Edge computing-supported methods for drone environmental monitoring and data processing

Abstract

This invention relates to the field of UAV communication technology, and particularly to a method for UAV environmental monitoring and data processing supported by edge computing. The method includes: acquiring time-series data streams collected by multi-source environmental sensing devices, compressing and processing them based on a linear projection compression model to obtain compressed feature vectors; calculating communication costs based on current heterogeneous network conditions and energy constraints using a preset communication cost function; and executing different data processing strategies based on different communication costs. This invention achieves efficient edge compression in embedded resource-constrained scenarios; realizes resource-aware data upload control through the communication cost function; and reconstructs the Kalman filter matrix structure based on a Bayesian state fusion method using compressed feature space, effectively supporting local state estimation and uncertainty control, and triggering immediate warnings when environmental indicators exceed limits, effectively enhancing the system's risk perception and response capabilities.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) communication technology, and particularly relates to a method for UAV environmental monitoring and data processing supported by edge computing. Background Technology

[0002] Currently, environmental monitoring systems generally adopt a centralized processing architecture centered on cloud computing. This architecture uses drones as aerial mobile sensing platforms, acquiring multimodal sensing data, including temperature and humidity, gas concentration, particulate matter (PM2.5), thermal radiation, and visible light images, through various environmental sensors. The collected raw data is transmitted back in real time to a remote cloud computing center via wireless communication (usually 4G or 5G links), where high-performance servers perform processing tasks such as data fusion, anomaly detection, and event recognition.

[0003] In this type of architecture, the flight missions performed by drones typically rely on pre-defined static path planning, i.e., waypoints and cruise paths are predefined before the mission begins. These paths are generally based on regional maps or historical mission experience, lacking the ability to dynamically adjust to sudden pollution events or temporary changes in targets. Furthermore, to reduce communication and processing load, some systems introduce simple data compression, image filtering, or resolution downsampling preprocessing operations on the drone side, but overall, the core perception, judgment, and decision-making processes still rely on the cloud.

[0004] This type of centralized architecture has certain advantages in terms of computing power, unified management and centralized analysis, but its system performance and practicality face significant challenges when facing dynamic environments, large-scale sensing areas and tasks with high timeliness requirements. Summary of the Invention

[0005] The purpose of this invention is to provide an edge computing-supported method for UAV environmental monitoring and data processing, aiming to solve the problem that existing technologies face significant challenges in terms of system performance and practicality when dealing with dynamic environments, large-scale sensing areas, and mission scenarios with high timeliness requirements.

[0006] This invention is implemented as follows: a method for environmental monitoring and data processing of unmanned aerial vehicles (UAVs) supported by edge computing, the method comprising:

[0007] The time-series data stream collected by multi-source environmental sensing devices is acquired, and compressed and processed based on a linear projection compression model to obtain the compressed feature vector;

[0008] The communication cost is calculated based on the current heterogeneous network conditions and energy constraints using a preset communication cost function.

[0009] When the communication cost exceeds the preset value, the cloud reporting process is executed, the compressed feature vector is transmitted to the cloud platform, the monitoring task instruction is generated through the cloud platform, and the linear projection compression model is trained and its parameters are updated.

[0010] When the communication cost is lower than the preset value, the observation vector is generated by the linear projection compression model, the observation matrix is ​​constructed, and it is replaced with the compressed form. The dynamic event decision engine is used for anomaly detection and response strategy reasoning.

[0011] Preferably, in the step of acquiring the time-series data stream collected by the multi-source environmental sensing device, the sources of the time-series data stream include at least an airborne temperature sensor, an air quality detector, and a humidity sensor, and are represented in vector form:

[0012]

[0013] in, It is a moment The state vector represents the various environmental information collected by the UAV through different sensors at that moment. for The temperature detected by the onboard temperature sensor at all times. for The air quality value detected by the air quality detector at any given time. for The humidity detected by the humidity sensor at any time. Represents a real number vector with three rows and one column.

[0014] Preferably, in the step of compressing and processing the vector based on a linear projection compression model to obtain the compressed feature vector, the linear projection compression model adopts the TinyML architecture, and its transformation relationship is expressed as follows:

[0015] ;

[0016] in, It is a projection matrix, with a size of 2 rows and 3 columns, which projects the input 3D features. Compressed into 2D features ,right Perform linear feature transformation. , representing the bias vector, has a size of 2 rows and 1 column, and generates the compressed feature vector. Used for subsequent edge collaboration tasks. This represents the compressed feature vector, which has a size of 2 rows and 1 column.

[0017] Preferably, the communication cost function is expressed as:

[0018]

[0019] in, This represents the communication cost, measuring the expense incurred by the drone in performing a communication task under its current state. The higher the value, the higher the communication cost. Indicates the channel congestion coefficient. This indicates the current remaining energy level of the drone. The weighting coefficient is used to control the impact of channel congestion on communication costs. Used to control the impact of surplus energy on communication costs.

[0020] Preferably, when the communication cost Higher than the preset value When the cloud reporting process is executed, the compressed feature vector is transmitted to the cloud platform. The cloud platform generates monitoring task instructions and trains and updates the parameters of the linear projection compression model. In this process, the cloud platform collects data from different UAV nodes, performs global fusion analysis of cross-regional environmental characteristics, and constructs a spatial distribution model and temporal evolution trend for urban pollution monitoring and forest fire situation identification. Based on the fusion analysis results and task scheduling requirements, the cloud platform generates dynamic response strategies, updated model parameters or new monitoring task instructions, and sends them to the corresponding UAV or edge node through the downlink.

[0021] Preferably, when the communication cost is lower than a preset value, in the step of generating observation vectors through a linear projection compression model, constructing an observation matrix, and replacing it with a compressed form, the multimodal observation data is linearly compressed on the edge device using a linear projection compression model to generate observation vectors with lower dimensionality. , observation matrix Replace with compressed format ,in, This is the original observation matrix, representing the set of multimodal, high-dimensional features collected by the UAV at the edge. The projection matrix is ​​generated by a linear projection compression model, which embeds feature dimensionality reduction and structure preservation information to transform high-dimensional features. Compressed into low-dimensional features , The compressed observation matrix is... The linear dimensionality reduction results in smaller feature dimensions while still retaining key features, which can be used for subsequent inference or collaborative tasks.

[0022] Preferably, in the step of mapping the compressed observation vector to the real state space during inference, the state estimation formula is expressed as:

[0023]

[0024] in:

[0025] The compressed observation features output by the edge nodes are obtained from multimodal sensing data through linear projection transformation. The observation matrix in compressed space, To compress the measurement noise covariance matrix in the observation space, and The state estimate and covariance before prediction, and The fused state estimates and covariance are given. The Kalman gain determines the estimated adjustment magnitude, where t is the time index, the current discrete time step, indicating which time point the UAV is at. yes The transpose of .

[0026] Preferably, in the step of anomaly detection and response strategy reasoning through a dynamic event decision engine, anomaly detection and response strategy reasoning for the target event are performed using a multi-source regularized fusion model, which is represented in the following modeling form:

[0027]

[0028] in, Indicates from the The state estimation output of each submodule The weight parameters corresponding to each submodule, This is the L2 regularization coefficient, used to suppress weight overfitting. It is the inferred probability result of abnormal events in environmental monitoring. It is a vector composed of all weights. , It is the squared L2 norm of the overall weight vector.

[0029] Preferably, during the execution of a mission by the UAV, when the state estimate value When the preset threshold is exceeded, the system will automatically respond with a preset strategy, which includes path replanning, coordinating node wake-up, or cloud task reporting.

[0030] Preferably, matrix Structural compression optimization is achieved through QR decomposition or singular value truncation, enabling compression operations to maintain information recognizability while having low computational overhead.

[0031] The edge computing-supported UAV environmental monitoring and data processing method provided by this invention has the following advantages:

[0032] In embedded resource-constrained scenarios, efficient edge compression is achieved by adapting the lightweight feature dimensionality reduction mechanism of the TinyML model;

[0033] A communication cost function that combines channel congestion and battery status is constructed to dynamically optimize the backhaul strategy and achieve resource-aware data upload control.

[0034] Based on the Bayesian state fusion method with compressed feature space, the Kalman filter matrix structure is reconstructed, which effectively supports local state estimation and uncertainty control.

[0035] A local mechanism for identifying and responding to abnormal events is established, which can trigger an early warning in real time when environmental indicators exceed the limits, effectively enhancing the system's risk perception and response capabilities. Attached Figure Description

[0036] Figure 1 This is an application environment architecture diagram of the edge computing-supported UAV environmental monitoring and data processing method provided in the embodiments of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] This invention primarily relates to the interdisciplinary fields of edge computing, the Internet of Things (IoT), and environmental monitoring. Specifically, it includes intelligent collaborative sensing by unmanned aerial vehicles (UAVs), edge data processing and compression methods, and multi-source information fusion and real-time response mechanisms in complex environments. These technologies not only represent current research hotspots in environmental sensing systems but also constitute key supporting technologies in practical scenarios such as smart cities, ecological monitoring, and disaster response.

[0039] Edge computing is an architecture that pushes computing resources and data processing capabilities down to the data-generating source (such as sensors or drone platforms), aiming to reduce network transmission latency, lower bandwidth consumption, and improve the system's real-time processing capabilities. This invention significantly improves the system's response efficiency and task autonomy to dynamic environmental changes by deploying computing tasks on the drone itself or edge nodes, providing a low-latency, high-efficiency solution for wide-area environmental monitoring scenarios.

[0040] The Internet of Things (IoT) is an infrastructure system for enabling interconnected devices and collaborative data sharing. By constructing a heterogeneous network architecture between drones, edge nodes, and ground control platforms, it forms a dynamic sensing and efficient communication-based task collaboration system. In this invention, the IoT architecture supports drone task scheduling and multi-source data fusion in a distributed multi-node environment, providing technical support for high spatiotemporal coverage and device linkage in environmental monitoring.

[0041] Environmental monitoring technology aims to conduct long-term, continuous, and high-precision observation and analysis of elements such as air, water, land, and meteorology in natural or man-made environments. This invention uses unmanned aerial vehicles (UAVs) as the core platform, integrating multimodal sensors to achieve refined collection of parameters such as pollutants, temperature, humidity, and thermal radiation. Combined with edge-side data preprocessing and intelligent judgment mechanisms, it improves the accuracy of identifying key environmental change events and the robustness of monitoring.

[0042] Intelligent Cooperative Perception for UAVs is one of the core innovative modules proposed in this invention. This technology relies on the DroEdge-EMS architecture and constructs an aerial distributed perception network through short-range wireless communication protocols (such as MAVLink) and dynamic task allocation mechanisms among multiple UAVs. The system supports real-time sharing of compressed sensing features and location information collected by each node, and combines lightweight inference models (such as the TinyML module) on edge nodes to perform joint analysis and event judgment on local areas. For example, when the PM2.5 value in a certain area is continuously abnormal, adjacent UAVs can trigger a "hotspot detection" cooperative mechanism, automatically adjusting their flight paths to increase coverage density, completing local risk confirmation and supplementary task sampling, thereby achieving rapid event response without relying on a central control node. This mechanism significantly improves the ability to perceive environmental changes and the system's robustness in complex terrain or signal blind spots.

[0043] Edge-side data processing is a crucial step in improving real-time performance and resource utilization efficiency. To address the challenges of large data volumes and limited communication resources collected by UAVs, this invention proposes an edge processing module integrating compression encoding, feature extraction, and task-specific filtering. This module can perform preliminary data filtering and encoding at the node level, effectively reducing communication load while preserving core information and improving the accuracy and efficiency of subsequent detection algorithms.

[0044] like Figure 1 The diagram shown illustrates the application environment architecture of the edge computing-supported UAV environmental monitoring and data processing method provided in this embodiment of the invention. The method includes:

[0045] The time-series data stream collected by multi-source environmental sensing devices is acquired, and then compressed and processed based on a linear projection compression model to obtain the compressed feature vector.

[0046] In this step, the sensing entry point of the present invention consists of a lightweight aggregator module deployed on a drone or a nearby edge node. Its core task is to significantly reduce the communication cost and energy consumption of raw environmental data during wireless transmission without sacrificing the ability to represent key features. This lightweight aggregator module receives data from an onboard temperature sensor (Temperature Sensor). Air quality detector ), humidity sensor (Humidity, The time-series data stream collected by multi-source environmental sensing devices, such as [devices name], is represented as a column vector.

[0047]

[0048] It is a moment The state vector represents the various environmental information collected by the UAV through different sensors at that moment. for The temperature detected by the onboard temperature sensor at all times. for The air quality value detected by the air quality detector at any given time. for The humidity detected by the humidity sensor at any time. Represents a real number vector with three rows and one column.

[0049] To achieve efficient data compression and processing on edge devices, the system introduces a linear projection compression model based on the TinyML architecture, whose transformation relationship is defined as:

[0050]

[0051] It is a projection matrix, with a size of 2 rows and 3 columns, which projects the input 3D features. Compressed into 2D features ,right Perform linear feature transformation. , representing the bias vector, has a size of 2 rows and 1 column, and generates the compressed feature vector. Used for subsequent edge collaboration tasks. This represents the compressed feature vector, which has a size of 2 rows and 1 column.

[0052] This transformation projects the original 3D perceptual vectors into a low-dimensional latent space, generating compressed feature vectors. Used for subsequent edge collaboration tasks. To ensure a balance between model expressiveness and compactness, the matrix... Structural compression optimization can be achieved through methods such as QR decomposition or Singular Value Decomposition (SVD), which allows the compression operation to maintain information recognizability while having low computational overhead, making it suitable for the low-power microcontroller unit (MCU) operating environment. This formula obtains 2D features from the original three features (temperature, air quality, and humidity) through linear projection, which can be used for subsequent edge collaborative computing (such as task allocation or target recognition).

[0053] The communication cost is calculated based on the current heterogeneous network conditions and energy constraints using a preset communication cost function.

[0054] This step involves the communication and state fusion phase, namely the UAV-Edge Collaborative Subsystem. As the core middleware layer of DroEdge-EMS, the UAV-Edge Collaborative Subsystem is responsible for the intelligent scheduling of data backhaul decisions and the collaborative prediction of multi-source state information. To address heterogeneous network conditions and energy constraints, the system introduces a communication cost function:

[0055]

[0056] in, This represents the current channel congestion coefficient (estimated by metrics such as packet loss rate and latency). This indicates the current remaining energy level of the drone. , This function is used to adjust the weighting coefficients for network load and energy consumption balance. It is embedded in the backhaul policy scheduler and serves as an optimization objective for intelligent decision-making on whether to perform data reporting, thereby extending flight endurance and system lifecycle while ensuring information freshness.

[0057] When the communication cost exceeds the preset value, the cloud reporting process is executed, the compressed feature vector is transmitted to the cloud platform, the monitoring task instruction is generated through the cloud platform, and the linear projection compression model is trained and its parameters are updated.

[0058] In this step, to achieve intelligent control of the communication scheduling strategy, the system sets an adjustable cost threshold parameter. When the calculated communication cost < At this time, the system triggers the cloud reporting process for compressed features. The compressed feature representation... Under suitable transmission conditions, the data is uploaded to a cloud platform or ground control center. This platform not only undertakes the centralized management of feature data but also serves as the remote intelligent hub of the DroEdge-EMS architecture, performing multiple key functions.

[0059] Firstly, the platform aggregates data from different drone nodes to conduct global fusion analysis of cross-regional environmental characteristics, constructing spatial distribution models and temporal evolution trends for macroscopic perception in scenarios such as urban pollution monitoring and forest fire situation identification.

[0060] Secondly, the platform supports periodic evaluation and retraining of edge models. With the support of centralized computing resources, the platform optimizes the accuracy of the models, adjusts parameters, or learns abnormal features to improve subsequent inference capabilities.

[0061] Third, based on the fusion analysis results and task scheduling requirements, the cloud can generate dynamic response strategies, updated model parameters, or new monitoring task instructions, and send them to the corresponding drones or edge nodes via downlink. This feedback process constitutes a closed-loop mechanism in the DroEdge-EMS system, from environmental perception, status uploading, central optimization to local feedback execution, realizing an intelligent control cycle of "perception-decision-feedback-re-perception," significantly enhancing the system's task adaptability and evolutionary optimization capabilities under long-term operation.

[0062] When the communication cost is lower than the preset value, the observation vector is generated by the linear projection compression model, the observation matrix is ​​constructed, and it is replaced with the compressed form. The dynamic event decision engine is used for anomaly detection and response strategy reasoning.

[0063] In this step, if the current communication cost exceeds a threshold, the data is retained locally, and the process transitions to edge-side state estimation and inference. This threshold can be dynamically adjusted based on mission urgency, historical network performance, or flight cycle time, thereby achieving a balance between energy efficiency, latency, and monitoring quality.

[0064] At the state estimation level, the DroEdge-EMS system introduces an improved Kalman Filter (KF) mechanism suitable for edge compression of feature spaces. Unlike traditional KF methods that estimate data on raw high-dimensional sensor data, this system first uses the TinyML module to linearly compress multimodal observation data on the edge device, generating observation vectors with lower dimensionality. Then, state fusion is performed within this compressed space. This processing paradigm fully demonstrates the advantages of edge computing's "on-site intelligence," namely, completing feature compression, state prediction, and anomaly identification locally, greatly reducing communication pressure and computational overhead.

[0065] To accommodate the state reasoning requirements of compressed space, the system will use the original observation matrix Replace with compressed format ,in, This is the projection matrix generated by the TinyML compressed model. This matrix embeds information about feature dimensionality reduction and structure preservation, ensuring that the compressed observation vectors can still be effectively mapped to the true state space during inference. This is the original observation matrix, representing the set of multimodal, high-dimensional features collected by the UAV at the edge. The projection matrix is ​​generated by the linear projection compression model and is used to package high-dimensional features. Compressed into low-dimensional features , The compressed observation matrix is... The linear dimensionality reduction results in smaller feature dimensions while still retaining key features, which can be used for subsequent inference or collaborative tasks.

[0066] The state estimation formula is shown below:

[0067] ;

[0068] in:

[0069] The compressed observation features output by the edge nodes are obtained from multimodal sensing data through linear projection transformation. The observation matrix in compressed space, To compress the measurement noise covariance matrix in the observation space, and The state estimate and covariance before prediction, and The fused state estimates and covariance are given. The Kalman gain determines the estimated adjustment magnitude, where t is the time index, the current discrete time step, indicating which time point the UAV is at. yes The transpose of .

[0070] This mechanism operates on edge devices with an estimation period of 200ms, supporting real-time state tracking in dynamic environments. Leveraging its information fusion capabilities within a compressed space, DroEdge-EMS not only improves the processing performance of edge nodes but also ensures the accuracy of state estimation under compression constraints, providing a practical and efficient intelligent inference capability for autonomous monitoring tasks in complex environments.

[0071] The upper layer of the system is a dynamic event decision engine, forming the intelligent control hub of the "perception-fusion-response" closed loop in DroEdge-EMS. This module aggregates the state estimation results and their uncertainty representations uploaded from the edge layer, and performs anomaly detection and response strategy reasoning for target events through a multi-source regularized fusion model. Its modeling form is as follows:

[0072]

[0073] in, Indicates from the The state estimation output of each submodule The weight parameters corresponding to each submodule, This is the L2 regularization coefficient, used to suppress weight overfitting. It is the inferred probability result of abnormal events in environmental monitoring. It is a vector composed of all weights. , It is the squared L2 norm of the overall weight vector.

[0074] In terms of emergency response, this module embeds an emergency triggering mechanism based on logical rules. For example, in an air quality anomaly detection task, when the state estimate... Exceeding the AQI threshold set by the World Health Organization (WHO) At that time, the system will trigger the following actions:

[0075]

[0076] The decision-making logic can be extended to other parameters (such as sudden temperature rise, sudden humidity drop, etc.) and supports response strategy configuration, including path replanning, collaborative node wake-up and cloud (i.e. cloud platform) task reporting, thereby realizing the self-organization, self-repair and scenario adaptation of unmanned systems.

[0077] In summary, this invention constructs a closed-loop unmanned monitoring system that integrates multimodal environmental perception, edge computing processing, and intelligent decision-making reasoning. Its core advantages are reflected in the following aspects:

[0078] First, the system can perform feature compression and preliminary perception judgment at the edge (such as the UAV itself or near-ground edge nodes), effectively reducing dependence on the cloud and significantly improving the system's real-time performance and task autonomy. Second, by integrating lightweight state estimation mechanisms (such as Kalman filters), DroEdge-EMS still exhibits excellent robustness and prediction accuracy in highly dynamic and uncertain environments, ensuring the temporal continuity and estimation reliability of key state variables.

[0079] Furthermore, the system's embedded communication cost modeling strategy comprehensively considers wireless channel status and remaining energy information, enabling dynamic optimization of communication strategies under resource constraints. This extends the system's task cycle and alleviates transmission bottlenecks. Finally, the introduction of an event-driven mechanism enables the system to respond rapidly to sudden risk factors. Through dynamic evaluation of monitored variables and threshold trigger judgment, the system can automatically execute various strategies, such as early warning push, task rescheduling, or multi-UAV coordinated response, when potential environmental anomalies are detected (e.g., excessive air quality, sudden temperature changes), demonstrating a high degree of intelligence and situational adaptability.

[0080] Overall, this invention, driven by edge intelligence, integrates key mechanisms such as perception compression, state prediction, resource optimization, and event reasoning, demonstrating strong adaptability, scalability, and operational efficiency in complex environmental monitoring tasks. This architecture represents a practical application of the "edge-collaboration-reasoning" paradigm in intelligent unmanned systems, providing theoretical support and a technological foundation for promoting the intelligent, low-power, and highly reliable operation of next-generation autonomous environmental monitoring systems.

[0081] Compared to traditional architectures that rely on centralized cloud computing, this invention is based on an edge computing paradigm. It preprocesses and compresses multimodal environmental data at the UAV itself or near-ground edge nodes, effectively reducing data backhaul frequency and bandwidth consumption, decreasing transmission requirements by approximately 60% compared to the original solution. The system achieves local intelligent inference through an embedded TinyML module, significantly reducing model parameter size and computational energy consumption, and improving deployment adaptability in resource-constrained scenarios. Traditional architectures typically delegate event response to remote servers, resulting in high decision latency. The DroEdge-EMS system, however, integrates state estimation and event recognition mechanisms, supporting real-time response and task scheduling at the UAV end, improving system real-time performance and task continuity. In terms of processing complexity, compared to the parameter-intensive and computationally intensive Transformer model, DroEdge-EMS achieves efficient operation through a lightweight structure, making it suitable for large-scale deployment and long-term operation of intelligent environmental monitoring tasks, with good scalability and task robustness.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for unmanned aerial vehicle environment monitoring and data processing supported by edge computing, characterized in that, The method comprises: Obtaining a time series data stream collected by a multi-source environment perception device, and compressing and processing the time series data stream based on a linear projection compression model to obtain a compressed feature vector; Calculating a communication cost based on a preset communication cost function, current heterogeneous network conditions and energy constraints; When the communication cost is higher than a preset value, a cloud reporting process is performed, the compressed feature vector is transmitted to a cloud platform, a monitoring task instruction is generated by the cloud platform, and the linear projection compression model is trained and parameter updated; When the communication cost is lower than the preset value, an observation vector is generated by the linear projection compression model, an observation matrix is constructed and replaced in a compressed form, and abnormal detection and response strategy reasoning are performed by a dynamic event decision engine; When the communication cost is lower than the preset value, the step of generating an observation vector by a linear projection compression model, constructing an observation matrix, and replacing it with a compressed form is performed by linearly compressing the multi-modal observation data on the edge device by the linear projection compression model to generate an observation vector with lower dimensions , the observation matrix is replaced with a compressed form , wherein is an original observation matrix, representing a multi-modal, high-dimensional feature set collected by the unmanned aerial vehicle on the edge side, is a projection matrix generated by the linear projection compression model, used to compress the high-dimensional features in into low-dimensional features in , is the compressed observation matrix. 2.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 1, characterized in that, In the step of obtaining the time series data stream collected by the multi-source environment perception device, the sources of the time series data stream at least include an airborne temperature sensor, an air quality detector and a humidity sensor, and the time series data stream is represented in a vector form: ； wherein, is a state vector at a time , representing a plurality of environment information collected by different sensors of the UAV at the time, is a temperature detected by an on-board temperature sensor at the time is an air quality value detected by an air quality detector at the time is a humidity detected by a humidity sensor at the time is a humidity detected by a humidity sensor at the time is a humidity detected by a humidity sensor at the time is a humidity detected by a humidity sensor at the time represents a three-by-one real vector. 3.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 2, characterized in that, In the step of compressing and processing the time series data stream based on the linear projection compression model to obtain the compressed feature vector, the linear projection compression model adopts a TinyML architecture, and a transformation relationship thereof is represented as: ； is a projection matrix with size 2 by 3 that compresses the input 3-dimensional feature into a 2-dimensional feature , performs a linear feature transformation on , , represents a bias vector with size 2 by 1 that generates the compressed feature vector for subsequent edge-coordination tasks, represents the compressed feature vector with size 2 by 1. 4.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 1, characterized in that, The communication cost function is represented as: ； wherein, represents a communication cost, measuring the cost paid by the UAV for performing the communication task in the current state, the larger the value, the higher the communication cost, represents a channel congestion coefficient, represents the remaining energy level of the current UAV, is a weight coefficient, used to control the influence degree of the channel congestion on the communication cost, the weight coefficient is used to control the influence degree of the remaining energy on the communication cost. 5.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 1, characterized in that, When the communication cost is higher than a preset value , a cloud reporting process is performed, and the compressed feature vector is transmitted to the cloud platform. In the steps of generating a monitoring task instruction by the cloud platform and training and parameter updating of the linear projection compression model, the cloud platform collects data from different unmanned aerial vehicle nodes, performs global fusion analysis of cross-regional environmental features, constructs a spatial distribution model and a time series evolution trend, and is used for city pollution monitoring and forest fire situation identification. Based on the fusion analysis result and the task scheduling demand, the cloud platform generates a dynamic response strategy, updated model parameters or a new monitoring task instruction, and sends them to the corresponding unmanned aerial vehicle or edge node through the downlink. 6.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 5, characterized in that, In the step of mapping the compressed observation vector to a real state space in an inference process, a state estimation formula is represented as: ； Wherein: The observation vector output for the edge node, obtained by linear projection transformation from the multi-modal perception data, The observation matrix in the compressed space, The measurement noise covariance matrix in the compressed observation space, And The state estimation and covariance before prediction, And The state estimation and covariance after fusion, The Kalman gain, which determines the adjustment amplitude of the estimation, t is the time index, the current discrete time step, represents the unmanned aerial vehicle at which time point, The identity matrix, the dimension of which is consistent with the state covariance matrix , used to keep the dimensions of matrix operations consistent, The transpose matrix of . 7.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 6, characterized in that, In the step of performing abnormal detection and response strategy reasoning by the dynamic event decision engine, abnormal detection and response strategy reasoning of a target event are performed by a multi-source regularization fusion model, and a modeling form is represented as: ； wherein, represents the state estimation output from the i-th sub-module, is the weight parameter corresponding to each sub-module, is the L2 regularization coefficient for inhibiting the weight overfitting phenomenon, is the inferred occurrence probability result of the abnormal event in the environment monitoring, is the vector composed of all weights, , is the square of the L2 norm of the overall weight vector.​ 8.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 7, characterized in that, In the process of executing tasks by the unmanned aerial vehicle, when the state estimation value exceeds a preset threshold, the system automatically responds to a preset strategy, which includes path re-planning, cooperative node wake-up, or cloud task reporting. 9.The edge computing supported unmanned aerial vehicle environment monitoring and data processing method according to claim 3, characterized in that, matrix The structural compression optimization is performed by QR decomposition or singular value truncation, so that the compression operation has low computational overhead while maintaining information discriminability.

Images