ReLU activation function and dropout mechanism, and the dropout rate is 0.2; The feature sequence output by the TCN is input into a double-layer LSTM, the number of hidden units of each layer of the LSTM is 128 and 64 respectively, and the time dependence relationship is modeled using tanh and sigmoid activation functions; Finally, the state estimation results are generated through a fully connected layer.

3. The method of claim 1, wherein, The composite loss function in step (3) is defined as: F=λ1A1+λ2A2+λ3A3+λ4A4 where: λ1 = 1.0, λ2 = 1.0, λ3 = 5.0, λ4 = 0.1; is the model estimated accelerometer bias; is the true accelerometer bias; is the model estimated gyroscope bias; is the true gyroscope bias; is the estimated attitude quaternion; is the true attitude quaternion; is the estimated velocity vector of the drone at time t; is the true velocity of the drone; is the model evaluated position vector of the drone at time t; is the true position corresponding to time t; N is the number of time steps.

4. The method of claim 1, wherein, In step (2), the linear interpolation method is used to align the sampling time of GPS and INS data to the IMU timestamp.

5. The method of claim 1, wherein, In step (4), the step length of the sliding time window is 10 seconds, and each window contains a sequence of unmanned aerial vehicle state information.

6. The method of claim 1, wherein, The real-time state estimation in step (5) meets the deployment requirements of the edge computing platform, the single inference time is less than 5ms, and it supports a control frequency of more than 100Hz.

7. The method of claim 3, wherein, The quaternion angle error is defined as: wherein in the formula: θ t is the estimated pose rotated to the real pose the minimum rotation angle required about a rotation axis.

8. The method of claim 1, wherein, In step (3), the proportion of the training set is not less than 70%, and the validation set is used for hyperparameter tuning and early stopping strategy.

9. The method of claim 1, wherein, The state estimation results in step (5) include three-dimensional position coordinates, three-axis velocity vectors, and pitch angle / roll angle / yaw angle attitude angles.

10. A drone control system, characterized by, The state estimation results output by any one of claims 1-9 are used to generate navigation instructions or closed-loop control signals in real time.

Images