Unmanned aerial vehicle security situation awareness method based on multi-source heterogeneous data fusion

Abstract

The invention discloses an unmanned aerial vehicle security situation awareness method based on multi-source heterogeneous data fusion, and belongs to the technical field of unmanned aerial vehicles, and the method comprises the steps: obtaining the multi-source heterogeneous data of a target unmanned aerial vehicle in real time, and carrying out the preprocessing of the multi-source heterogeneous data; performing feature extraction on the preprocessed multi-source heterogeneous data according to data types to generate multi-source heterogeneous features; performing conversion processing on the multi-source heterogeneous features through a heterogeneous feature self-combination algorithm to generate fusion features, and introducing CLS nodes to generate single time step diagram structure data; cLS node vector representation is extracted from the single time step graph structure data through a graph neural network; paying attention to key data through an attention mechanism for CLS node vector representation to generate a comprehensive feature vector; and performing security situation awareness classification on the comprehensive feature vector through a CatBoost algorithm to generate a security situation awareness result. The dynamic change of the flight environment can be responded in real time, and the flight safety and the situation prediction capability of the unmanned aerial vehicle are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and particularly to a method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion. Background Art

[0002] With the wide application of unmanned aerial vehicles in civilian and commercial fields, especially in the low-altitude flight field, the flight safety problem of unmanned aerial vehicles has gradually become the focus. Currently, the safety situation awareness of unmanned aerial vehicles faces many challenges: First, there are various complex interference factors in the low-altitude environment, such as meteorological changes, electromagnetic interference, etc., which makes it difficult to predict the flight trajectory and state changes of unmanned aerial vehicles; Second, the acquisition and processing of multi-source heterogeneous data in the low-altitude intelligent network is highly complex, including flight data, environmental data from different sensors, and communication information with other unmanned aerial vehicles or ground control stations. The fusion and analysis of these data are difficult; Finally, the current safety situation awareness technology has certain delays and misjudgment problems when dealing with the high-speed dynamic changes of unmanned aerial vehicle trajectories and behaviors, and it is difficult to provide accurate threat predictions and alarms in real time.

[0003] In existing methods, most unmanned aerial vehicle safety situation awareness systems focus on single data sources or low-dimensional data processing, and it is difficult to cope with the challenges brought by complex and changeable low-altitude environments and multi-source heterogeneous data, resulting in weak perception capabilities of the system for the flight state and potential risks of unmanned aerial vehicles and being unable to comprehensively grasp the safety situation. In addition, traditional situation awareness methods are difficult to adapt to dynamic flight environments, especially during low-altitude flights, and cannot be updated and adjusted in real time, resulting in inaccurate risk predictions and response strategies. At the same time, most existing safety situation awareness and early warning mechanisms rely on static analysis and do not fully utilize real-time data for dynamic monitoring. Especially when multiple unmanned aerial vehicles fly in coordination, it is difficult to accurately identify potential safety threats such as collision risks or trajectory anomalies. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion, so as to solve the technical problems such as single data source, low data processing efficiency, and incomplete situation awareness existing in the existing unmanned aerial vehicle safety perception methods.

[0005] To achieve the above object, the present invention is implemented by the following technical solutions:

[0006] The present invention provides a method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion, including:

[0007] Real-time acquisition of multi-source heterogeneous data of the target unmanned aerial vehicle and preprocessing of the multi-source heterogeneous data;

[0008] Extract features from the preprocessed multi-source heterogeneous data according to the data type to generate multi-source heterogeneous features;

[0009] Generate fused features by performing transformation processing on the multi-source heterogeneous features through a heterogeneous feature self-combination algorithm;

[0010] Introduce a CLS node into the fused features to construct an adjacency matrix, and generate single-time-step graph structure data;

[0011] Extract the CLS node vector representation from the single-time-step graph structure data through a graph neural network;

[0012] Generate a comprehensive feature vector by focusing on key data for the CLS node vector representation through an attention mechanism;

[0013] Perform security situation awareness classification on the comprehensive feature vector through the CatBoost algorithm to generate security situation awareness results.

[0014] Optionally, the multi-source heterogeneous data includes: ADS-B data, remote identification data, alarm configuration data and alarm information data, task-related data, and device basic data.

[0015] Optionally, the preprocessing of the multi-source heterogeneous data includes: missing value filling, outlier detection and elimination, noise removal, normalization, standardization, and time serialization.

[0016] Optionally, the feature extraction from the preprocessed multi-source heterogeneous data according to the data type includes: extracting time-domain features from time series data; extracting spatial features from position data and motion data; extracting image features from image data and video data.

[0017] Optionally, the transformation processing of the multi-source heterogeneous features through the heterogeneous feature self-combination algorithm includes:

[0018] Denote the multi-source heterogeneous features at the th time step as , is the number of features, is the th feature in ;

[0019] Initialize a learnable vector representation for the multi-source heterogeneous features at each time step, is the dimension of the vector representation, is the th feature in ;

[0020] Successively calculate the vector representation Possibility score matrix between features , the possibility score matrix The elements in are as follows:

[0021]

[0022] Wherein, , is the possibility score calculation function;

[0023] Calculate the association matrix according to the possibility score matrix :

[0024]

[0025] Wherein, is the element activation function, is the learnable inductive bias, is the constant threshold;

[0026] According to the association matrix , perform heterogeneous feature combination on multi-source heterogeneous features through matrix multiplication:

[0027]

[0028]

[0029] Wherein, is the concatenation operation, is the selection function, is The th feature in is the after heterogeneous feature combination;

[0030] Map the features after heterogeneous feature combination to a dense and uniform -dimensional latent space to obtain the fused features at the th time step.

[0031] Optionally, introducing a CLS node to the fused features to construct an adjacency matrix and generating single-time-step graph structure data includes:

[0032] Initialize the CLS node to -dimensional space vector , define the subscript set of the nodes of the space vector as , and the subscript set of the nodes of the fused features is ;

[0033] Bi - directionally connect the CLS node with each node of the fused features to generate an adjacency matrix , the adjacency matrix where the element is:

[0034]

[0035] Based on the spatial vector and the adjacency matrix construct the single - time - step graph - structured data .

[0036] Optionally, the extraction of the CLS node vector representation from the single - time - step graph - structured data by the graph neural network includes:

[0037] Initialize the mean of the vector representations of the two nodes connected by each edge in the single - time - step graph - structured data as the vector representation of each edge through the graph neural network;

[0038] Concatenate the vector representations of each edge and the two nodes it connects, and concatenate the vector representations of each node and the edges connected to it. The concatenation result completes message passing and aggregation through a multi - layer perceptron:

[0039]

[0040]

[0041] In the formula, is the vector representation of the th node, is the vector representation of the edge between the th and th nodes, is the neighbor node of the th node,

[0042] Update the vector representations of the nodes and edges:

[0043]

[0044]

[0045] After multiple iterative updates, obtain the CLS node vector representation.

[0046] Optionally, the generation of the comprehensive feature vector by focusing on key data for the CLS node vector representation through the attention mechanism includes:

[0047] At each time step, first calculate the query value and the key value The similarity, and the score is obtained through dot product :

[0048]

[0049] For the score Normalize it through the Softmax function to obtain the attention weights at each time step :

[0050]

[0051] In the formula, is the th time step;

[0052] For the CLS node vector representation at each time step Perform weighted aggregation to obtain the comprehensive feature vector :

[0053]

[0054] In the formula, is the th attention weight and CLS node vector representation at the time step.

[0055] Optionally, the generating the security situation awareness result by classifying the comprehensive feature vector through the CatBoost algorithm includes:

[0056] Construct and initialize the CatBoost model, and use the comprehensive feature vector as the training sample;

[0057] Calculate the gradient of the CatBoost model for each training sample, and update the CatBoost model using the gradient. The gradient is:

[0058]

[0059] In the formula, is the loss function, is the th training sample 's target value, is round of the CatBoost model's prediction for the training sample , is the step size;

[0060] In each round of training, construct rounds of decision trees to minimize the loss function and correct the error of the previous round;

[0061]

[0062] In the formula, is the desired function, the output of the newly trained decision tree;

[0063] According to the round of decision tree obtain the prediction of the round of CatBoost model for the training samples : :

[0064]

[0065] After multiple rounds of training, the outputs of all decision trees are used to obtain the final security situation awareness result through weighted summation or voting.

[0066] Optionally, the security situation awareness result includes collision risk, trajectory anomaly, communication interruption, system failure, environmental interference, and normal operation.

[0067] Compared with the prior art, the beneficial effects achieved by the present invention:

[0068] The present invention provides a method for unmanned aerial vehicle (UAV) security situation awareness based on multi-source heterogeneous data fusion, which collects multi-source heterogeneous data and preprocesses it to improve data quality. By designing a heterogeneous feature self-combination algorithm, the data is converted into single-time-step graph-structured data, and spatial and temporal features are extracted by combining a graph neural network and an attention mechanism to generate a comprehensive feature vector. The CatBoost algorithm is used to classify the flight state and potential risks of the UAV, realizing accurate and efficient security situation awareness. It solves the problems of single data source, low processing efficiency, and incomplete situation recognition in current UAV security situation awareness. In summary, the present invention can respond in real time to the dynamic changes in the flight environment, enhance the flight safety and situation prediction ability of the UAV, and is applicable to the security monitoring and early warning of UAVs in complex low-altitude environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a schematic flowchart of a method for UAV security situation awareness based on multi-source heterogeneous data provided by an embodiment of the present invention;

[0070] Figure 2 is a schematic structural diagram of a UAV security situation awareness system based on multi-source heterogeneous data provided by an embodiment of the present invention;

[0071] Figure 3 is a multi-source heterogeneous data fusion structure diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and cannot be used to limit the protection scope of the present invention.

[0073] Embodiment 1:

[0074] As Figure 1 shown, a method for unmanned aerial vehicle (UAV) safety situation awareness based on multi-source heterogeneous data fusion provided by an embodiment of the present invention includes the following steps:

[0075] Step S1: Real-time obtain multi-source heterogeneous data of the target UAV and preprocess the multi-source heterogeneous data.

[0076] Specifically, in this embodiment, the multi-source heterogeneous data includes but is not limited to: ADS-B data, remote identification data, alarm configuration data, alarm information data, task-related data, and device basic data.

[0077] The ADS-B data is derived from the automatic dependent surveillance-broadcast system and can transmit key information such as the position, heading, speed, and altitude of the UAV in real time. The specific fields include UAV number, task ID, mode, flight number, flight speed, heading, position (latitude and longitude), altitude, signal strength, etc. The system can monitor the UAV status in real time through these data and can detect and handle potential flight anomalies in a timely manner.

[0078] The remote identification data (REMOTE_ID_DATA) is obtained from the Remote ID system that transmits the identity and position information of the UAV, and can provide detailed operation status and environmental interaction information of the UAV to ensure its safety during flight. The specific fields include UAV number, task ID, sensor ID, received signal strength, timestamp, communication channel, data type (such as Bluetooth or WIFI), status, direction, speed, position, etc.

[0079] The alarm configuration data (ALARM_CONFIG) is used to define the alarm rules and policies in the monitoring system to ensure that the system can generate alarm notifications in a timely manner when specific conditions or thresholds are triggered. The specific fields include alarm name, monitoring metrics, trigger conditions, alarm level, alarm information content, and the configured enabled status.

[0080] The alarm information data (ALARM_INFO) records various alarm events generated by the system during the monitoring process. The specific fields include UAV number, alarm name, monitoring metrics, trigger conditions, alarm level, alarm information, etc. Through these data, the system can trace historical alarm records, analyze alarm patterns, optimize alarm policies, and improve the intelligence and response speed of the system.

[0081] Task-related data (UAV_DATA) contains detailed flight data related to the UAV's tasks. The specific fields include UAV number, voltage, current, remaining battery level, positioning data, speed, flight angle, etc. The system can analyze this data to adjust flight parameters in a timely manner to ensure the successful completion of the task.

[0082] Device basic data (UNIT_INFO) stores detailed information about the UAV and its related devices. The specific fields include UAV number, device type, model, size, operating system, manufacturer, network support, weight, etc.

[0083] The above data sources provide the real-time status, flight trajectory, operation status, environmental interaction, and alarm information of the UAV, ensuring that the system can comprehensively and accurately grasp the operation of the UAV.

[0084] Specifically, in this embodiment, the preprocessing of multi-source heterogeneous data includes, but is not limited to: missing value imputation, outlier detection and removal, noise removal, normalization, standardization, and time series.

[0085] Missing value imputation: In UAV data analysis, especially in the case of multi-sensor integration, missing values are inevitable. To effectively handle these missing values, for time series data, linear interpolation is used to fill in the missing values, which can smoothly restore the trend of the missing part. At the same time, for samples or sensor data with severe missing values, the deletion method is used to remove these invalid data from the dataset to prevent them from having an adverse impact on subsequent analysis and modeling.

[0086] Outlier detection and removal: In UAV sensor data, outliers may be caused by factors such as sensor failures, environmental interference, or data recording errors. To improve data quality and avoid the negative impact of outliers on model training, statistical methods are used for outlier detection. Specifically, the Z-score method is used to identify data points that deviate significantly from the mean, or the interquartile range (IQR) method is used to detect outliers outside the data distribution range. These methods can effectively remove unreasonable data and ensure the accuracy of subsequent analysis and modeling.

[0087] Noise removal: Sensor data is often disturbed by environmental noise. For example, the readings of temperature and humidity sensors may be affected by environmental factors such as wind speed. To effectively reduce noise, a low-pass filter and a moving average method are added. By using a low-pass filter, high-frequency noise in the signal is removed to smooth data fluctuations; the moving average method is used to calculate the average value within the window to reduce random fluctuations in the data, thereby improving the stability and accuracy of the data.

[0088] Normalization: Compress the data into a unified scale range. Min-max normalization is adopted, which maps the data between 0 and 1 to eliminate the dimensional differences between features and convert different features into the same scale. For UAV sensor data such as air pressure and temperature, normalization can avoid the imbalance problem caused by different dimensions between different sensors and keep the contribution of each feature to the model balanced.

[0089] Standardization: Convert the data into a distribution with a mean of 0 and a standard deviation of 1, so that each feature has the same scale. The standardization process subtracts the mean and divides by the standard deviation to ensure that features with different numerical ranges have the same impact on the model. For numerical features such as the speed and acceleration of UAVs, standardization can eliminate the interference of large-value features, thereby improving the stability and accuracy of the algorithm.

[0090] The goals of data normalization and standardization are to compress the data into a unified scale range to more accurately extract features from various collected data.

[0091] Time series: Perform time series processing on UAV data to ensure that the data collected at the same moment is consistent in the time series. It mainly consists of three parts: time series alignment and interpolation, time series sliding window, and time series decomposition.

[0092] Time series alignment and interpolation: Since the data collected by different sensors usually have different timestamps, it is necessary to align them to a unified time point when processing time series data to prevent information loss caused by inconsistent timestamps.

[0093] Time series sliding window: Divide the data set into time windows of a fixed size (such as every 5 seconds or 10 data points) to analyze the short-term changes in the data. As the window moves forward, non-overlapping or overlapping subsequences are gradually extracted to help capture the short-term dependencies in the time series and improve the prediction ability of the model.

[0094] Time series decomposition: By decomposing the original time series data into multiple components, it helps to better understand the internal structure of the data, extract the long-term changes and seasonal periodic fluctuations in UAV data, provide valuable insights for subsequent modeling, and help identify flight patterns and behaviors.

[0095] Step S2: Extract features from the preprocessed multi-source heterogeneous data according to the data type to generate multi-source heterogeneous features.

[0096] Specifically, in this embodiment, the multi-source heterogeneous features include but are not limited to time domain features, spatial features, and image features.

[0097] The time-domain features are statistics extracted from the original time series data of the drone, including mean, maximum value, minimum value, standard deviation, skewness, kurtosis, etc. For the flight data of the drone, the time-domain features can provide key information about the flight state, speed change, etc. For example, the standard deviation of speed can reflect the flight stability, and the maximum acceleration can reveal the intensity of the flight.

[0098] The spatial features are features extracted from the position information and motion data of the drone. For example, by analyzing the coverage area of the flight path, path complexity, and uniformity of position distribution, the flight pattern and efficiency can be revealed, and similar flight patterns can be identified by analyzing the spatial features.

[0099] The image features are an important part of the data collected by the drone, especially playing a key role in tasks such as environmental monitoring and target detection. The convolutional neural network (CNN) technology is used to extract high-level features from image data or video data, providing reliable and efficient feature support for tasks such as drone target recognition, flight trajectory, and safety situation awareness.

[0100] Step S3: Use the heterogeneous feature self-combination algorithm to perform transformation processing on the multi-source heterogeneous features to generate fused features.

[0101] Automatically combine the multi-source heterogeneous features, mine the deep connections between the features, and eliminate the heterogeneity.

[0102] Specifically, in this embodiment, the transformation processing of the multi-source heterogeneous features by the heterogeneous feature self-combination algorithm includes:

[0103] Step S3.1: Denote the multi-source heterogeneous features at the th time step as , is the number of features, is the th feature in

[0104] Step S3.2: Initialize a learnable vector representation for the multi-source heterogeneous features at each time step, is the dimension of the vector representation, is the th feature in

[0105] Step S3.3: Calculate the possibility score matrix between the features in the vector representation in turn. The element in the possibility score matrix is:

[0106]

[0107] In the formula, , is the possibility score calculation function.

[0108] Step S3.4: Calculate the correlation matrix according to the possibility score matrix :

[0109]

[0110] In the formula, is the element activation function, such as the sigmoid function, is the learnable inductive bias, is the constant threshold for controlling the number of relevant features;

[0111] Step S3.5: Perform heterogeneous feature combination on the multi-source heterogeneous features through matrix multiplication according to the correlation matrix :

[0112]

[0113]

[0114] In the formula, is the concatenation operation, is the selection function, is the th feature in is the after heterogeneous feature combination;

[0115] Step S3.6: Map the features after heterogeneous feature combination to a dense and uniform -dimensional latent space through the feature mapper of the MLP structure to obtain the fused features at the th time step.

[0116] Step S4: Introduce a CLS node into the fused features to construct an adjacency matrix and generate single-time-step graph structure data.

[0117] After having the node vectors, it is also necessary to truly convert the data into graph structure data. Here, the adjacency matrix of the graph is defined.

[0118] Specifically, in this embodiment, introducing a CLS node into the fused features to construct an adjacency matrix and generate single-time-step graph structure data includes:

[0119] Initialize the CLS node to a -dimensional space vector , and to distinguish the CLS node from ordinary feature nodes, define the space vector The subscript set of the nodes is , and the subscript set of the nodes with fused features is ;

[0120] Since the CLS node is a global node and other normal nodes are local feature nodes, in order to better interact local feature information into the global feature, a bidirectional connection is made between the CLS node and each node with fused features to generate an adjacency matrix , and the element in the adjacency matrix is:

[0121]

[0122] In addition to the connection between the CLS node and the ordinary feature nodes, the connection between the ordinary feature nodes also needs to be considered. Since the feature self-combination algorithm has considered the correlation information between features, a minimum connected subgraph is used to construct the connection between the ordinary feature nodes to avoid excessive information redundancy and resource consumption. Based on the spatial vector and the adjacency matrix to construct the single time-step graph structure data .

[0123] Step S5: Extract the vector representation of the CLS node from the single time-step graph structure data through a graph neural network.

[0124] Specifically, in this embodiment, extracting the vector representation of the CLS node from the single time-step graph structure data through a graph neural network includes:

[0125] Initializing the mean value of the vector representations of the two nodes connected by each edge in the single time-step graph structure data as the vector representation of each edge through a graph neural network;

[0126] Concatenating the vector representations of each edge and the two nodes connected to it, and concatenating the vector of each point and the edge connected to it. The concatenation result completes message passing and aggregation through a multi-layer perceptron:

[0127]

[0128]

[0129] In the formula, is the vector representation of the th node, is the vector representation of the edge between the th nodes, is the neighbor node of the th node, is the multi-layer perceptron;

[0130] Update the vector representations of nodes and edges:

[0131]

[0132]

[0133] After multiple iterative updates, the vector representation of the CLS node is obtained.

[0134] Step S6: Generate a comprehensive feature vector by focusing on key data for the vector representation of the CLS node through the attention mechanism.

[0135] In the multi-source data fusion model of the unmanned aerial vehicle, introducing the attention mechanism enables the model to automatically focus on the most critical data and capture the context time attributes to improve the prediction performance.

[0136] Specifically, in this embodiment, generating a comprehensive feature vector by focusing on key data for the vector representation of the CLS node through the attention mechanism includes:

[0137] At each time step, first calculate the query value and the similarity with the key value to obtain the score through the dot product:

[0138]

[0139] Normalize the score through the Softmax function to obtain the attention weight at each time step:

[0140]

[0141] where is the th time step;

[0142] Perform weighted aggregation on the vector representation of the CLS node at each time step to obtain the comprehensive feature vector :

[0143]

[0144] where is the attention weight and the vector representation of the CLS node at the th time step.

[0145] With such attention weights, it helps the model to automatically fuse the effective feature information of the current time step and the previous time step when processing multi-source data of the drone, forming a comprehensive feature vector that enhances the model's expressive ability, thereby improving the risk identification accuracy and robustness of the downstream prediction model.

[0146] Step S7: Perform security situation awareness classification on the comprehensive feature vector through the CatBoost algorithm to generate security situation awareness results.

[0147] Specifically, in this embodiment, the security situation awareness results include collision risk, trajectory anomaly, communication interruption, system failure, environmental interference, and normal operation.

[0148] Collision risk: By monitoring the distance, relative speed, and flight trajectory between drones, judge whether there is a collision risk.

[0149] Trajectory anomaly: Analyze the flight trajectory of the drone to identify abnormal behaviors such as deviation from the predetermined path, sudden acceleration or deceleration.

[0150] Communication interruption: Continuously track the communication connection between the drone and the ground station or other drones to identify whether there is a communication interruption or signal weakening problem.

[0151] System failure: Monitor the power system status of the drone through sensor data to identify problems such as engine failure and insufficient battery power.

[0152] Environmental interference: Analyze the overall change trend of the drone to predict the changes in the surrounding environment, and judge whether these factors affect the normal flight of the drone.

[0153] To achieve the classification of these risks, the embodiment of the present invention uses the CatBoost algorithm to classify the comprehensive feature vector. CatBoost is a machine learning method based on gradient boosting, which specifically optimizes the processing of categorical features and can efficiently perform classification tasks. The algorithm iteratively optimizes the model by constructing a series of decision trees, gradually reducing the error, and finally improving the classification accuracy.

[0154] The specific process includes:

[0155] Construct and initialize the CatBoost model, and use the comprehensive feature vector as the training sample;

[0156] Calculate the gradient of the CatBoost model for each training sample, and update the CatBoost model using the gradient. The gradient is:

[0157]

[0158]

[0159] In the formula, is the loss function, is the th training sample target value, is round of the CatBoost model's prediction for the training sample ; is the step size, and the optimal value can be selected through cross-validation.

[0160] In each round of training, construct rounds of decision trees to minimize the loss function and correct the error of the previous round;

[0161]

[0162] In the formula, is the expectation function, the output of the newly trained decision tree;

[0163] After that, perform multiple iterations. In each round of iteration, the new decision tree corrects the error of the previous round, gradually improves the classification performance, continuously updates the model, and obtains the final CatBoost model :

[0164]

[0165] In the formula, is the prediction of the decision tree for the category , is the input comprehensive feature vector.

[0166] The CatBoost algorithm can directly process categorical features without additional preprocessing, which makes its application in complex drone data very efficient.

[0167] The CatBoost algorithm has strong robustness. Even in the face of noisy or partially missing data, CatBoost can maintain high stability and accuracy.

[0168] The CatBoost algorithm supports parallel computing and can accelerate the training and prediction processes in a multi-core computing environment.

[0169] Through adaptive techniques, CatBoost can effectively reduce the risk of overfitting during training and ensure the generalization ability of the model.

[0170] As the number of decision trees increases, the CatBoost model can significantly improve classification accuracy and effectively reduce the risk of overfitting. Each decision tree is trained based on different data and features, so it has strong generalization ability.

[0171] As the diversity and complexity of data sources increase, CatBoost can effectively extract valid information from multi-dimensional features and adapt to different situation prediction scenarios.

[0172] In summary, the method for unmanned aerial vehicle (UAV) safety situation awareness based on multi-source heterogeneous data fusion provided by the embodiments of the present invention includes: a) Situation element awareness: The data acquisition layer collects data in real time from multiple data sources, providing real-time status, flight trajectory, operation status, environmental interaction and warning information of the UAV, ensuring that the system can comprehensively monitor the operation of the UAV. b) Situation understanding: The feature extraction layer realizes spatio-temporal feature extraction, enabling the system to analyze multi-dimensional information of the UAV in terms of flight trajectory, speed change and environmental status. Introducing the Attention mechanism can make the model focus on key data, so as to automatically focus on important features affecting flight safety during the data analysis process. Finally, the system integrates these analysis results to form a comprehensive understanding of the current situation, providing a reliable basis for further situation prediction and risk assessment. c) Situation prediction: The present invention uses the CatBoost algorithm to classify the risk of the comprehensive feature vector, distinguishing the normal and abnormal states of the UAV and its cluster. The solution provided by the embodiments of the present invention can respond to the dynamic changes of the flight environment in real time, enhance the flight safety and situation prediction ability of the UAV, and is applicable to the safety monitoring and early warning of the UAV in a complex low-altitude environment.

[0173] Embodiment 2:

[0174] As Figure 2 shown, the embodiments of the present invention provide a UAV safety situation awareness system based on multi-source heterogeneous data fusion, including:

[0175] A multi-source heterogeneous data layer, configured to obtain multi-source heterogeneous data of a target UAV in real time.

[0176] The multi-source heterogeneous data layer is a basic component in the entire UAV safety situation awareness system. Its core task is to efficiently and accurately collect data from multiple data sources, ensure the real-time nature and integrity of the data, and provide a solid foundation for subsequent data processing, feature extraction, fusion and safety situation awareness. Specifically, the main data sources involved in the data acquisition layer in the present invention include ADS-B data, remote identification data, alarm configuration data and alarm information, etc.

[0177] The multi-source data fusion layer is configured to preprocess the multi-source heterogeneous data, extract features from the preprocessed multi-source heterogeneous data according to data types to generate multi-source heterogeneous features; perform transformation processing on the multi-source heterogeneous features through a heterogeneous feature self-combination algorithm to generate fusion features; introduce a CLS node into the fusion features to construct an adjacency matrix, and generate single-time-step graph structure data; extract the CLS node vector representation from the single-time-step graph structure data through a graph neural network; and generate a comprehensive feature vector by focusing on key data through an attention mechanism for the CLS node vector representation.

[0178] Data processing is responsible for cleaning, normalizing, time-seriesizing, and feature extracting the data, and passing the processed data to the fusion layer to ensure the quality and consistency of the data.

[0179] As Figure 3 shown, the multi-source data fusion layer fuses multi-source heterogeneous data to generate a comprehensive feature vector. It uses a feature self-combination algorithm to extract implicit associations between heterogeneous features, converts the heterogeneous data into graph structure data, combines with a GNN network to extract the global CLS node vector representation, and focuses on key data through an Attention mechanism to generate a comprehensive feature vector.

[0180] The security situation awareness layer is configured to perform security situation awareness classification on the comprehensive feature vector through the CatBoost algorithm to generate a security situation awareness result.

[0181] The security situation awareness layer monitors and understands the real-time state of a specific environment or system by collecting, analyzing, and interpreting information from multiple data sources. It mainly includes: situation element awareness, situation understanding, and situation prediction. Situation element awareness collects data from multiple sensors; situation understanding analyzes and interprets these data to form a comprehensive understanding of the current state; situation prediction predicts the future situation development based on the current understanding and historical data, so as to take measures in advance to deal with potential threats.

[0182] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0183] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0184] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0185] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0186] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for UAV safety situation awareness based on multi-source heterogeneous data fusion, characterized in that Including: Obtaining multi-source heterogeneous data of a target UAV in real time and preprocessing the multi-source heterogeneous data; Extracting features from the preprocessed multi-source heterogeneous data according to data types to generate multi-source heterogeneous features; Performing transformation processing on the multi-source heterogeneous features through a heterogeneous feature self-combination algorithm to generate fusion features; Introducing a CLS node to the fusion features to construct an adjacency matrix and generating single-time-step graph structure data; Extracting a CLS node vector representation from the single-time-step graph structure data through a graph neural network; Focusing on key data through an attention mechanism for the CLS node vector representation to generate a comprehensive feature vector; Performing security situation awareness classification on the comprehensive feature vector through a CatBoost algorithm to generate a security situation awareness result.

2. The method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion according to claim 1, wherein The multi-source heterogeneous data includes: ADS-B data, remote identification data, alarm configuration data and alarm information data, mission-related data, and device basic data.

3. The method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion according to claim 1, characterized in that The preprocessing of the multi-source heterogeneous data includes: missing value filling, outlier detection and elimination, noise removal, normalization, standardization, and time serialization.

4. The method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion according to claim 1, wherein The feature extraction from the preprocessed multi-source heterogeneous data according to data types includes: extracting time-domain features from time series data; extracting spatial features from position data and motion data; extracting image features from image data and video data.

5. The method for UAV security situation awareness based on multi-source heterogeneous data fusion according to claim 1, characterized in that The transformation processing of the multi-source heterogeneous features through a heterogeneous feature self-combination algorithm includes: Denote the multi-source heterogeneous features at the -th time step as , being the number of features, being the -th feature in . Initialize a learnable vector representation for the multi-source heterogeneous features at each time step , is the dimension of the vector representation, is the th feature in Calculate the vector representations sequentially The likelihood score matrix between features in , the likelihood score matrix The elements in are as follows: In the formula, , is a possibility score calculation function; According to the probability score matrix Calculate the correlation matrix : wherein, is the element activation function, is the learnable inductive bias, is the constant threshold; According to the association matrix , heterogeneous feature combination of multi-source heterogeneous features is performed through matrix multiplication: Wherein, is a splicing operation, is a selection function, is the th feature in after heterogeneous feature combination ; The features after combining heterogeneous features are mapped to a dense and uniform dimensional latent space through a feature mapper to obtain the fused features at the th time step.

6. The method for unmanned aerial vehicle security situation awareness based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The introducing of a CLS node to the fusion features to construct an adjacency matrix and generating single-time-step graph structure data includes: Initialize the CLS node as dimensional space vector , define the subscript set of the nodes of the space vector as , and the subscript set of the nodes of the fusion feature is ; Bi-directionally connect the CLS node with each node of the fused feature to generate an adjacency matrix , the adjacency matrix where the element is as follows: Based on the spatial vector and the adjacency matrix construct the single time-step graph structure data .

7. The method for UAV security situation awareness based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The extracting of a CLS node vector representation from the single-time-step graph structure data through a graph neural network includes: Initializing the mean of the vector representations of the two nodes connected by each edge in the single-time-step graph structure data as the vector representation of each edge through a graph neural network; Concatenating the vector representations of each edge and the two nodes connected thereto and concatenating the vector of each point and the edge connected thereto, and the concatenation result completes message passing and aggregation through a multi-layer perceptron: In the formula, is the vector representation of the th node, is the vector representation of the edge between the th nodes, is the neighbor node of the th node, is a multi-layer perceptron; Updating the vector representations of nodes and edges: After multiple iterative updates, a CLS node vector representation is obtained.

8. The method for unmanned aerial vehicle safety situation awareness based on multi-source heterogeneous data fusion according to claim 1, wherein The focusing on key data through an attention mechanism for the CLS node vector representation to generate a comprehensive feature vector includes: At each time step, first calculate the query value and the similarity with the key value to obtain a score through dot product : Fraction Normalize through the Softmax function to obtain the attention weights at each time step : wherein, is the th time step; The CLS node vector representation for each time step is weighted and aggregated to obtain a comprehensive feature vector : wherein, is the attention weight and the CLS node vector representation at the -th time step.

9. The method for drone safety situation awareness based on multi-source heterogeneous data fusion according to claim 1, characterized in that The performing of security situation awareness classification on the comprehensive feature vector through a CatBoost algorithm to generate a security situation awareness result includes: Constructing and initializing a CatBoost model and using the comprehensive feature vector as a training sample; Calculate the gradient of the CatBoost model for each training sample, and update the CatBoost model using the gradient. The gradient is as follows: In the formula, is the loss function, is the th training sample target value, is round CatBoost model's prediction for the training sample , is the step size; In each round of training, construct round decision trees to minimize the loss function and correct the error of the previous round; wherein, is the desired function, the output of the newly trained decision tree; According to round decision tree obtain the prediction of the round CatBoost model for the training samples as follows :[[]]END]] After multiple rounds of training, the outputs of all decision trees are obtained through weighted summation or voting to obtain the final security situation awareness result.

10. The method for UAV security situation awareness based on multi-source heterogeneous data fusion according to claim 1, wherein, The security situation awareness result includes collision risk, trajectory anomaly, communication interruption, system failure, environmental interference, and normal operation.

Images