Unmanned aerial vehicle positioning and tracking risk identification method, storage medium and application

Abstract

The invention belongs to the technical field of automatic driving, and particularly relates to an unmanned aerial vehicle positioning and tracking risk identification method, a storage medium and application. State data in the flight process of the unmanned aerial vehicle are received in real time; after data preprocessing and standardization, recombining state data by adopting a sliding time window, and constructing a state tensor format of the unmanned aerial vehicle at the next moment; performing feature extraction by adopting a one-dimensional convolutional neural network in combination with a bidirectional long-short-term memory network; and feature integration and final state prediction. According to the method, aiming at the multi-dimensional motion state information of the unmanned aerial vehicle, a prediction relation is constructed by adopting a sliding time window sampling and sliding mode, and an efficient sequence space and time two-dimensional feature extraction mechanism is combined, so that the prediction accuracy and stability are improved, and the safety of an unmanned aerial vehicle control system and the real-time accuracy of decision making are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and specifically relates to a method for identifying risks of unmanned aerial vehicle positioning and tracking, a storage medium and an application thereof. Background Art

[0002] Existing UAV attitude recognition technologies mostly rely on inertial measurement units (IMUs) or visual sensors for real-time attitude estimation, but these traditional methods have relatively poor robustness and adaptability in complex environments, especially in high-dynamic flight or when there is a lot of environmental interference, which can easily lead to errors and affect flight safety. Therefore, how to accurately identify risks and respond in a timely manner based on the real-time data of the UAV's own attitude is a core technical issue in improving the safety of UAV flight.

[0003] At present, research on UAV attitude recognition mainly focuses on sensor data processing and algorithm optimization. Traditional attitude recognition methods are based on inertial measurement unit data and estimate the attitude of the UAV through sensors such as accelerometers and gyroscopes. However, these methods are easily affected by factors such as sensor noise, drift error, and magnetic field interference during flight, especially in fast flight and dynamically changing environments, and their accuracy is difficult to guarantee. Although the vision-based attitude estimation method can provide relatively accurate spatial positioning information, its performance is significantly reduced in low light, occlusion or complex background. Therefore, traditional methods often cannot fully meet the high-precision requirements for attitude recognition in complex environments.

[0004] With the rapid development of deep learning technology, gesture recognition algorithms based on artificial intelligence have gradually become an effective solution. In particular, the long short-term memory network LSTM, due to its advantages in processing time series data, can overcome the shortcomings of traditional methods in processing long-term dependent data in dynamic environments. The LSTM network can analyze the temporal patterns in the historical flight data of drones and identify the potential risks of gestures during flight. For example, LSTM can predict the possible gesture anomalies in the next few seconds by analyzing the gesture data of the drone at a certain moment and combining its historical flight status, so as to respond in advance. In addition, combined with multi-sensor data fusion technology, LSTM can effectively fuse data from different sensors such as IMU, visual sensors, GPS, etc., to further improve the accuracy and robustness of gesture risk identification. Although these methods have improved the accuracy of gesture recognition to a certain extent, there are still some problems. First, deep learning models usually require a large amount of labeled data for training, and the collection and labeling of these data are cumbersome and costly. Especially in special flight scenarios, the lack of labeled data may affect the training effect of the model. Secondly, the computational complexity of the LSTM model is high, which may cause processing delays for drone systems with high real-time requirements, affecting the timeliness of flight decisions. In addition, although multi-sensor fusion technology can reduce the error of a single sensor, the error characteristics and data format differences of different sensors may still lead to inconsistencies in the data fusion process, thereby affecting the accuracy and robustness of the recognition results. In short, the existing methods still face challenges such as large data requirements, poor real-time performance, and high difficulty in sensor fusion, and need further optimization and innovation to adapt to the requirements of complex flight environments. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a drone positioning and tracking risk identification method, a storage medium and an application.

[0006] The technical solution adopted by the present invention to solve the technical problem is as follows: a method for identifying risks of positioning and tracking unmanned aerial vehicles, the steps are as follows: S1. Data reception: real-time reception of the status data of the drone during flight; S2. Data preprocessing. After data preprocessing and standardization, the state data is reorganized using a sliding time window, and a state tensor format representing the state data of the drone at the next moment is constructed with a fixed-length window as the input data. S3, feature extraction, first use the convolution kernels from multiple one-dimensional convolutional neural networks to extract features of different spatial dimensions of the input data, then use the Bi-LSTM network to further extract time series features from the convolved data to capture the long-term dependency information in the state sequence, and combine the multi-head attention mechanism to splice the output to obtain enhanced feature representation; S4, feature integration and final state prediction, the final integration of features is performed through the long short-term memory network LSTM, and the state prediction value or risk classification result of the drone at the next moment is output.

[0007] Preferably, in step S1, the status data includes but is not limited to heading angle, roll angle, yaw rate, total speed, throttle opening and rudder angle.

[0008] Preferably, in step S2, the state quantity of the drone at the next moment is first represented by a nonlinear discrete mapping based on the multidimensional state data, and then the state data is normalized using the z-score standardization method; and then a sliding time window is used to slide from beginning to end to sequentially form new samples for processing.

[0009] Preferably, in step S2, according to the heading angle , yaw angular velocity , Roll Angle , total speed , throttle opening and rudder angle , the state of the drone at the next moment It is expressed as: ; in, express The state quantity at the moment; express The input amount at time, is its nonlinear discrete mapping; Considering that the UAV motion state has a periodic law, a sliding time window sampling is adopted, and the UAV state information prediction relation is expressed as: ; That is, drone model identification and prediction input: ; Where: , which is processed by sliding time window The state of the drone at the moment; It is a prediction model for drones, which can predict the state at the next moment based on the current state and control input; It is the state quantity after sliding time window processing; It is the input quantity after sliding time window processing.

[0010] Preferably, in step S3, multiple groups of one-dimensional convolutional neural networks are used to perform multi-dimensional feature extraction according to different dimensions of the state data.

[0011] Preferably, in step S3, the Bi-LSTM is composed of multiple LSTM units, each unit includes a forget gate, an input gate and an output gate to perform information screening and updating; the combination layer performs vector superposition on the output states calculated by the previous and next LSTM units.

[0012] Preferably, in the multi-head attention mechanism, scaled dot product attention is adopted, and the weight value obtained by the dot product operation of the query matrix and the key matrix is ​​normalized using the Softmax layer, and then the weighted sum is performed with the value matrix to obtain the attention result.

[0013] Preferably, the safety risk threshold is set using a baseline method, and the risk envelope distribution of each state variable is calculated and determined for risk classification.

[0014] A storage medium stores a program capable of executing the above-mentioned drone positioning and tracking risk identification method.

[0015] The application of the drone positioning and tracking risk identification method deploys the above-mentioned storage medium in a modular manner in an embedded target platform or in an external computing module.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention clearly proposes a UAV positioning and tracking risk identification method based on enhanced Bi-LSTM, which can effectively extract state time series features and has strong prediction accuracy and robustness. The method obtains the state data of the UAV in real time, determines the safe range of the UAV's motion state, sets the safety risk threshold by using the baseline method for the multi-dimensional motion state information of the UAV, clarifies the boundary of the risk area, and calculates and determines the risk envelope distribution of each state variable, so as to identify potential risks; 2. The prediction relationship constructed by sliding time window sampling and sliding can effectively capture the time series dynamic characteristics of drone status data and improve the accuracy and stability of prediction; 3. The method of combining one-dimensional convolutional neural network and enhanced Bi-LSTM network for feature extraction is used. The one-dimensional convolutional network is used to effectively extract the spatial features between different time series, and the enhanced Bi-LSTM captures the long-term time-dependent features of the sequence. The two work together to form an enhanced bidirectional long short-term memory network model IBLSTM with an efficient sequence space and time dual-dimensional feature extraction mechanism, which effectively reduces the risk of UAV autonomous driving systems in complex environments and improves the safety of UAV control systems and the real-time accuracy of decision-making; In summary, the risk identification method in this application can effectively capture the temporal dynamic characteristics of UAV status data, improve the accuracy and stability of prediction, effectively reduce the risk of UAV autonomous driving systems in complex environments, and improve the safety of UAV control systems and the real-time accuracy of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the application of the sliding time window method in the present invention; Figure 2 It is a schematic diagram of a one-dimensional convolutional neural network in the present invention; Figure 3 It is a schematic diagram of the LSTM unit structure in the present invention; Figure 4 It is a schematic diagram of the Bi-LSTM structure in the present invention; Figure 5 Schematic diagram of the multi-head self-attention mechanism in the present invention; Figure 6 It is a structural diagram of the IBLSTM model in the present invention; Figure 7 This is a bar chart of the heading angle prediction accuracy of the three models; Figure 8 This is a bar chart of the accuracy of yaw rate prediction for the three models; Fig. 9 It is a bar chart of the accuracy of roll angle prediction of the three models. DETAILED DESCRIPTION

[0018] In order to facilitate the understanding of the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0019] The UAV positioning and tracking risk identification method based on enhanced Bi-LSTM has the following steps: Step 1: UAV model identification and data preprocessing.

[0020] The status data of UAVs are collected through real flight experiments. After data preprocessing and standardization, the sliding time window technology is used to reorganize the status data to construct data samples for risk identification model training.

[0021] Data collection and standardization: Use real drones or computer simulations to conduct experiments under different environments and factors, obtain the drone's real state information, position information, and positioning tracking error data from the state data, and design different risk areas through different tracking point locations. The drone motion model state information includes: heading angle , yaw angular velocity , Roll Angle , total speed , Input: Throttle opening , rudder angle The multi-input and multi-output high-dimensional system, drone model recognition and prediction problem relationship: ; by express The state quantity at the moment; express The input amount at time, is a nonlinear discrete mapping.

[0022] First, a large amount of data on the status and location information of drones collected in real environments is preprocessed, and the status information is normalized using the z-score standardization method: ;in, is the original data set, is the mean of the original data set, is the standard deviation of the original data set, is the normalized value.

[0023] Sliding time window to reorganize data: Consider that the UAV motion state has a periodic law, that is, With obvious periodicity, considering the time correlation of state data, the following Figure 1 The sliding time window method shown processes the UAV motion state time series obtained through experiments.

[0024] The sliding time window technology is used to process the UAV state sequence data, that is, the original data is processed with a fixed length (such as ) slides forward from beginning to end at a certain step length, and each slide forms a new data sample. The data in each window is used to construct the input features, and the next state at the end of the window is used as the prediction target of the model. In this way, the time series dynamic characteristics of the drone state data can be effectively captured, and the accuracy and stability of the model prediction can be improved.

[0025] Model Identification and Prediction: After the sliding time window processing, the state sequence of the drone will be presented in a more time-series form. The drone state information model after the sliding time window algorithm can be expressed as: ; That is, the relationship between drone model identification and prediction problem is rewritten as: ; Where: , which is processed by sliding time window The state of the drone at the moment; It is a prediction model for drones, which can predict the state at the next moment based on the current state and control input; It is the state quantity after sliding time window processing; It is the input quantity after sliding time window processing.

[0026] Through this method, the model can not only capture the time dependence of the drone state data, but also effectively perform continuous single-step predictions, gradually approaching the real motion state of the drone. This can achieve accurate identification of the drone motion model and provide a basis for subsequent risk identification and control decisions.

[0027] Step 2: Feature extraction and optimization A one-dimensional convolutional neural network (1D-CNN) is used to extract spatial dimension features of the state sequence data obtained in step 1, effectively reducing the data dimension and optimizing the feature representation, thereby improving the calculation efficiency of subsequent models.

[0028] Considering that the spatial attributes of the state information of the UAV motion model have multi-dimensional characteristics, in order to improve the calculation speed, 1D-CNN is used for optimization to reorganize the features between the state information at different times. Figure 2 As shown in the figure, 1D-CNN performs convolution operations on the input data along the time axis by sliding the convolution kernel to extract local features between different time steps.

[0029] The convolution kernel (filter) slides along the input data on the time axis, and the dot product between each local area (time window) and the convolution kernel is calculated to obtain the convolution feature map. Each convolution kernel extracts a specific feature, such as the changing trend or local fluctuation of the drone state at a certain moment. By using multiple convolution kernels, feature information of different granularities can be extracted, which helps to describe the motion characteristics of the drone in different time periods.

[0030] The feature information processed by 1D-CNN will provide more compact and highly recognizable input data for subsequent models, which will help improve the training efficiency and prediction accuracy of subsequent models. Through this step, the features of the original drone state data are effectively extracted and optimized, providing clearer input data for capturing time-dependent features in subsequent steps, thereby enhancing the performance of the model.

[0031] Step 3: Time series feature extraction and Long Short-Term Memory (LSTM) application An enhanced bidirectional long short-term memory (Bidirectional Long Short-Term Memory, Bi-LSTM) network is used to further extract time series features from the convolved data and capture the long-term dependency information in the state sequence. Compared with the traditional recursive neural network (RNN), it can better capture the time correlation in the sequence and improve the accuracy of the model's prediction of state change trends.

[0032] LSTM (Long Short-Term Memory) is a special type of RNN. The LSTM unit structure is as follows Figure 3 As shown in Figure 1, it is often used to process and predict time series data. The core advantage of LSTM is that it can effectively capture long-term dependencies and avoid the gradient vanishing problem common in traditional RNNs. Each LSTM unit consists of the following key components: ForgetGate: Forget Gate Decide which information needs to be discarded from the cell state. As shown in the following formula represents the sigmoid activation function, is the forget gate weight matrix, is the forget gate bias matrix. It receives the hidden state of the previous moment and the current input , and generates an output between 0 and 1 through the sigmoid activation function. The output represents the proportion of information retained, 0 means completely forgotten, and 1 means completely retained: ; Input Gate: Input Gate Determines how the current information should be added to the cell state. It first generates an update signal through the sigmoid activation function to determine which values ​​will be updated; at the same time, the tanh activation function generates a candidate value that represents the impact of the current input on the cell state: ; Among them, tanh represents the hyperbolic tangent activation function, is the input gate weight matrix, is the input gate bias matrix.

[0033] Update cell state: Update the cell state through the combination of forget gate and input gate (multiplication unit). The cell state is the core of the LSTM unit. It stores important information about the current moment and the past moment and passes it in the sequence: ; ; in, is the cell state at the current moment, is the cell state at the previous moment, is the candidate value of the input gate, is the LSTM multiplication unit weight matrix, is the multiplication unit bias matrix.

[0034] Output Gate: Output Gate Determine the hidden state (i.e. output) of the next moment. It generates the hidden state based on the current cell state, which is ultimately used for calculation at the next moment: ; ; in, is the hidden state at the current moment, which serves as the input for the next moment and the final model output. is the weight matrix of the LSTM unit output gate, is the bias matrix of the output gate.

[0035] like Figure 4 As shown in the figure, Bi-LSTM is an improved LSTM network that can capture the bidirectional dependencies of time series more comprehensively by adding back propagation layers and combination layers on the basis of traditional LSTM. Specifically, Bi-LSTM extracts information from both the past (forward propagation) and the future (backward propagation) during the processing of time series. The input of each LSTM unit includes the state information of the current moment as well as the state information of the past and the future, so that the time series characteristics in the data can be captured more accurately. Bi-LSTM consists of multiple LSTM units, each of which contains structures such as forget gates, input gates, and output gates, which are responsible for deciding which information needs to be retained, which information needs to be forgotten, and how to update the memory unit.

[0036] The added combination layer calculates the output state of the previous and next LSTM units Perform vector superposition: ; in, Represents the output of the Bi-LSTM layer at time t; The output states calculated for the forward and backward LSTM units respectively.

[0037] Step 4: Multi-Head Self-Attention (MHSA) enhances feature extraction capabilities The features output by the Bi-LSTM network are introduced into the multi-head self-attention mechanism (MHSA) to further enhance the model's ability to understand long-term sequence information, reduce the loss of long-term sequence information, and improve the accuracy of feature extraction.

[0038] Overview of multi-head self-attention mechanism: As a variant of the attention mechanism, the multi-head attention mechanism (MHSA) uses multiple independent attention weights to obtain the correlation of the sequence, and captures the dependencies between different ranges through each independent head. Finally, the calculation results of these multiple attention heads are spliced ​​or weighted to form the final output. This mechanism enables the model to understand the structure of sequence data from multiple perspectives and improve the flexibility and robustness of information processing.

[0039] In step 3, Bi-LSTM has successfully captured the bidirectional temporal dependency of time series data, but there may still be some information loss or insufficient processing for modeling long-term dependency information. To this end, this step inputs the calculation results of Bi-LSTM into MHSA to reduce the correlation loss of the sequence in long-term transmission and improve the ability to extract features in the time dimension.

[0040] MHSA is applied on the sequence feature output after Bi-LSTM processing, calculating the attention score of each time step and weighting the input features according to these scores. Specifically, for each time step, MHSA calculates the similarity between this time step and other time steps, and dynamically adjusts the weight of the feature according to the calculation results.

[0041] MHSA combined with Bi-LSTM: like Figure 5 As shown, the scaled dot product attention (ScaledDot-ProductAttention) uses the Softmax layer to convert the query matrix ( ) and the key matrix ( ) dot product operation to obtain the weight value and normalize it, and then compare it with the value matrix ( ) is weighted and summed to obtain the attention result. The calculation process of MHSA usually includes the following steps: Query, key, and value calculation: First, the features output by Bi-LSTM Transformed into query (Query), key (Key), and value (Value) matrices through linear transformation (weight matrix).

[0042] in, ; ; ; .

[0043] Attention score calculation: Then, the dot product similarity between the query and the key is calculated and normalized by the softmax function to get the attention weight: ; in, It is the dimension of the key vector, which is used to scale the dot product result to avoid the value being too large.

[0044] Calculation and merging of multiple heads: Multiple independent attention heads are used to calculate their respective attention results in parallel. Finally, the outputs of all heads are concatenated to obtain an enhanced feature representation: The calculation result of Bi-LSTM is: , , , They are , , MHSA is calculated based on a single independent attention head, and the calculation results of n attention heads are concatenated to obtain the total attention result. , calculated as follows: ; ; in , , Respectively represent Attention Head , , The weight matrix, is the MHSA weight matrix.

[0045] Through the processing of the MHSA mechanism, the model can focus on the key time points in the sequence more accurately and capture time-dependent information of different ranges from different angles (multiple attention heads).

[0046] The feature representation after MHSA processing has stronger representation ability, which can effectively enhance the modeling ability of subsequent models for long-term dependent information, and provide more detailed and efficient feature input for drone state prediction. The introduction of the multi-head self-attention mechanism enables the model to better handle long-term dependencies when processing complex time series data, avoid information loss, and improve the ability to pay attention to information in different time periods. By combining MHSA with Bi-LSTM, the accuracy and robustness of drone state prediction can be significantly improved.

[0047] Step 5: Feature integration and final state prediction.

[0048] The LSTM network is used to integrate the features and output the predicted state value of the drone at the next moment. This step realizes the real-time prediction of the future state of the drone.

[0049] The LSTM with 128 hidden units is used to transfer features, and the calculation result of the last LSTM unit is output as 4-dimensional data as the risk prediction result. The structure of the enhanced Bi-LSTM model (IBLSTM model) proposed in the invention is as follows: Figure 6 As shown in Figure 1, it consists of an input layer, a CNN layer, a Bi-LSTM layer, a MHSA layer, and a decoding layer. The structure is as follows: Figure 6 shown.

[0050] Motion model recognition and prediction can be regarded as a sequence regression problem. At each sampling point, the state at the next moment is predicted based on the current state. In order to evaluate the prediction accuracy, the mean square error (MSE) is selected as the reference performance evaluation indicator, which is calculated as follows: ; The root mean square error (RMSE), which is more sensitive to data outliers, is selected as the core performance evaluation indicator, and its calculation is as follows: ; in, , They are respectively The predicted and actual values ​​of samples, is the sample size.

[0051] Step 6: Model training and cross-validation The EuRoCMAVDataset dataset was processed and different data segments were annotated by experts to ensure that the dataset is suitable for risk classification tasks. This annotation enables us to provide the training network with samples with clear labels so that the model can learn the characteristics of drone states under different risk levels. The annotated dataset was used to train the IBLSTM model and its generalization ability and prediction accuracy under different flight conditions were evaluated through cross-validation.

[0052] After completing the model structure design, in order to train the posture risk recognition neural network model based on the enhanced Bi-LSTM structure, EuRoCMAVDataset was selected as the training data source. EuRoCMAVDataset is a drone micro-flight dataset provided by the Swiss Federal Institute of Technology (ETHZurich). It contains high-frequency IMU data (acceleration, angular velocity), attitude truth, position data and image data collected in multiple typical indoor flight scenes. It has the characteristics of multi-modality, multi-scene, high time accuracy, etc., and is suitable for drone posture estimation, state prediction and risk identification tasks.

[0053] In order to adapt to the risk classification task, the training data selected the IMU data and attitude state quantities containing different flight trajectories and action types in the EuRoC dataset as feature input, including heading angle, yaw angular velocity, roll angle, three-axis acceleration, three-axis angular velocity and timestamp information. First, the original data was cleaned and synchronized, and the sliding time window algorithm was used to restructure the time series data to form input and output paired samples. Subsequently, the data was normalized using the z-score normalization method to improve the stability and generalization ability of model training.

[0054] Next, expert data labeling was performed. Experts labeled each data segment with the corresponding risk level based on the status changes in different time periods during the flight. The risk level is divided into six levels, increasing from level 1 (lowest risk) to level 6 (highest risk). This labeling process is done manually to ensure the accuracy and reliability of the data labels. A total of 400 sets of data were labeled, each of which contains the status characteristics of the drone in a specific time period and the corresponding risk level label.

[0055] After the dataset is labeled, each data segment not only contains the status information of the drone, but also corresponds to a clear risk level. This allows us to classify based on different risk levels when training the model, so that the network can learn the characteristics of different risk levels and ultimately improve the accuracy of risk identification.

[0056] The neural network model is trained in the Matlab Deep Learning Toolbox environment. The network structure includes an input layer, a one-dimensional convolution layer (for extracting state space features), a bidirectional LSTM layer (for capturing the time-dependent characteristics of state evolution), a multi-head self-attention mechanism layer (for enhancing the ability to model long-term dependencies), and an output decoding layer. The model output is the risk level of the current drone, and the training goal is to minimize the error between the predicted value and the actual state.

[0057] During the training process, the model performance was evaluated using indicators such as MSE (mean square error) and RMSE (root mean square error). The model parameters were updated through the Adam optimizer, and the dynamic monitoring chart of the training process was used to evaluate the model convergence. The hyperparameter settings were initial learning rate 0.001, mini-batch size 128, maximum learning batch 200, learning rate decay 0.15, and decay cycle 20.

[0058] The proposed IBLSTM model is cross-validated using multiple subsequences in the annotated EuRoCMAVDataset to verify its generalization ability under different flight conditions. The MSE and RMSE indicators are used to compare and analyze the IBLSTM, the LSTM model commonly used in prediction and classification tasks, and the support vector machine (SVM) regression model. , yaw angular velocity and roll angle The risk classification results of these three posture data are shown in Table 1. The prediction accuracy of the three models under different types of data is shown in Table 1. Figure 7-9 The results show that the IBLSTM model can effectively extract state time series features compared with other networks, has strong prediction and classification accuracy and robustness, and provides an algorithmic basis for subsequent system deployment and risk assessment.

[0059] The results of cross-validation further prove that the IBLSTM model has good generalization ability under different flight conditions and can effectively process time series data in various flight environments.

[0060] Table 1 Comparison of prediction performance of different models in EuRoCMAVDataset

[0061] Step 7: Model deployment and application This paper introduces the practical deployment method of the IBLSTM model, including simulation verification in the Simulink environment and the actual application solution on the plug-in computing module, to provide real-time risk identification capabilities for the UAV autonomous driving system.

[0062] Model deployment and usage plan, after completing the IBLSTM model training based on one-dimensional convolutional neural network, bidirectional long short-term memory network (Bi-LSTM) and multi-head attention mechanism (MHSA), the model can be applied to the system integration of posture risk recognition during autonomous flight of drones in the following two ways: One way is to import the trained deep neural network model into the Simulink environment. The trained network structure is introduced into the Simulink model in a supported format (such as DAGNetwork or dlnetwork) through Matlab Deep Learning Toolbox, and the model is encapsulated and simulated by combining the sliding time window construction module, data preprocessing module and prediction output module in Simulink. Furthermore, by using Simulink Coder to generate C / C++ code, the model can be deployed on an embedded target platform or in software-in-the-loop (SiL) testing, and can be used in conjunction with the PX4 flight control system for joint simulation verification or actual operation.

[0063] Another way is to deploy the IBLSTM model to an external computing module. The external computing module is an independent embedded intelligent processing device with the ability to run an operating system and deep learning reasoning, such as JetsonNano, RaspberryPi or an industrial-grade edge computing terminal. The model deployment process includes the following steps: Use the exportONNXNetwork function in Matlab or the Python-side conversion tool to export the trained IBLSTM model to ONNX or other common formats; deploy an adaptive reasoning framework (such as PyTorch) in the external processing device and load the model file; build a data acquisition interface, and the external module establishes a connection with the PX4 flight control system through a communication protocol such as a serial port, UDP or MAVROS to receive real-time status information of the drone during flight, including heading angle, roll angle, yaw angular velocity, total speed, throttle opening, rudder angle, etc.; the received status The information enters the sliding time window cache module and is constructed into an input tensor format consistent with the model training stage; the preprocessed state data is input into the IBLSTM model for real-time reasoning, and the next moment state prediction value or risk classification result is output; the recognition result is fed back to the PX4 flight control system through the return link (serial port, UDP data packet or ROS topic) for strategy correction of the mission controller or attitude controller; according to actual application requirements, the risk level threshold or anomaly detection trigger mechanism can be set. When the output result exceeds the set risk limit, the PX4 can be linked to execute control logic such as attitude stabilization, mission interruption, and return.

[0064] Among them, the baseline method is used to collect a large amount of state information and location information of drones in real environments through experiments, and statistical analysis is used to determine the normal fluctuation range of each state variable, and this is used as the baseline to form a risk envelope, clarify the boundaries of the risk area, and identify potential risks. The risk envelope is specifically manifested in setting upper and lower limit thresholds for each motion state variable. When the state data detected in real time exceeds the threshold range, it is determined that there is an abnormal risk, thereby realizing the accurate extraction and clustering of risk information of drone state data.

[0065] As an auxiliary risk identification unit of the main flight control system, the plug-in module does not change the original controller structure inside PX4, but only provides additional risk perception and information feedback functions. It has the characteristics of structural decoupling, flexible deployment, and strong processing capability. It is suitable for actual scenarios where resource-constrained platforms cannot run complex neural network models.

[0066] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for identifying risks of drone positioning and tracking, characterized in that: Here are the steps: S1. Data reception: real-time reception of the status data of the drone during flight; S2. Data preprocessing. After data preprocessing and standardization, the state data is reorganized using a sliding time window, and a state tensor format representing the state data of the drone at the next moment is constructed with a fixed-length window as the input data. S3, feature extraction, first use the convolution kernels from multiple one-dimensional convolutional neural networks to extract features of different spatial dimensions of the input data, then use the Bi-LSTM network to further extract time series features from the convolved data to capture the long-term dependency information in the state sequence, and combine the multi-head attention mechanism to splice the output to obtain enhanced feature representation; S4, feature integration and final state prediction, the final integration of features is performed through the long short-term memory network LSTM, and the state prediction value or risk classification result of the drone at the next moment is output.

2. The method for identifying risk of UAV positioning and tracking according to claim 1, characterized in that: In step S1, the status data includes but is not limited to heading angle, roll angle, yaw rate, total speed, throttle opening and rudder angle.

3. The method for identifying risk of UAV positioning and tracking according to claim 2, characterized in that: In step S2, the state of the drone at the next moment is first represented by a nonlinear discrete mapping based on the multidimensional state data, and then the z-score standardization method is used to normalize the state data; then a sliding time window is used to slide from the beginning to the end to form new samples in sequence.

4. The method for identifying risk of UAV positioning and tracking according to claim 3, characterized in that: In step S2, According to the heading angle , yaw angular velocity , Roll Angle , total speed , throttle opening and rudder angle , the state of the drone at the next moment It is expressed as: ； in, express The state quantity at the moment; express The input amount at time, is its nonlinear discrete mapping; Considering that the UAV motion state has a periodic law, a sliding time window sampling is adopted, and the UAV state information prediction relation is expressed as: ； That is, drone model identification and prediction input: ; Where: , which is processed by sliding time window The state of the drone at the moment; It is a prediction model for drones, which can predict the state at the next moment based on the current state and control input; It is the state quantity after sliding time window processing; It is the input quantity after sliding time window processing.

5. The method for identifying risk of UAV positioning and tracking according to claim 4, characterized in that: In step S3, multiple groups of one-dimensional convolutional neural networks are used to perform multi-dimensional feature extraction according to different dimensions of the state data.

6. The method for identifying risk of UAV positioning and tracking according to claim 5, characterized in that: In step S3, the Bi-LSTM is composed of multiple LSTM units, each of which contains a forget gate, an input gate, and an output gate to perform information screening and updating; the combination layer vectorizes the output states calculated by the previous and next LSTM units.

7. The method for identifying risk of UAV positioning and tracking according to claim 6, characterized in that: In the multi-head attention mechanism, scaled dot product attention is adopted. The Softmax layer is used to perform dot product operation on the query matrix and the key matrix to obtain the weight value, which is then normalized and weightedly summed with the value matrix to obtain the attention result.

8. The method for identifying risk of UAV positioning and tracking according to claim 7, characterized in that: The baseline method is used to set the safety risk threshold, and the risk envelope distribution of each state variable is calculated and determined for risk classification.

9. A storage medium, characterized in that A program capable of executing the drone positioning and tracking risk identification method as described in any one of claims 1-8 is stored.

10. Application of the risk identification method for drone positioning and tracking, characterized in that: The storage medium as described in claim 9 is modularly deployed in an embedded target platform or in an external computing module.

Images