An artificial intelligence-based unmanned aerial vehicle VOCs monitoring data management system

Abstract

This invention provides an AI-based drone-based VOCs monitoring data management system, belonging to the interdisciplinary field of environmental monitoring and artificial intelligence. Addressing the problems of limited coverage, poor timeliness, low data quality, and weak collaborative capabilities in traditional VOCs monitoring, the system includes a drone platform, a VOCs monitoring module, an edge intelligent processing module, a communication transmission module, a cloud data analysis and comparison module, and an intelligent early warning module. Through multi-sensor fusion and real-time calibration, edge-cloud collaborative AI analysis, multi-drone collaborative scheduling, and improved federated learning, it achieves high-precision VOCs data acquisition, real-time processing, concentration prediction, tiered early warning, and anomaly tracing, effectively solving the pain points of traditional monitoring and providing efficient decision support for environmental regulation.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and artificial intelligence, specifically to an AI-based unmanned aerial vehicle (UAV) VOCs monitoring data management system. Background Technology

[0002] With the rapid development of industrial production, VOCs, as one of the main sources of air pollution, have made emission monitoring and pollution source tracing a core requirement of environmental governance. Currently, mainstream VOCs monitoring technologies mainly rely on two types of schemes:

[0003] One type is a monitoring system based on fixed sites, such as the "VOCs online monitoring device" disclosed in patent CN215218726U. This type of device can only achieve fixed-point monitoring, with limited coverage, and cannot meet the monitoring needs of large areas and complex terrain.

[0004] Another type is mobile monitoring technology based on drones. A representative example is the patent "An Atmospheric VOCs Sampling Device, Sampling and Analysis Method Based on a Drone" (patent number: CN111781030B) published by Nanjing University et al. Its core technical features are: using a Gaussian diffusion model for source intensity inversion, a preset fixed flight trajectory, offline sampling via a SUMMA canister, and data processing entirely centralized in the cloud. This technology has three major drawbacks: 1. The fixed trajectory cannot adapt to dynamic changes in VOCs concentration, resulting in low monitoring coverage and a high rate of missed detections; 2. Offline sampling and centralized cloud processing lead to data latency exceeding 300ms, failing to meet real-time source tracing requirements; 3. The Gaussian diffusion model inversion accuracy is only ≤82%, and the pollution source location error often exceeds 20m, limiting its practicality.

[0005] Meanwhile, traditional drone monitoring systems generally adopt a centralized cloud computing architecture (as described in the CSDN blog "End-Edge-Cloud Three-Level Intelligent Collaboration Platform"). This architecture suffers from low resource utilization (only 60%-70%), high end-to-end latency (>100ms), and lack of real-time AI processing capabilities on the edge side. When facing multi-drone collaborative monitoring scenarios, it is prone to problems such as data transmission congestion and computing resource overload, further reducing monitoring efficiency and traceability accuracy.

[0006] To address this, an AI-based VOCs monitoring data management system for drones is proposed. Summary of the Invention

[0007] The present invention aims to solve the problems mentioned in the background art by providing an artificial intelligence-based drone VOCs monitoring data management system.

[0008] The specific technical solution is as follows: A drone-based VOCs monitoring data management system, comprising a drone platform, a VOCs monitoring module, an edge intelligent processing module, a communication transmission module, a cloud data analysis and comparison module, and an intelligent early warning module, wherein:

[0009] The drone platform is used to carry the VOCs monitoring module, the edge intelligent processing module, and its own flight control and positioning components. The positioning components are used to obtain the real-time location information of the drone platform.

[0010] The signal output terminal of the VOCs monitoring module is connected to the signal input terminal of the edge intelligent processing module, which is used to collect VOCs concentration data and environmental parameter data in the target area and transmit the collected data to the edge intelligent processing module.

[0011] The signal output terminal of the edge intelligent processing module is connected to the first signal terminal of the communication transmission module. It is used to process, extract features and perform AI analysis on the data transmitted by the VOCs monitoring module. Based on a preset lightweight AI model, it performs real-time anomaly screening on the data and only uploads the original sequence of data segments marked as abnormal or with significant trend changes to the communication transmission module to generate preprocessed data, feature data and AI analysis results.

[0012] The second signal terminal of the communication transmission module communicates bidirectionally with the first signal terminal of the cloud data analysis and comparison module to realize the transmission of raw data, preprocessed data, AI analysis results and control commands between the UAV platform and the cloud data analysis and comparison module.

[0013] The cloud-based data analysis and comparison module has built-in national or local VOCs emission standards. Its second signal terminal is connected to the signal input terminal of the intelligent early warning module. It is used to store, manage and analyze the data transmitted by the communication transmission module, and compare the analyzed VOCs concentration data with the built-in VOCs emission standards to determine whether the emission standards are exceeded.

[0014] The intelligent early warning module is used to generate early warning information and push it to designated objects based on the exceedance judgment results of the cloud data analysis and comparison module and the AI ​​analysis results.

[0015] As a preferred embodiment of the present invention, the VOCs monitoring module includes at least two types of VOCs sensors, an environmental parameter sensor, and a self-calibration unit. The VOCs sensors include at least two of the following: a PID photoionization sensor, a miniature FID flame ionization sensor, and a MOS metal oxide semiconductor sensor, used to collect VOCs concentration data of benzene series compounds, alkanes, and alkenes within the target area. The environmental parameter sensor includes a temperature and humidity sensor and a pressure sensor, used to collect real-time temperature and humidity data and pressure data of the target area. The self-calibration unit has a built-in miniature standard VOCs gas storage component with a replaceable gas tank design and a capacity adapted to the UAV payload. It is used to automatically trigger the calibration process during the ground preparation phase before UAV takeoff. If the sensor data drift exceeds a preset threshold (±5%) during flight, in-flight calibration is triggered during the UAV hovering operation interval. By comparing the measured value of the VOCs sensor for the standard VOCs gas with the standard value, the drift error of the VOCs sensor is corrected.

[0016] In a preferred embodiment of the present invention, the edge intelligent processing module includes a data preprocessing unit, a feature extraction unit, and an AI analysis unit. The signal input terminal of the data preprocessing unit is connected to the signal output terminal of the VOCs monitoring module, and is used to perform real-time noise reduction, accurate calibration, and normalization processing on the raw VOCs concentration data. The signal output terminal of the feature extraction unit is connected to the signal input terminal of the AI ​​analysis unit, and is used to extract time-domain features, frequency-domain features, and time-frequency joint features from the preprocessed data. The time-domain features include mean, variance, and peak value; the frequency-domain features include Fourier transform coefficients; and the time-frequency joint features include wavelet transform energy entropy. The AI ​​analysis unit has a built-in multimodal AI model, which is used to predict the VOCs concentration change trend within a preset time period based on the extracted features, and to identify abnormal VOCs emission events.

[0017] As a preferred embodiment of the present invention, the denoising process of the data preprocessing unit includes sequentially performing moving average filtering and wavelet denoising processing on the original VOCs concentration data. The window length of the moving average filtering is set according to the flight speed of the UAV platform. The calibration process of the data preprocessing unit adopts a compensation model based on environmental parameters. The compensation model uses temperature, humidity, and air pressure data collected by environmental parameter sensors as input variables and includes linear terms, quadratic terms, and interaction terms. The specific formula is as follows:

[0018] ;

[0019] Where: C cal The calibrated VOCs concentration value; C rawThe original VOCs concentration value after noise reduction, in ppm or ppb; T, H, and P are the real-time collected temperature, humidity, and air pressure values, respectively, from temperature, humidity, and air pressure sensors; T0, H0, and P0 are the temperature, humidity, and air pressure values ​​under standard environmental parameters (25℃, 50%RH, 1013hPa); a1 to a9 are compensation coefficients obtained by fitting through standard gas experiments, for example: a1: first-order influence coefficient of temperature; a4: second-order influence coefficient of temperature; a7: interaction term coefficient between temperature and humidity.

[0020] The normalization process of the data preprocessing unit adopts the maximum-minimum normalization method to map the calibrated VOCs concentration data to a preset numerical range.

[0021] In a preferred embodiment of the present invention, the communication transmission module includes a UAV-side communication component and a ground-side communication component; the UAV-side communication component is integrated into the UAV platform and includes a 5G communication unit (supporting URLLC mode) and a BeiDou short message communication unit; the ground-side communication component is connected to the cloud-based data analysis and comparison module and includes a LoRa communication unit and a satellite communication unit; the UAV-side communication component and the ground-side communication component transmit data using a combination of data serialization and transmission protocol, wherein control commands are serialized using Protobuf and transmitted via the MQTT protocol, raw data and preprocessed data are serialized using Protobuf and transmitted via the 5G URLLC protocol, and emergency warning information generated by the intelligent warning module is serialized using Protobuf and transmitted via the BeiDou short message protocol.

[0022] As a preferred embodiment of the present invention, the cloud-based data analysis and comparison module includes a data storage unit, a data analysis unit, and a data visualization unit. The data storage unit adopts a combined architecture of relational database and time-series database for hierarchical storage of raw data, preprocessed data, AI analysis result data, and metadata. The metadata includes sensor calibration records, AI model version information, and national or local VOCs emission standards. The signal input terminal of the data analysis unit is connected to the signal output terminal of the data storage unit, and is used to perform statistical analysis, spatial interpolation analysis, and VOCs diffusion model coupling analysis (HYSPLIT model coupled with CFD computational fluid dynamics model) on the stored data. Simultaneously, it compares the analyzed VOCs concentration data with the VOCs emission standards in the metadata and outputs the exceedance judgment result. The signal input terminal of the data visualization unit is connected to the signal output terminal of the data analysis unit, and is used to display the exceedance judgment result in the form of a GIS map heat map, the VOCs concentration change trend in the form of a time trend curve, and the warning information in the form of a warning list.

[0023] In a preferred embodiment of the present invention, the intelligent early warning module includes an early warning level determination unit and an early warning push unit. The signal input terminal of the early warning level determination unit is connected to the second signal terminal of the cloud data analysis and comparison module, and is used to generate graded early warning signals based on the exceedance judgment results of the cloud data analysis and comparison module and the abnormal emission events identified by the AI ​​analysis unit, combined with preset rules. The graded early warning signals include blue early warning, yellow early warning, and red early warning. The signal input terminal of the early warning push unit is connected to the signal output terminal of the early warning level determination unit, and is used to push the graded early warning signals to the terminal devices of environmental regulatory personnel via SMS and mobile applications.

[0024] As a preferred embodiment of the present invention, the UAV platform supports collaborative operation of multiple UAVs, employing a hexacopter UAV model. The cloud-based data analysis and comparison module further includes a multi-UAV collaborative scheduling unit. The first signal terminal of the multi-UAV collaborative scheduling unit is connected to the second signal terminal of the communication transmission module, used to receive real-time status data from each UAV platform. The real-time status data includes remaining flight time, current location, and the amount of data collected. The second signal terminal of the multi-UAV collaborative scheduling unit is connected to the flight control components of each UAV platform, used to dynamically allocate monitoring grids based on the VOCs risk level of the monitoring area and the real-time status data of each UAV platform. The multi-UAV collaborative scheduling unit is also used to fuse the monitoring data of the same time stamp and adjacent locations transmitted by multiple UAV platforms using a Kalman filter algorithm to eliminate spatial measurement errors. When a single UAV platform fails and loses connection, the monitoring task of the failed UAV platform is automatically assigned to other online UAV platforms.

[0025] In a preferred embodiment of the present invention, the edge intelligent processing module and the cloud data analysis and comparison module achieve collaborative optimization of AI model parameters through a federated learning framework. The cloud data analysis and comparison module further includes a federated learning center unit. The first signal terminal of the federated learning center unit is connected to the edge intelligent processing module of each UAV platform and is used to receive local AI model gradient data uploaded by each edge intelligent processing module. The local AI model gradient data is generated based on the desensitized data collected by each UAV platform. The second signal terminal of the federated learning center unit is connected to each edge intelligent processing module and is used to weight the local model gradient according to the actual data signal-to-noise ratio (rather than the historical drift coefficient) after self-calibration of each UAV VOCs monitoring module. The higher the signal-to-noise ratio, the greater the weight. The FedAvg algorithm is used to aggregate the local AI model gradient data to generate global AI model parameters, and the global AI model parameters are sent to each edge intelligent processing module to update the multimodal AI model in the AI ​​analysis unit.

[0026] As a preferred embodiment of the present invention, the cloud data analysis and comparison module further includes an anomaly tracing unit; the signal input end of the anomaly tracing unit is connected to the signal output end of the intelligent early warning module, and is used to extract the timestamp of the early warning event and the monitoring data within a preset time period before and after the early warning signal is triggered by the intelligent early warning module; the anomaly tracing unit is also used to calculate the VOCs diffusion path corresponding to the early warning event by combining the VOCs meteorological inversion model (HYSPLIT+CFD coupled model) and lock the potential emission source area (positioning error ≤50m); the signal output end of the anomaly tracing unit is connected to the data storage unit, and is used to call the historical enterprise emission list of the potential emission source area stored in the data storage unit, match the VOCs component characteristics in the monitoring data with the VOCs component information in the enterprise emission list, locate the suspicious emission enterprise, and push the information of the suspicious emission enterprise synchronously with the early warning signal.

[0027] The present invention has the following beneficial effects:

[0028] 1. More reliable data quality: Solves problems of inaccurate and inconsistent data.

[0029] The system utilizes multiple VOCs sensors (PID + miniature FID) to cover various pollutants, avoiding blind spots in single-sensor monitoring. The self-calibration unit combines conventional ground calibration with air threshold-triggered calibration, along with an environmental parameter compensation model that includes interactive terms, to eliminate interference from temperature, humidity, and air pressure. This ensures that monitoring data from different time periods and environments reflects the true VOCs concentration, providing a reliable foundation for subsequent analysis.

[0030] 2. More efficient response: Solves problems of poor real-time performance and inability to predict.

[0031] The edge intelligent processing module completes data preprocessing and AI analysis locally, without waiting for centralized cloud processing. Abnormal emission identification and short-term concentration prediction can be completed quickly. Based on the 5G URLLC protocol, the end-to-end latency is ≤150ms, and the early warning response speed is significantly improved compared with the traditional mode. At the same time, it can predict the concentration trend in the next few hours, allowing regulators to deploy control measures in advance to prevent pollution from spreading.

[0032] 3. More comprehensive monitoring coverage: solves the problems of limited coverage and gaps in monitoring.

[0033] Multi-drone collaborative scheduling dynamically allocates monitoring grids based on risk levels, with higher-risk areas receiving denser monitoring. Combined with the high payload capacity and fault compensation mechanism of hexacopter drones, it avoids monitoring gaps caused by insufficient battery life of single drones, ensuring that there are no monitoring gaps in complex areas (such as large industrial parks), with a coverage area far exceeding that of fixed stations and single drones.

[0034] 4. More precise decision support: solving the problems of fragmented and difficult-to-use information.

[0035] The cloud-based data analysis and comparison module integrates multi-source data and transforms abstract data into intuitive heat maps and trend curves through spatial interpolation and HYSPLIT+CFD coupled diffusion simulation. Combined with tiered early warning and anomaly tracing, it provides regulatory personnel with full-chain information on "where the pollution is, how much the concentration is, whether it exceeds the standard, and who the source is," eliminating the need for manual sifting through massive amounts of data and significantly improving decision-making efficiency.

[0036] 5. Enhanced system adaptability: Solves the problems of models becoming outdated easily and being limited by specific scenarios.

[0037] While protecting data privacy, the federated learning framework optimizes AI models by weighting and aggregating multi-region monitoring experience based on calibrated data signal-to-noise ratio. Even in new scenarios (such as different types of chemical plants or remote mountainous areas), the models can quickly adapt and avoid a decline in prediction accuracy after long-term use. At the same time, multiple types of communication modules (5G / BeiDou / LoRa) are adapted to different environments, ensuring stable operation in areas with and without public networks, making it applicable to a wider range of scenarios. Attached Figure Description

[0038] Figure 1 A schematic diagram illustrating the architecture of an AI-based UAV VOCs monitoring data management system provided in an embodiment of the present invention;

[0039] Figure 2 A comparison chart of VOCs concentration trends;

[0040] Figure 3 This is a comparison chart of early warning response times;

[0041] Figure 4 A comparison chart of data consistency deviations;

[0042] Figure 5 This is a graph showing the efficiency of multi-machine collaborative monitoring.

[0043] (Attached image description:) Figure 2 The diagram shows the concentration changes of the system monitoring data and traditional fixed stations over a 24-hour period. The solid blue line represents the system data, and the dashed red line represents the traditional data. Figure 3 Comparing the response times of the system (solid green line) and the traditional method (dashed purple line) in 10 abnormal events, the average response time of the system is significantly shortened; Figure 4 The measurement deviation of calibrated (sky blue) and uncalibrated (salmon) data is displayed in different time periods, and the deviation after calibration is stable within 2%. Figure 5 Comparing the effective duration of a single drone (blue) and multi-drone collaboration (red) over a 5-day monitoring period, multi-drone collaboration achieves 2.5 times the effective duration of a single drone. Detailed Implementation

[0044] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0045] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0046] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0047] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0048] Reference Figure 1-5 The present invention provides an artificial intelligence-based drone VOCs monitoring data management system, such as... Figures 1-5 As shown, it includes a drone platform, a VOCs monitoring module, an edge intelligent processing module, a communication transmission module, a cloud data analysis and comparison module, and an intelligent early warning module, among which:

[0049] The drone platform is used to carry the VOCs monitoring module, the edge intelligent processing module, and its own flight control and positioning components. The positioning components are used to obtain the real-time location information of the drone platform.

[0050] The signal output terminal of the VOCs monitoring module is connected to the signal input terminal of the edge intelligent processing module to collect VOCs concentration data and environmental parameter data in the target area, and transmit the collected data to the edge intelligent processing module.

[0051] The signal output end of the edge intelligent processing module is connected to the first signal end of the communication transmission module. It is used to process, extract features and perform AI analysis on the data transmitted by the VOCs monitoring module. Based on the preset lightweight AI model, it performs real-time anomaly screening on the data and only uploads the original sequence of data segments marked as abnormal or with significant trend changes to the communication transmission module to generate preprocessed data, feature data and AI analysis results.

[0052] The second signal terminal of the communication transmission module communicates bidirectionally with the first signal terminal of the cloud data analysis and comparison module to realize the transmission of raw data, preprocessed data, AI analysis results and control commands between the UAV platform and the cloud data analysis and comparison module;

[0053] The cloud-based data analysis and comparison module has built-in national or local VOCs emission standards. Its second signal terminal is connected to the signal input terminal of the intelligent early warning module. It is used to store, manage and analyze the data transmitted by the communication transmission module, and compare the analyzed VOCs concentration data with the built-in VOCs emission standards to determine whether the emission standards are exceeded.

[0054] The intelligent early warning module is used to generate early warning information and push it to designated objects based on the results of the cloud data analysis and comparison module's judgment of exceeding the standard and the results of AI analysis.

[0055] By establishing a closed-loop architecture encompassing a drone platform, a VOCs monitoring module, an edge intelligent processing module, a communication transmission module, a cloud data analysis and comparison module, and an intelligent early warning module, the signal connections and data flows of each module are clearly defined, solving the problems of decentralized and delayed response in traditional VOCs monitoring. The modules collaborate to achieve a complete process from data acquisition to early warning push. Edge processing reduces cloud pressure, cloud-based emission standards directly meet regulatory needs, and the early warning module breaks down barriers between monitoring data and enforcement actions. Ultimately, this transforms VOCs monitoring from passive acquisition to proactive early warning, and from decentralized to integrated systems, providing end-to-end technical support for environmental regulation.

[0056] Specifically, in this embodiment, the VOCs monitoring module includes at least two types of VOCs sensors, an environmental parameter sensor, and a self-calibration unit. The VOCs sensors include at least two of the following: a PID photoionization sensor, a miniature FID flame ionization sensor, and a MOS metal oxide semiconductor sensor. Component differentiation is achieved through "differentiated response characteristics + data fusion": the PID photoionization sensor has significantly higher sensitivity (detection coefficient ≥ 0.95) for aromatic VOCs such as benzene series compounds than for alkanes; the miniature FID flame ionization sensor exhibits higher linearity (R0.95) for hydrocarbons such as alkanes and alkenes. 2(≥0.98) is superior to benzene series compounds; the feature extraction unit of the edge intelligent processing module distinguishes between benzene series compounds, alkanes, and alkenes by comparing the peak response ratio and response time difference of the two types of sensors, combined with a preset component feature library (identification accuracy ≥92%), and is used to collect VOCs concentration data of benzene series compounds, alkanes, and alkenes in the target area; the environmental parameter sensors include temperature and humidity sensors and air pressure sensors, used to collect real-time temperature, humidity, and air pressure data of the target area; the miniature FID sensor integrates a miniature safety hydrogen cylinder (capacity 50mL, pressure 0.5MPa, weight ≤150g), adopts a low-pressure slow-release design, and the hydrogen consumption is ≤1mL / min, which is suitable for the flight endurance requirements of drones; the hydrogen cylinder and the sensor chamber are connected by a one-way solenoid valve, which automatically checks for leaks before flight and monitors the hydrogen pressure in real time during flight. When the pressure is lower than 0.1MPa, a low pressure alarm is triggered, and the system switches to P simultaneously. The ID single-sensor operating mode ensures continuous monitoring. The self-calibration unit has a built-in miniature standard VOCs gas storage component with a replaceable gas tank design. The capacity is adapted to the UAV payload and is used to automatically trigger the calibration process during the ground preparation phase before UAV takeoff. During the ground preparation phase: the standard gas introduction process is automatically triggered, and the deviation δ0 between the sensor's measured value and the standard value is recorded. If δ0 > ±3%, the initial calibration is completed through a linear correction algorithm. During flight, an in-flight calibration is triggered every 30 minutes (executed when the UAV is hovering, taking ≤10 seconds). 50ppm isobutylene standard gas is introduced through the miniature standard gas tank, and the deviation δ1 between the measured value and the standard value is calculated. If δ1 > ±5%, the sensor is determined to have drift. The sensor output value is corrected through a proportional-integral correction algorithm. By comparing the measured value of the VOCs sensor with the standard value of the standard VOCs gas, the drift error of the VOCs sensor is corrected.

[0057] By combining multiple types of VOCs sensors, covering various pollutants such as benzene compounds, alkanes, and olefins, blind spots in single-sensor monitoring are avoided. Environmental parameter sensors collect temperature, humidity, and air pressure data, providing crucial information for subsequent calibration. The self-calibration unit employs a combination of conventional ground calibration and in-flight threshold-triggered calibration, coupled with a miniature replaceable gas tank. This adapts to the payload requirements of UAVs without adding extra endurance burden, while effectively correcting sensor drift errors. The synergy of these three components ensures stable and consistent monitoring data across different environments and time periods, providing reliable raw data for subsequent analysis.

[0058] Specifically, in this embodiment, the edge intelligent processing module includes a data preprocessing unit, a feature extraction unit, and an AI analysis unit. The signal input of the data preprocessing unit is connected to the signal output of the VOCs monitoring module, and is used to perform real-time denoising, accurate calibration, and normalization processing on the raw VOCs concentration data. The signal input of the feature extraction unit is connected to the signal output of the data preprocessing unit, and the signal output is connected to the signal input of the AI ​​analysis unit. It is used to extract time-domain features, frequency-domain features, and time-frequency joint features from the preprocessed data. The time-domain features include mean, variance, and peak value; the frequency-domain features include Fourier transform coefficients; and the time-frequency joint features include wavelet transform energy entropy. The AI ​​analysis unit has a built-in multimodal AI model, which is used to predict the VOCs concentration change trend within a preset time period based on the extracted features, and to identify abnormal VOCs emission events.

[0059] The edge intelligent processing module integrates data preprocessing, feature extraction, and AI analysis functions. The data preprocessing unit first optimizes the raw data, while the feature extraction unit comprehensively captures the time-domain, frequency-domain, and time-frequency joint features of VOCs concentration, providing multi-dimensional support for AI analysis. A lightweight AI model filters out abnormal data in real time, uploading only key segments to reduce data transmission volume. This design allows abnormal emission identification and concentration trend prediction to be completed quickly at the edge, without relying on centralized cloud processing, improving response speed. At the same time, comprehensive feature extraction makes the AI ​​analysis results more consistent with reality.

[0060] Specifically, in this embodiment, the denoising process of the data preprocessing unit includes sequentially performing moving average filtering and wavelet denoising on the original VOCs concentration data. The window length of the moving average filtering is set according to the flight speed of the UAV platform. The calibration process of the data preprocessing unit adopts a compensation model based on environmental parameters. The compensation model uses temperature, humidity, and air pressure data collected by environmental parameter sensors as input variables and includes linear terms, quadratic terms, and interaction terms. The specific formula is as follows:

[0061] ;

[0062] Where: C cal The calibrated VOCs concentration value; C raw The original VOCs concentration value after noise reduction, in ppm or ppb; T, H, and P are the real-time collected temperature, humidity, and air pressure values, respectively, from temperature, humidity, and air pressure sensors; T0, H0, and P0 are the temperature, humidity, and air pressure values ​​under reference environmental conditions (usually set as standard conditions: 25℃, 50%RH, 1013hPa); a1 to a9 are compensation coefficients obtained by fitting through standard gas experiments, for example: a1: first-order influence coefficient of temperature; a4: second-order influence coefficient of temperature; a7: interaction term coefficient between temperature and humidity.

[0063] The normalization process of the data preprocessing unit adopts the maximum-minimum normalization method to map the calibrated VOCs concentration data to a preset numerical range.

[0064] The combined use of moving average filtering and wavelet denoising effectively removes different types of noise, preserving more effective data features compared to single denoising methods. The 9-parameter compensation model includes linear, quadratic, and interactive terms, fully considering the impact of temperature, humidity, and air pressure on monitoring results and eliminating environmental interference. Max-min normalization unifies the data format, avoiding analytical biases caused by differences in data magnitude from different sensors. This series of preprocessing operations makes VOCs concentration data closer to true values, laying a high-quality data foundation for subsequent feature extraction and AI model calculations.

[0065] Derivation of the equation: This equation is based on multivariate quadratic regression theory and aims to capture the nonlinear influence and interaction of environmental parameters (temperature, humidity, air pressure) on VOCs sensor readings. The derivation process is as follows:

[0066] 1. Experimental design: In a controlled environment chamber, different combinations of temperature (-10℃ to 50℃), humidity (20% to 95%RH), and air pressure (900hPa to 1100hPa) were set up, and standard VOCs gas (such as isobutylene) of known concentrations were introduced.

[0067] 2. Data Acquisition: Record the sensor output value C under different environmental conditions. raw Compared with the actual standard concentration C std .

[0068] 3. Model Establishment: Assuming a quadratic relationship between sensor error and environmental parameter deviation, and introducing interaction terms to capture the interaction effect, the following generalized model is established:

[0069] ;

[0070] Where f(T,H,P) is the compensation function.

[0071] 4. Coefficient Fitting: The experimental data are fitted using the least squares method, and the coefficients a1 to a9 are solved to minimize the sum of squared residuals.

[0072] ;

[0073] 5. Model Validation: Cross-validation is used to check the model's prediction error and ensure robustness.

[0074] This model is superior to the linear model because it better reflects the response characteristics of the sensor in real-world complex environments.

[0075] Example: Taking a flight monitoring test as an example:

[0076] The sensor reads the original concentration Craw =45.6ppm; Environmental parameters: temperature T=35℃, humidity H=80%RH, air pressure P=1005hPa; Reference conditions: T0=25℃, H0=50%RH, P0=1013hPa;

[0077] Compensation coefficients (example values): a1=0.002, a2= -0.001, a3=0.0005, a4=0.0001, a5=0.00005.

[0078] a6= - 0.00001, a7= - 0.0002, a8=0.0001, a9=0.00005.

[0079] Substitute into the equation: C cal =45.6×[1+0.002×(35 - 25) - 0.001×(80- 50)+······]=45.6×1.012=46.15ppm.

[0080] The calibrated concentration is closer to the true value, eliminating the positive bias under high temperature and high humidity conditions.

[0081] Technical effect

[0082] 1. Improve data accuracy: By using nonlinear terms and interaction terms, environmental interference can be compensated more accurately, reducing false alarms and missed alarms.

[0083] 2. Enhance data consistency: Make monitoring data in different environments comparable and avoid data fluctuations caused by diurnal temperature differences and seasonal changes.

[0084] 3. Reduced calibration frequency: Compared to the simple linear model, this model is more robust and extends the effective calibration cycle of the sensor.

[0085] 4. Improved model generalization ability: Applicable to various VOCs sensors (PID, FID, MOS), enhancing the overall reliability of the system.

[0086] Working principle and process:

[0087] The equation operates according to the following procedure during the calibration process of the data preprocessing unit:

[0088] 1. Data Input: Obtain the denoised concentration C from the VOCs monitoring module. raw T, H, P are obtained from environmental sensors.

[0089] 2. Parameter call: Read the reference values ​​T0, H0, P0 and compensation coefficients a1 to a9 from memory.

[0090] 3. Calculate the compensation factor: Calculate the compensation factor based on the equation:

[0091] f=1+a1(T - T0)+·······+a9(H - H0)(P - P0).

[0092] 4. Output calibration value: Calculate C cal =C raw ×f.

[0093] 5. Transfer to the next unit: The calibrated data is sent to the feature extraction unit for subsequent analysis.

[0094] This process is embedded in an edge intelligence processing module and runs in real time, ensuring that each frame of data is environmentally compensated, providing high-quality input for AI analysis.

[0095] Specifically, in this embodiment, the communication transmission module includes a UAV-side communication component and a ground-side communication component. The UAV-side communication component is integrated into the UAV platform and includes a 5G communication unit (supporting URLLC mode) and a BeiDou short message communication unit. The ground-side communication component is connected to the cloud-based data analysis and comparison module and includes a LoRa communication unit and a satellite communication unit. The communication between the UAV-side and the ground-side adopts a "protocol-data type" matching design: low-bandwidth data (control commands, UAV status data, alarm prompts) is serialized using Protobuf and transmitted via the LoRa private protocol (transmission rate 1-5kbps, transmission distance ≥3km); large data streams (raw VOCs concentration data, pre-processed data, etc.) are transmitted via the LoRa private protocol. Data processing and AI analysis results are serialized using Protobuf and compressed using LZ4 (compression ratio ≥ 3:1) and then transmitted via the 5G URLLC protocol (bandwidth ≥ 100Mbps, end-to-end latency ≤ 150ms). Emergency warning information is serialized using Protobuf and transmitted via a dual-link system of "BeiDou short message (main channel) + 5G SMS (backup channel)". Control commands are serialized using Protobuf and transmitted via the MQTT protocol, while raw and pre-processed data are serialized using Protobuf and transmitted via the 5G URLLC protocol. Emergency warning information generated by the intelligent warning module is serialized using Protobuf and transmitted via the BeiDou short message protocol.

[0096] The communication transmission module integrates multiple communication units. 5G URLLC ensures high-speed transmission in public network environments, while BeiDou short message service and satellite communication adapt to scenarios without public networks, ensuring uninterrupted data transmission in different environments. It employs a combination of data serialization and transmission protocols, using Protobuf serialization to compress data volume and matching different data types with the optimal transmission protocol (MQTT ensures command reliability, 5G URLLC improves data transmission efficiency, and BeiDou short message service prioritizes emergency warnings). This design achieves reliable, efficient, and seamless data transmission, avoiding transmission bottlenecks caused by a single communication method or protocol.

[0097] Specifically, in this embodiment, the cloud-based data analysis and comparison module includes a data storage unit, a data analysis unit, and a data visualization unit. The data storage unit adopts a combined architecture of relational database and time-series database to store raw data, preprocessed data, AI analysis results data, and metadata in layers. The metadata includes sensor calibration records, AI model version information, and national or local VOCs emission standards. The signal input terminal of the data analysis unit is connected to the signal output terminal of the data storage unit. It is used to perform statistical analysis, spatial interpolation analysis, and VOCs diffusion model coupling analysis (HYSPLIT model coupled with CFD computational fluid dynamics model) on the stored data. At the same time, it compares the VOCs concentration data obtained from the analysis with the VOCs emission standards in the metadata and outputs the exceedance judgment result. The signal input terminal of the data visualization unit is connected to the signal output terminal of the data analysis unit. It is used to display the exceedance judgment result in the form of a GIS map heat map, the VOCs concentration change trend in the form of a time trend curve, and the warning information in the form of a warning list.

[0098] The data storage unit adopts a hierarchical storage architecture, classifying and managing different types of data for easy querying and tracing. The data analysis unit combines HYSPLIT and CFD models for coupled analysis, considering both atmospheric diffusion patterns and the influence of local regional fluid dynamics, making the judgment of exceedances and diffusion simulation more accurate. The data visualization unit displays data according to data type using heatmaps, time trend curves, and warning lists, intuitively presenting pollution distribution, concentration changes, and warning information. This design makes complex data easier to interpret, providing regulatory personnel with comprehensive information on pollution distribution, concentration trends, and exceedance situations, improving decision-making efficiency.

[0099] Specifically, in this embodiment, the intelligent early warning module includes an early warning level determination unit and an early warning push unit. The signal input terminal of the early warning level determination unit is connected to the second signal terminal of the cloud data analysis and comparison module, and is used to generate graded early warning signals based on the exceedance judgment results of the cloud data analysis and comparison module and the abnormal emission events identified by the AI ​​analysis unit, combined with preset rules. The graded early warning signals include blue early warning, yellow early warning and red early warning. The signal input terminal of the early warning push unit is connected to the signal output terminal of the early warning level determination unit, and is used to push the graded early warning signals to the terminal devices of environmental regulatory personnel via SMS and mobile applications.

[0100] The early warning level determination unit classifies early warnings into different levels based on the degree of exceedance and the type of abnormal event, allowing regulatory personnel to quickly distinguish the level of urgency and avoid wasting resources or overlooking urgent situations by treating all warnings the same. The early warning push unit pushes information through multiple channels such as SMS and mobile applications, ensuring that regulatory personnel can receive early warning content in a timely manner regardless of whether they are at the monitoring center. This design shortens the time from anomaly detection to initiation of response, making early warning responses more targeted and improving the flexibility and timeliness of environmental supervision.

[0101] Specifically, in this embodiment, the drone platform supports collaborative work of multiple drones, employing a hexacopter drone model. The cloud-based data analysis and comparison module also includes a multi-drone collaborative scheduling unit. The first signal terminal of the multi-drone collaborative scheduling unit is connected to the second signal terminal of the communication transmission module, used to receive real-time status data from each drone platform. The real-time status data includes remaining flight time, current location, and the amount of data collected. The multi-drone collaborative scheduling unit dynamically allocates the "monitoring-return-to-recharge" sequence based on the remaining flight time of each drone (threshold set at 15 minutes). High-risk area monitoring tasks adopt a "rotational replenishment" mode to ensure that when a single drone returns for battery replacement, other drones automatically cover its monitoring. The system uses a grid to prevent monitoring interruptions; when three drones work together, the cumulative effective monitoring time per day can reach 8 hours (8.7 times the endurance of a single drone); the second signal terminal of the multi-drone collaborative scheduling unit is connected to the flight control components of each drone platform to dynamically allocate monitoring grids based on the VOCs risk level of the monitoring area and the real-time status data of each drone platform; the multi-drone collaborative scheduling unit is also used to fuse the monitoring data of the same time stamp and adjacent positions transmitted by multiple drone platforms using the Kalman filter algorithm to eliminate spatial measurement errors; when a single drone platform fails and loses connection, the monitoring task of the failed drone platform is automatically assigned to other online drone platforms.

[0102] Employing a hexacopter UAV enhances payload capacity and flight stability, adapting to the needs of carrying multiple devices. A multi-UAV collaborative scheduling unit dynamically allocates monitoring grids based on the risk level of the monitored area, allowing for denser monitoring of high-risk areas. Simultaneously, a Kalman filter algorithm is used to fuse multi-UAV data, eliminating spatial measurement errors. A fault compensation mechanism automatically assigns tasks when a single UAV loses contact, preventing monitoring gaps. This design solves the problems of limited coverage and insufficient endurance of a single UAV, enabling continuous and uniform monitoring of large, complex areas, and improving the comprehensiveness and reliability of monitoring coverage.

[0103] Specifically, in this embodiment, the edge intelligence processing module and the cloud data analysis and comparison module achieve collaborative optimization of AI model parameters through a federated learning framework. The cloud data analysis and comparison module also includes a federated learning center unit. The first signal terminal of the federated learning center unit is connected to the edge intelligence processing module of each UAV platform and is used to receive local AI model gradient data uploaded by each edge intelligence processing module. The local AI model gradient data is generated based on the desensitized data collected by each UAV platform. The second signal terminal of the federated learning center unit is connected to each edge intelligence processing module and is used to weight the local model gradient according to the actual data signal-to-noise ratio after self-calibration of each UAV VOCs monitoring module. The higher the signal-to-noise ratio, the greater the weight. The FedAvg algorithm is used to aggregate the local AI model gradient data to generate global AI model parameters, and the global AI model parameters are sent to each edge intelligence processing module to update the multimodal AI model in the AI ​​analysis unit.

[0104] By employing a federated learning framework, edge devices only upload local model gradients rather than raw data, effectively protecting data privacy and resolving the data silos and privacy leaks inherent in traditional centralized model training. The federated learning central unit weights gradients based on the calibrated data signal-to-noise ratio, giving higher weight to high-quality data, and then aggregates these gradients using the FedAvg algorithm to generate global model parameters. This design allows each edge model to learn from multi-region monitoring experience, adapting to different scenarios and preventing a decline in prediction accuracy after long-term use. Simultaneously, it highlights the contribution of high-quality data and improves the analytical performance of the global model.

[0105] Specifically, in this embodiment, the cloud data analysis and comparison module also includes an anomaly tracing unit; the signal input end of the anomaly tracing unit is connected to the signal output end of the intelligent early warning module, and is used to extract the timestamp of the early warning event and the monitoring data within a preset time period before and after when the intelligent early warning module triggers the early warning signal; the anomaly tracing unit is also used to calculate the VOCs diffusion path corresponding to the early warning event by combining the VOCs meteorological inversion model (HYSPLIT+CFD coupled model) and lock the potential emission source area (positioning error ≤50m); the signal output end of the anomaly tracing unit is connected to the data storage unit, and is used to call the historical enterprise emission list of the potential emission source area stored in the data storage unit, match the VOCs component characteristics in the monitoring data with the VOCs component information in the enterprise emission list, locate the suspicious emission enterprise, and push the information of the suspicious emission enterprise and the early warning signal simultaneously.

[0106] After an alert is triggered, the anomaly tracing unit accurately extracts monitoring data from key time periods, avoiding blind searching through massive amounts of historical data. It combines the HYSPLIT and CFD coupled models to invert diffusion paths, fully considering atmospheric transport patterns and local hydrodynamic influences to narrow down the range of potential emission sources. By matching the VOCs component characteristics in the enterprise's emission inventory, it accurately locates suspected emitting enterprises. This design solves the problem of traditional anomaly tracing—"knowing there is an anomaly, but not knowing who emitted it"—providing a clear direction for environmental law enforcement and improving the accuracy of source tracing and enforcement.

[0107] The path planning algorithm employs the reinforcement learning PPO (Proximal Policy Optimization) algorithm. Specific parameters and implementation details are as follows:

[0108] State space: S = [current position (x, y, z), remaining battery power (0-100%), VOCs concentration gradient (ppm / m²), covered area (m²)] 2 Meteorological parameters (wind speed / wind direction)];

[0109] Action space: A = [flight direction (0-360°), flight speed (5-15m / s), sampling frequency (5-20Hz)];

[0110] Reward function: R = 0.6 × coverage improvement rate + 0.3 × source tracing error reduction rate - 0.1 × energy consumption ratio;

[0111] Training parameters: learning rate 1e-4, batch size 64, training iterations 500 rounds, global parameters are synchronized to the cloud every 10 seconds via federated learning;

[0112] Execution logic: The edge layer takes the environmental status into real time, outputs the optimal action command, and dynamically adjusts the flight path to ensure that the monitoring coverage is maximized under the constraint of endurance (endurance time ≥ 55 minutes).

[0113] The specific interaction logic of the system modules is as follows:

[0114] The three-level collaborative interaction process of "device-edge-cloud":

[0115] End layer (UAV): PID sensor + GPS + weather sensor (sampling frequency 10Hz) collect multimodal data and transmit it to the edge module through MIPI-CSI interface;

[0116] Edge layer: After receiving data, preprocessing and AI inference are completed within 10ms to generate preliminary pollution source location results and path adjustment instructions, and the compressed data (1 packet every 2 seconds) is uploaded to the cloud simultaneously;

[0117] In the cloud: Alibaba Cloud ECS servers receive edge data, store it in a MySQL database (supporting 100,000 data points / day), incrementally train the AI ​​model through a GPU cluster (NVIDIA A100) (updating model parameters every 24 hours), and then distribute the optimized model to the edge layer;

[0118] Interaction protocol: The end-to-edge communication adopts the LoRa proprietary protocol (transmission distance ≥3km), and the edge-to-cloud communication adopts 5GURLLC, with a total end-to-end latency ≤150ms.

[0119] Specifically, in this embodiment, the "gradient-weighted optimization" of the improved federated learning refers to introducing a sensor drift correction coefficient as the basis for weighting the local model gradient on the basis of the standard FedAvg algorithm. By dynamically adjusting the contribution of the gradients of different UAV edge nodes, the adaptability of the global model to differences in data quality is improved.

[0120] Mathematical model: Suppose the system contains N UAV edge nodes, and the local model gradient of the i-th node is g. i The drift correction coefficient fed back by the VOCs sensor self-calibration unit is w. i (Value range [0,1], the smaller the drift, the better w) i The closer the gradient is to 1, the better the global gradient aggregation formula becomes:

[0121] ;

[0122] in:

[0123] n i Let be the number of local training samples for the i-th node;

[0124] w i These are gradient weights;

[0125] G represents the aggregated global model gradient.

[0126] Algorithm flow:

[0127] 1. Initialization: The cloud-based federated learning center distributes the initial global model parameters. To all edge nodes;

[0128] 2. Local Training: Each node trains a local model using desensitized monitoring data and calculates the gradient g. i Simultaneously, the self-calibration unit outputs a drift correction coefficient w. i ;

[0129] 3. Gradient upload: Each node uploads g i w i and sample size n i Up to the cloud;

[0130] 4. Weighted aggregation: The global gradient G is calculated in the cloud according to the above formula, and the global model parameters are updated. (η is the learning rate);

[0131] 5. Model Deployment: The updated global model parameters are deployed from the cloud to each node, and the process is repeated iteratively until the model converges.

[0132] 2. Specific calculation basis for weights:

[0133] Gradient weight w i The dynamic calculation based on the sensor drift correction coefficient follows these steps:

[0134] 1. Drift correction coefficient acquisition: Each UAV's self-calibration unit triggers calibration every 2 hours, calculating the deviation rate between the sensor's measured values ​​and the standard gas values. ;

[0135] 2. Weight Mapping: The bias rate is converted into gradient weights using a non-linear function, as shown in the formula:

[0136] ;

[0137] Where k=5 (determined through experimental fitting to ensure w when the deviation rate is 5%) i ≈0.78, when the deviation rate is 20%, w i ≈0.37);

[0138] 3. Weight normalization: If there exists a node w i If the value is ≤0.1 (significant sensor drift), set its weight to 0.1 (to avoid discarding data completely), and then normalize the weights of all nodes to ensure that the total weight is reasonable.

[0139] 3. Technical Principles and Advantages:

[0140] Main problem: Traditional FedAvg applies equal weight to the gradients of all nodes. If some drone sensors drift significantly (poor data quality), their gradients will interfere with the global model optimization, leading to a decrease in prediction accuracy.

[0141] Technical principle: Drift correction coefficient Directly reflects data quality — The smaller the value, the more reliable the data, and the corresponding gradient weight w. i The larger the value, the greater the contribution of high-quality data to the global model;

[0142] Advantages: No additional hardware costs are required. The negative impact of low-quality data can be suppressed through algorithm optimization, while retaining the effective information of all nodes. It is especially suitable for scenarios with inconsistent sensor states in multi-drone collaborative monitoring.

[0143] 4. Supporting experimental data:

[0144] A comparative experiment was conducted on three drones (node ​​A: drift rate 2%, node B: drift rate 8%, node C: drift rate 15%) in a chemical industrial park. The results are as follows:

[0145] index Standard FedAvg algorithm Improved Federated Learning (Gradient Weighting) Improvement effect Model convergence iterations 25 rounds 18 rounds Convergence speed improved by 28% VOCs concentration prediction accuracy 82.3% 91.7% Accuracy improved by 9.4%. Abnormal emission identification false alarm rate 11.5% 4.2% False alarm rate reduced by 63.5%

[0146] Experimental results show that the improved algorithm effectively reduces the interference of low-quality data by weighting the gradients of high-quality nodes, enabling the model to converge faster and significantly improving prediction accuracy.

[0147] The reproducibility details of the multimodal AI model are as follows:

[0148] 1. The input data modality is clearly defined.

[0149] The input to the multimodal AI model consists of three core modalities, all extracted using a 5-minute sliding window, as detailed below:

[0150] Modal 1: VOCs concentration time series data (10 dimensions) - real-time concentration values, mean, variance, peak value, minimum value, maximum value, rate of change, cumulative increment, fluctuation frequency, and stationarity index of benzene series compounds, alkanes, and alkenes;

[0151] Mode 2: Environmental parameter data (5 dimensions) – real-time values, average values, and changes in temperature, humidity, and air pressure;

[0152] Modal 3: Spatial location data (3D) - GPS coordinates (longitude, latitude), UAV flight altitude.

[0153] 2. Detailed Network Architecture Design (GRU and Transformer Integration)

[0154] The model adopts a three-layer architecture of "modal feature extraction - cross-modal fusion - temporal prediction", as follows:

[0155] (1) Modal feature extraction layer:

[0156] Concentration time series: A 2-layer GRU network (64 hidden units, dropout=0.2) was used to extract time-dependent features and output a 64-dimensional feature vector;

[0157] Environmental parameters: A single fully connected network (5-dimensional input → 32-dimensional output, ReLU activation function) is used to extract environmental impact features;

[0158] Spatial location data: A single convolutional layer (kernel size 3, output 32 dimensions) + pooling layer are used to extract spatial distribution features;

[0159] (2) Cross-modal fusion layer:

[0160] The feature vectors of the three modalities (64+32+32=128 dimensions) are concatenated and input into the Transformer encoder (4 layers, 8 attention heads, 128 hidden layer dimensions). The self-attention mechanism captures the correlation between modalities (such as the correlation between temperature and concentration changes) and outputs 128-dimensional fused features.

[0161] (3) Time series prediction layer:

[0162] A 1-layer GRU network (32 hidden units) + 2-layer fully connected network (128-dimensional → 64-dimensional → 1-dimensional) is used to output the predicted VOCs concentration values ​​for the next 1 / 3 / 6 hours.

[0163] Anomaly detection branch: After feature fusion, a fully connected network layer (128-dimensional → 2-dimensional, activation function Softmax) is connected in parallel to output the "normal / abnormal" classification result.

[0164] 3. Training dataset construction method

[0165] (1) Data source: Monitoring data of a chemical industrial park for one month, including 1Hz sampling data from 3 drones (a total of about 26,000 valid samples).

[0166] (2) Data labeling:

[0167] (2.1) Prediction task: The average VOCs concentration for the next hour is used as the label;

[0168] (2.2) Anomaly identification task: According to the national standard GB37822-2019, the samples with excessive concentrations are marked as "abnormal" and the rest are marked as "normal".

[0169] (3) Data partitioning: The data was divided into training set, validation set and test set in a ratio of 7:2:1. The sliding window method was used to expand the sample (window step size 5 minutes).

[0170] (4) Data preprocessing: All input data are mapped to the [0,1] interval using max-min normalization, and missing values ​​are filled by linear interpolation.

[0171] 4. Criteria for Selecting Key Hyperparameters

[0172] Hyperparameters Value Selection Criteria GRU hidden unit count 64 Multiple experiments have verified that: a value less than 64 indicates insufficient feature extraction, while a value greater than 64 indicates model overfitting; a value of 64 strikes a balance between accuracy and efficiency. Transformer layer number 4 floors The fusion effect is insufficient with 3 layers or less, the computational cost increases dramatically with 5 layers or more, and 4 layers can fully capture cross-modal correlations. Learning rate 0.001 A cosine annealing strategy is adopted, with an initial learning rate of 0.001 to avoid initial oscillations, which are then gradually reduced in the later stages to achieve stable convergence. batchsize 32 Adapted to the computing power of the Jetson Nano edge computing unit, achieving an optimal balance between 32-bit GPU memory and training efficiency. Number of training iterations 100 rounds The model converged after 80 epochs. An additional 20 epochs were trained to verify its generalization ability, and no overfitting was observed.

[0173] The following is a specific implementation example of an AI-based drone VOCs monitoring data management system:

[0174] I. Technical Solution

[0175] This embodiment focuses on VOCs monitoring in industrial parks and constructs a complete AI-based drone-based VOCs monitoring data management system, which includes the following components:

[0176] 1. Unmanned Aerial Vehicle (UAV) Platform

[0177] It employs a hexacopter drone, equipped with flight control components (including an IMU inertial measurement unit) and positioning components (GPS+RTK combined positioning, with a positioning accuracy within 1 meter). The fuselage is equipped with a shock-absorbing gimbal to stabilize the VOCs monitoring module and prevent interference from flight vibrations on sensor detection. The drone has a full-load endurance of 55 minutes (including sensors, edge computing unit, communication module, and hydrogen tank), and supports mid-flight battery swapping (battery swapping time ≤ 3 minutes) or fixed-point hovering refueling.

[0178] 2. VOCs monitoring module

[0179] Sensor combination: Integrates a PID photoionization sensor (detection range 0-200ppm) and a miniature FID flame ionization sensor (detection range 0-1000ppm), and achieves component differentiation through "differentiated response characteristics + data fusion": the PID has a higher sensitivity to benzene series compounds (detection coefficient ≥0.95) than alkanes, and the FID has a better linearity to alkanes (R²≥0.98) than benzene series compounds. At the edge end, category differentiation is achieved by combining the peak response ratio, response time difference and feature library (identification accuracy ≥92%). It can simultaneously collect concentration data of benzene series compounds and alkanes VOCs.

[0180] Environmental parameter sensors: DHT22 temperature and humidity sensor (measurement range -40~80℃, 20%~95%RH) and BMP388 barometric pressure sensor (measurement range 300~1100hPa).

[0181] Miniature FID hydrogen supply: Integrated miniature safety hydrogen cylinder (capacity 50mL, pressure 0.5MPa, weight ≤150g), hydrogen consumption ≤1mL / min, automatic leak detection before flight, alarm triggered and switch to PID single sensor mode when the pressure is lower than 0.1MPa.

[0182] Self-calibration unit: Built-in 100ppm standard isobutylene miniature gas cylinder (weight ≤300g), automatically calibrates once on the ground before takeoff, records the deviation δ0, and if δ0>±3%, it is linearly corrected; during flight, it hovers and calibrates once every 30 minutes (time ≤10 seconds), calculates the deviation δ1, and if δ1>±5%, it is corrected by proportional integral algorithm.

[0183] 3. Edge intelligent processing module

[0184] Employing the Jetson Nano edge computing unit (4-core ARM CPU, 128-core GPU), it integrates the following functional units:

[0185] Data preprocessing unit: Moving average filtering (window length 10 sampling points) and wavelet denoising (using db4 wavelet basis, decomposed into 3 levels) are implemented using Python scripts; the calibration model adopts a 9-parameter quadratic polynomial, and the compensation coefficients are obtained by fitting standard gas experiments (a1=0.002, a2=-0.001, a3=0.0005, a4=0.0001, a5=0.00005, a6=-0.00001, a7=-0.0002, a8=0.0001, a9=0.00005).

[0186] Feature extraction unit: Extracts the mean, variance, and peak value (time domain) within a 5-minute sliding window, the first 8 low-frequency coefficients after Fourier transform (frequency domain), and the energy entropy of the 4 frequency bands of wavelet transform (time-frequency joint), forming a 20-dimensional feature vector;

[0187] AI Analysis Unit: Deploys a multimodal AI model, uses a GRU network (2 layers, 64 hidden units) to fuse features, and a Transformer encoder (4 layers, 8 attention heads) to perform time-series prediction, outputting the concentration trend for the next 1 / 3 / 6 hours; identifies abnormal emissions through the isolated forest algorithm (100 trees).

[0188] 4. Communication transmission module

[0189] Drone terminal: integrates a 5G module (supports SA / NSA dual mode, enables URLLC mode) and a Beidou short message module (communication rate 1200bps).

[0190] Ground-based deployment: Deploy LoRa gateways (communication range 3km) and satellite receiving terminals;

[0191] Transmission configuration: Low-bandwidth data (control commands, status data) is serialized using Protobuf and transmitted via LoRa protocol; large data streams (raw sensor data, preprocessed data) are serialized using Protobuf and compressed using LZ4 and transmitted via 5G URLLC protocol; red alert information is serialized using Protobuf and transmitted via BeiDou short message + 5G SMS dual-link transmission to ensure transmission reliability and efficiency in different scenarios.

[0192] 5. The cloud-based data analysis and comparison module calculates the compensation factor based on the equation.

[0193] Deployed on an Alibaba Cloud server (8-core CPU, 16GB memory), including:

[0194] Data storage units: MySQL database stores metadata (sensor calibration records, enterprise emission inventories), and InfluxDB time-series database stores monitoring data (sampling frequency 1Hz);

[0195] Data Analysis Unit: The concentration heat map of a 50m×50m grid is generated by Kriging interpolation, and the diffusion path is simulated by coupling HYSPLIT and CFD models. The "Standard for the Control of Unorganized Emissions of Volatile Organic Compounds" (GB37822-2019) is built in as the basis for judging the exceedance.

[0196] Data visualization unit: Web-based GIS map interface development, real-time display of drone trajectory and concentration heat map (color gradient: green <20ppm, yellow 20-50ppm, red >50ppm), supports querying 72-hour concentration curves for any point;

[0197] Multi-drone collaborative scheduling unit: When deploying 3 drones, the industrial park is divided into high / medium / low risk zones based on historical emission data. The high-risk zone (such as the area around a chemical workshop) is assigned a grid density of 20m×20m, the medium-risk zone is 50m×50m, and the low-risk zone is 100m×100m. The "monitoring-return to home for resupply" sequence is dynamically allocated based on the remaining flight time of each drone (threshold 15 minutes), and a round-robin resupply mode is adopted to avoid monitoring interruption.

[0198] Federated Learning Center Unit: The model update is triggered every Sunday morning, aggregating the local model gradients of 5 monitoring sites, and updating the global model parameters using the FedAvg algorithm after weighting the signal-to-noise ratio of the data from each site.

[0199] Anomaly Source Tracing Unit: After the warning is triggered, it automatically extracts data from the previous and next hours, calls real-time wind field data from the meteorological station, and uses the HYSPLIT+CFD coupled model to invert the diffusion path, locks the potential emission source area (error ≤ 50m), and matches the VOCs component characteristics in the enterprise's emission inventory (e.g., a chemical plant mainly emits toluene, and the proportion of benzene series compounds in the corresponding monitoring data is > 70%).

[0200] 6. Intelligent Early Warning Module

[0201] Warning level determination: Blue warning (concentration exceeds the standard by less than 10% or exceeds the standard 3 hours after prediction), Yellow warning (concentration exceeds the standard by 10%-50% or slight abnormal emissions are detected), Red warning (concentration exceeds the standard by more than 50% or sudden abnormal emissions).

[0202] Push method: Warning information is sent to the mobile phones of supervisors via SMS interface, and pop-up prompts are displayed simultaneously on the web and APP. Red warnings are accompanied by on-site handling suggestions (such as closing valves and starting exhaust gas treatment devices).

[0203] II. Working Principle

[0204] The system workflow revolves around a closed loop: data acquisition, edge processing, cloud analysis, early warning and source tracing, and model optimization.

[0205] 1. Data acquisition phase: The UAV flies along a preset route (100 meters altitude, 5 m / s speed). The VOCs sensor collects concentration data once per second, the environmental sensor records temperature, humidity and air pressure simultaneously, the positioning component outputs latitude and longitude information in real time, and all raw data is transmitted to the edge computing unit via serial port.

[0206] 2. Edge processing stage: After receiving the data, the edge intelligent processing module first smooths the fluctuations caused by airflow disturbances through moving average filtering, and then removes high-frequency noise through wavelet denoising; combined with the temperature, humidity and air pressure data at the same time, it corrects the deviation caused by environmental interference through a 9-parameter calibration model; after extracting multi-dimensional features, it inputs them into the AI ​​model to predict the concentration trend and determine whether there are abnormal emissions (such as the concentration rising above the baseline value within 30 seconds). The processing results are packaged and uploaded through the communication module.

[0207] 3. Cloud-based analysis phase: After receiving the data, the cloud-based data analysis and comparison module stores the data in layers according to the original data, preprocessed data, and prediction results; the data analysis unit merges the data from multiple drones to generate a global concentration heat map, and compares it with national standards to determine whether it exceeds the standard; the data visualization unit transforms the results into intuitive charts for regulatory personnel to view.

[0208] 4. Early warning and source tracing stage: When an excessive or abnormal event occurs, the intelligent early warning module generates and pushes a graded early warning based on the severity of the event; the abnormal source tracing unit is activated simultaneously, and combines meteorological data with a coupled model to invert the pollution source, match the enterprise's emission characteristics, locate the suspected pollution source, and attach it to the early warning information.

[0209] 5. Model Optimization Phase: The local model gradients trained on anonymized data from each site are summarized weekly. A global model is generated by weighted aggregation based on the data signal-to-noise ratio through federated learning. This model is then distributed to each drone edge unit to update the AI ​​analysis model and improve its adaptability to complex scenarios.

[0210] III. Technical Effects

[0211] 1. Improved data quality: The combination of PID and miniature FID sensors covers multiple categories of VOCs, avoiding blind spots in single sensor monitoring; the self-calibration unit combines ground and air calibration with a 9-parameter compensation model to reduce the impact of temperature, humidity, air pressure and sensor drift, ensuring data consistency under different environmental conditions, with measurement deviation stabilized within 2% after calibration.

[0212] 2. Enhanced real-time performance: Edge computing units complete data preprocessing and AI analysis locally, reducing the amount of raw data transmission. Combined with the 5G URLLC protocol, end-to-end latency is ≤150ms. Abnormal event identification and short-term prediction are completed at the edge, improving the early warning response speed by more than 60% compared to the traditional cloud processing mode.

[0213] 3. Monitoring coverage optimization: Multi-drone collaborative scheduling dynamically allocates monitoring grids according to risk level, and high-risk areas receive denser monitoring. The high payload design and fault compensation mechanism of the hexacopter UAV ensure that there are no gaps in monitoring. Compared with single UAV monitoring, it achieves continuous and uniform coverage of a large area, and the coverage efficiency is improved by 2.5 times.

[0214] 4. Enhanced Decision Support: The cloud-based data analysis and comparison module integrates multi-source data and presents pollution distribution and trends through spatial interpolation and HYSPLIT+CFD coupled diffusion simulation; the visualization interface transforms abstract data into intuitive charts, and combined with tiered early warning and anomaly tracing, it provides regulatory personnel with full-chain information support from problem discovery to source location, improving law enforcement efficiency by more than 80%.

[0215] 5. Improved Model Adaptability: Under the premise of protecting data privacy, the federated learning framework optimizes the AI ​​model by aggregating multi-region monitoring experience based on the signal-to-noise ratio of the data. This enables the AI ​​analysis unit to maintain stable prediction and anomaly identification capabilities in new monitoring scenarios (such as different types of industrial parks), with a prediction accuracy of ≥91% and a false alarm rate of ≤4.2%.

[0216] IV. Experimental Data

[0217] A comparative experiment was conducted in a chemical industrial park (area 5 km²), deploying three UAVs from this system and traditional fixed-site monitoring (8 sites) in parallel operation for one month. The results showed:

[0218] 1. Coverage: The system can cover 95% of the park area, including workshop blind spots and tank area perimeters that cannot be reached by fixed stations, while fixed stations can only cover 30% of the core area.

[0219] 2. Abnormal response: A total of 12 abnormal emission events were detected. The system issued an early warning within an average of 2 minutes after the event occurred. In 10 of these events, the emission companies were accurately identified through the abnormal source tracing unit. Traditional methods require manual investigation and take an average of more than 3 hours.

[0220] 3. Data consistency: In comparisons at the same locations, the deviation between the calibrated monitoring data and the laboratory sampling and analysis results is ≤2%, which is 85% lower than the deviation of the uncalibrated raw data, and the measurement deviation remains stable at different time periods (20℃ diurnal temperature difference).

[0221] 4. Multi-drone collaboration effect: When three drones work together, the effective daily monitoring time reaches 2.5 times that of a traditional single drone, the data spatial resolution (50m grid) is 4 times higher than that of a single drone, and when one drone suddenly fails, the system automatically assigns tasks without monitoring interruption.

[0222] 5. Model optimization results: After 4 weeks of federated learning updates, the AI ​​model reduced the concentration prediction bias under complex operating conditions (such as rainy days and nights) by 35% compared with the initial model, and the false alarm rate of abnormal emission identification decreased from 11.5% to 4.2%.

[0223] Experimental Results Recording Table

[0224] Test metrics This application technology Traditional technology (CN111781030B) Pure cloud computing architecture Accuracy of pollution source tracing (average) 95.3% (error ≤ 5m) 82.1% (error ≤ 20m) 88.7% (error ≤ 15m) Path planning efficiency (covering 100,000 square meters) Flight time: 42 minutes; Energy consumption: 68Wh Flight time: 75 minutes; Energy consumption: 112Wh Flight time: 63 minutes; Energy consumption: 95Wh Data processing latency (end-to-end) 89ms 320ms 187ms Sampling error at monitoring points (ppm) ±0.03ppm ±0.12ppm ±0.08ppm Battery life (fully loaded) 105 minutes 82 minutes 91 minutes

[0225] Experimental results show that this system can effectively solve the problems of limited coverage, delayed response, and insufficient data reliability of traditional methods in VOCs monitoring in industrial parks. It provides more comprehensive, timely, and accurate technical support for environmental supervision. It is significantly better than existing technologies in core indicators such as source tracing accuracy, path efficiency, and delay control, verifying the superiority and practicality of the technical solution.

[0226] How to use

[0227] The use of this system needs to cover the entire lifecycle of deployment, planning, execution, analysis, response, and maintenance. The specific steps are as follows:

[0228] 1. System Deployment: Hardware Setup and Software Initialization

[0229] Hardware configuration: Assemble a hexacopter drone (equipped with VOCs sensors, edge computing units, and communication modules), deploy ground LoRa gateways and satellite receiving terminals around the monitoring area, and install a data management platform (including database and visualization interface) on a cloud server.

[0230] Software initialization: Deploy preprocessing scripts and initial parameters of AI models in the edge computing unit, configure communication protocols (Protobuf+MQTT / 5GURLLC) and VOCs emission standard thresholds in the cloud, and install and bind an early warning APP on the mobile phones of regulatory personnel.

[0231] 2. Task Planning: Set up monitoring plans as needed.

[0232] Regulatory personnel can delineate monitoring areas (such as a chemical industrial park) through a cloud-based visual interface. The system automatically divides the area into high / medium / low-risk zones based on historical emission data (high-risk zones include the area around workshops, and low-risk zones include the park boundary). Then, based on the drone's endurance and the grid density of the risk zone (20m×20m for high-risk zones and 100m×100m for low-risk zones), the system automatically generates a multi-drone collaborative flight path. The path nodes can also be manually adjusted (e.g., focusing on a specific discharge outlet).

[0233] 3. Data Acquisition and Execution: Autonomous Monitoring by UAVs

[0234] Before starting the drone, the self-calibration unit is ground-calibrated. After starting, the drone flies autonomously along the planned path. The VOCs monitoring module collects data in real time, and the edge intelligent processing module completes preprocessing and AI analysis simultaneously. During flight, the drone transmits the preprocessed data, AI prediction results, and its own status (battery level, location) back to the cloud through the communication module. If the battery level is below 20%, the system automatically triggers a return-to-home command to avoid losing contact.

[0235] 4. Data Processing and Analysis: Cloud Integration and Visualization

[0236] After receiving data in the cloud, it automatically completes hierarchical storage and multi-machine data fusion to generate a global concentration heat map and historical trend curves for single points (such as 72-hour concentration changes). The data analysis unit combines the HYSPLIT+CFD coupled model to simulate the possible direction and range of pollution diffusion and updates the results to the visualization interface in real time. Supervisors can view monitoring data at any time and location via computer or mobile phone, and can also filter the distribution of specific VOCs components (such as benzene series compounds).

[0237] 5. Early warning response: tiered handling and source tracing

[0238] Warning reception: When the system issues a warning, the regulatory personnel will receive a pop-up window on their mobile APP and a text message containing the warning level, the location of the incident, a comparison of the current concentration with the standard value, and the suspected emission enterprise;

[0239] On-site response: For a blue alert, personnel can be assigned to track the trend online; for a yellow alert, personnel can be dispatched to the site for inspection; for a red alert, an emergency response should be initiated immediately (such as notifying the company to suspend production and investigating the leak point).

[0240] Source tracing and investigation: Based on the list of suspicious enterprises in the early warning information, regulatory personnel can carry portable detectors to the site for verification, and take legal action after confirming the emission source.

[0241] 6. Model updates and maintenance: Ensuring long-term system stability

[0242] Model update: The system automatically triggers the federated learning process every week, without manual intervention, to update the AI ​​model parameters;

[0243] Equipment maintenance: Regularly check the status of the drone's battery and sensors, replace the miniature calibration gas canister of the self-calibration unit monthly, and verify the signal strength of the communication module quarterly to ensure that all components are working properly;

[0244] Data maintenance: Regularly back up the cloud database and clean up raw data older than one year (retain pre-processed data and analysis results) to avoid excessive storage pressure.

[0245] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based drone VOCs monitoring data management system, characterized in that, It includes a drone platform, a VOCs monitoring module, an edge intelligent processing module, a communication transmission module, a cloud data analysis and comparison module, and an intelligent early warning module, among which: The drone platform is used to carry the VOCs monitoring module, the edge intelligent processing module, and its own flight control and positioning components. The positioning components are used to obtain the real-time location information of the drone platform. The signal output terminal of the VOCs monitoring module is connected to the signal input terminal of the edge intelligent processing module, which is used to collect VOCs concentration data and environmental parameter data in the target area and transmit the collected data to the edge intelligent processing module. The signal output terminal of the edge intelligent processing module is connected to the first signal terminal of the communication transmission module. It is used to process, extract features and perform AI analysis on the data transmitted by the VOCs monitoring module. Based on a preset lightweight AI model, it performs real-time anomaly screening on the data and only uploads the original sequence of data segments marked as abnormal or with significant trend changes to the communication transmission module to generate preprocessed data, feature data and AI analysis results. The second signal terminal of the communication transmission module communicates bidirectionally with the first signal terminal of the cloud data analysis and comparison module to realize the transmission of raw data, preprocessed data, AI analysis results and control commands between the UAV platform and the cloud data analysis and comparison module. The cloud-based data analysis and comparison module has built-in national or local VOCs emission standards. Its second signal terminal is connected to the signal input terminal of the intelligent early warning module. It is used to store, manage and analyze the data transmitted by the communication transmission module, and compare the analyzed VOCs concentration data with the built-in VOCs emission standards to determine whether the emission standards are exceeded. The intelligent early warning module is used to generate early warning information and push it to designated objects based on the exceedance judgment results of the cloud data analysis and comparison module and the AI ​​analysis results; The VOCs monitoring module includes at least two types of VOCs sensors, an environmental parameter sensor, and a self-calibration unit. The VOCs sensors include at least two of the following: a PID photoionization sensor, a miniature FID flame ionization sensor, and a MOS metal oxide semiconductor sensor, used to collect VOCs concentration data for benzene series compounds, alkanes, and alkenes within the target area. The environmental parameter sensor includes a temperature and humidity sensor and a pressure sensor, used to collect real-time temperature, humidity, and pressure data for the target area. The self-calibration unit has a built-in miniature standard VOCs gas storage component with a replaceable gas tank design, its capacity adapted to the UAV payload. It automatically triggers the calibration process during the ground preparation phase before UAV takeoff. If the sensor data drift exceeds a preset threshold of ±5% during flight, in-flight calibration is triggered during UAV hovering operations. By comparing the measured values ​​of the VOCs sensor for standard VOCs gas with the standard values, the drift error of the VOCs sensor is corrected. The edge intelligent processing module includes a data preprocessing unit, a feature extraction unit, and an AI analysis unit. The signal input of the data preprocessing unit is connected to the signal output of the VOCs monitoring module, and is used to perform real-time noise reduction, accurate calibration, and normalization processing on the raw VOCs concentration data. The signal input of the feature extraction unit is connected to the signal output of the data preprocessing unit, and the signal output is connected to the signal input of the AI ​​analysis unit. It is used to extract time-domain features, frequency-domain features, and time-frequency joint features from the preprocessed data. The time-domain features include mean, variance, and peak value. The frequency-domain features include Fourier transform coefficients. The time-frequency joint features include wavelet transform energy entropy. The AI ​​analysis unit has a built-in multimodal AI model, which is used to predict the VOCs concentration change trend within a preset time period based on the extracted features, and to identify abnormal VOCs emission events.

2. The AI-based UAV VOCs monitoring data management system according to claim 1, characterized in that, The denoising process of the data preprocessing unit includes sequentially performing moving average filtering and wavelet denoising on the original VOCs concentration data. The window length of the moving average filtering is set according to the flight speed of the UAV platform. The calibration process of the data preprocessing unit adopts a compensation model based on environmental parameters. The compensation model uses temperature, humidity, and air pressure data collected by environmental parameter sensors as input variables and includes linear, quadratic, and interactive terms. The specific formula is as follows: ； Where: C cal The calibrated VOCs concentration value; C raw The original VOCs concentration value after noise reduction, in ppm or ppb; T, H, and P are the real-time collected temperature, humidity, and air pressure values, respectively, from temperature, humidity, and air pressure sensors; T0, H0, and P0 are the temperature, humidity, and air pressure values ​​under standard environmental parameter conditions; a1 to a9 are the compensation coefficients obtained by fitting through standard gas experiments; The normalization process of the data preprocessing unit adopts the maximum-minimum normalization method to map the calibrated VOCs concentration data to a preset numerical range.

3. The AI-based UAV VOCs monitoring data management system according to claim 1, characterized in that, The communication transmission module includes a UAV-side communication component and a ground-side communication component. The UAV-side communication component is integrated into the UAV platform and includes a 5G communication unit and a BeiDou short message communication unit. The ground-side communication component is connected to the cloud-based data analysis and comparison module and includes a LoRa communication unit and a satellite communication unit. The UAV-side communication component and the ground-side communication component transmit data using a combination of data serialization and transmission protocols. Control commands are serialized using Protobuf and transmitted via the MQTT protocol. Raw and pre-processed data are serialized using Protobuf and transmitted via the 5GURLLC protocol. Emergency warning information generated by the intelligent warning module is serialized using Protobuf and transmitted via the BeiDou short message protocol.

4. The AI-based UAV VOCs monitoring data management system according to claim 1, characterized in that, The cloud-based data analysis and comparison module includes a data storage unit, a data analysis unit, and a data visualization unit. The data storage unit employs a combined relational and time-series database architecture to hierarchically store raw data, preprocessed data, AI analysis results, and metadata. The metadata includes sensor calibration records, AI model version information, and national or local VOCs emission standards. The signal input of the data analysis unit is connected to the signal output of the data storage unit. It performs statistical analysis, spatial interpolation analysis, and VOCs diffusion model coupling analysis on the stored data. Simultaneously, it compares the analyzed VOCs concentration data with the VOCs emission standards in the metadata and outputs the exceedance judgment result. The signal input of the data visualization unit is connected to the signal output of the data analysis unit. It displays the exceedance judgment result in the form of a GIS map heatmap, the VOCs concentration change trend in the form of a time trend curve, and the warning information in the form of a warning list.

5. The AI-based UAV VOCs monitoring data management system according to claim 1, characterized in that, The intelligent early warning module includes an early warning level determination unit and an early warning push unit. The signal input terminal of the early warning level determination unit is connected to the second signal terminal of the cloud data analysis and comparison module, and is used to generate graded early warning signals based on the exceedance judgment results of the cloud data analysis and comparison module and the abnormal emission events identified by the AI ​​analysis unit, combined with preset rules. The graded early warning signals include blue, yellow and red warnings. The signal input terminal of the early warning push unit is connected to the signal output terminal of the early warning level determination unit, and is used to push the graded early warning signals to the terminal devices of environmental regulators via SMS and mobile applications.

6. The AI-based UAV VOCs monitoring data management system according to claim 1, characterized in that, The drone platform supports collaborative operation of multiple drones, employing a hexacopter drone model. The cloud-based data analysis and comparison module further includes a multi-drone collaborative scheduling unit. The first signal terminal of the multi-drone collaborative scheduling unit is connected to the second signal terminal of the communication transmission module, used to receive real-time status data from each drone platform. The real-time status data includes remaining flight time, current location, and the amount of data collected. The second signal terminal of the multi-drone collaborative scheduling unit is connected to the flight control components of each drone platform, used to dynamically allocate monitoring grids based on the VOCs risk level of the monitoring area and the real-time status data of each drone platform. The multi-drone collaborative scheduling unit is also used to fuse the monitoring data of the same time stamps and adjacent locations transmitted by multiple drone platforms using a Kalman filter algorithm to eliminate spatial measurement errors. When a single drone platform fails and loses connection, the monitoring task of the failed drone platform is automatically assigned to other online drone platforms.

7. The AI-based UAV VOCs monitoring data management system according to claim 1, characterized in that, The edge intelligence processing module and the cloud data analysis and comparison module achieve collaborative optimization of AI model parameters through a federated learning framework. The cloud data analysis and comparison module also includes a federated learning center unit. The first signal terminal of the federated learning center unit is connected to the edge intelligence processing module of each UAV platform and is used to receive local AI model gradient data uploaded by each edge intelligence processing module. The local AI model gradient data is generated based on the desensitized data collected by each UAV platform. The second signal terminal of the federated learning center unit is connected to each edge intelligence processing module and is used to weight the local model gradient according to the actual data signal-to-noise ratio after self-calibration of each UAV VOCs monitoring module. The higher the signal-to-noise ratio, the greater the weight. The FedAvg algorithm is used to aggregate the local AI model gradient data to generate global AI model parameters, and the global AI model parameters are sent to each edge intelligence processing module to update the multimodal AI model in the AI ​​analysis unit.

8. The AI-based UAV VOCs monitoring data management system according to any one of claims 1-7, characterized in that, The cloud-based data analysis and comparison module also includes an anomaly tracing unit; the signal input end of the anomaly tracing unit is connected to the signal output end of the intelligent early warning module, and is used to extract the timestamp of the early warning event and the monitoring data within a preset time period before and after when the intelligent early warning module triggers the early warning signal; the anomaly tracing unit is also used to calculate the VOCs diffusion path corresponding to the early warning event by combining the VOCs meteorological inversion model, and to lock the potential emission source area; The signal output terminal of the anomaly tracing unit is connected to the data storage unit. It is used to call the historical enterprise emission list of potential emission source areas stored in the data storage unit, match the VOCs component characteristics in the monitoring data with the VOCs component information in the enterprise emission list, locate the suspicious emission enterprises, and push the information of the suspicious emission enterprises and the early warning signal simultaneously.

Images