Sensor-based unmanned aerial vehicle intelligent inspection monitoring method and system

Abstract

The application discloses the technical field of inspection monitoring, and discloses an unmanned aerial vehicle intelligent inspection monitoring method and system based on a sensor, which comprises the following steps: acquiring an initial inspection data set; obtaining a feature data set; generating a unified target feature data set; performing abnormality detection on the target feature data set to acquire an abnormal inspection area data set; performing safety level division on the abnormal inspection area to obtain a division result; performing risk screening on a safety risk area in the division result to obtain a plurality of high-risk point types; when the change threshold of one of the high-risk point types reaches a preset threshold, a high-risk occurrence zone is preliminarily determined; the unmanned aerial vehicle is controlled to perform fixed-point hovering to perform key monitoring and refined inspection on the high-risk occurrence zone, and a secondary inspection data set is acquired; an analysis result is obtained; it is confirmed again that the current inspection area is in the high-risk occurrence zone, and corresponding emergency response early warning strategies and regulation and control strategies are generated, so that the inspection area can be monitored in real time and accurately.

Description

Technical Field

[0001] This invention relates to the field of inspection and monitoring technology, and more specifically, to a sensor-based intelligent inspection and monitoring method and system for unmanned aerial vehicles (UAVs). Background Technology

[0002] As a crucial technology for modern industrial monitoring, unmanned aerial vehicle (UAV) intelligent inspection directly impacts inspection efficiency and safety assurance levels through its autonomous perception, intelligent decision-making, and precise early warning capabilities in complex environments. Traditional inspection and monitoring practices primarily rely on periodic manual inspections, single-sensor data collection, or simple fixed-threshold alarm mechanisms. These methods fail to achieve dynamic and precise monitoring and assessment of abnormal states in the inspected area. They struggle to uncover the spatiotemporal correlations and anomaly evolution patterns among multi-dimensional feature data; anomaly detection mechanisms lack intelligent analysis capabilities, hindering accurate division of areas with different safety levels and dynamic screening of high-risk locations; and, more critically, existing technologies have failed to establish a complete closed-loop control process from initial inspection to re-inspection and early warning, resulting in a disconnect between anomaly identification and emergency response, severely limiting the relevance and timeliness of early warning strategies. Consequently, UAV inspection and monitoring remain largely in a passive response state, hindering the shift from reactive to proactive prevention.

[0003] Therefore, how to detect anomalies through features in multi-source data, conduct dynamic security assessments of the anomaly detection results, screen high-risk locations and perform detailed inspections of anomaly areas, and perform in-depth analysis of the detailed inspection data to enable emergency early warning decisions, and achieve real-time accurate monitoring and proactive early warning have become urgent technical problems to be solved. Summary of the Invention

[0004] This invention provides a sensor-based intelligent inspection and monitoring method and system for unmanned aerial vehicles (UAVs), which solves the problem in the prior art that it is difficult to detect anomalies through features in multi-source data, and further performs dynamic security assessment on the anomaly detection results. At the same time, it conducts high-risk point screening and fixed-point fine inspection of anomaly areas, and performs in-depth analysis of fine inspection data to enable emergency early warning decision-making, thereby achieving real-time accurate monitoring and proactive early warning.

[0005] This invention provides a sensor-based intelligent inspection and monitoring method and system for unmanned aerial vehicles (UAVs), comprising:

[0006] Firstly, a sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles (UAVs) includes:

[0007] Collect multi-source data from the inspection area, and perform timestamp marking, coordinate positioning and format standardization to form a structured initial inspection dataset;

[0008] Features of various types of data in the initial inspection dataset are extracted and standardized to form a feature dataset with a unified dimension.

[0009] The feature dataset is cleaned, time-series aligned, spatially calibrated, normalized, and standardized. Correlation analysis between features is established to generate a fused target feature dataset.

[0010] The target feature dataset is input into a pre-trained anomaly detection model to identify abnormal feature data that deviates from the normal baseline pattern and record its spatiotemporal location information to form an anomaly inspection area dataset.

[0011] Based on the abnormal inspection area dataset, and considering the type, intensity, and duration of abnormal features, combined with geographical environmental sensitivity and historical accident data, a comprehensive risk score is calculated for each abnormal area. Based on the score range, the area is divided into four safety levels: low risk, medium risk, high risk, and extremely high risk, and the division results are output.

[0012] From the division results, areas with high-risk and extremely high-risk safety levels are selected; the selected areas are further subdivided into grids, and the hazard index of each grid point is calculated; a hazard index threshold is set, and points exceeding the threshold are identified as high-risk points, and a list of high-risk points is generated; a dynamic change threshold is set for each type of high-risk point in the list.

[0013] When the rate of change of parameters at a high-risk location reaches its corresponding change threshold, the area where the location is located is preliminarily determined to be a high-risk area.

[0014] In response to the preliminary judgment, the drone is controlled to fly to the high-risk area and hover at fixed points according to the preset multi-dimensional inspection path at different flight altitudes, observation positions and time nodes to collect high-resolution fine data. The key points are repeatedly observed to obtain time-series data and integrated to form a secondary inspection dataset.

[0015] The secondary inspection dataset is analyzed using deep learning models and data analysis algorithms to obtain analytical results that include hazard source type identification, hazard level quantitative assessment, anomaly pattern comparison, and development trend prediction.

[0016] The preliminary judgment is verified based on the analysis results: if the verification is successful, the current area is confirmed as a high-risk area, and corresponding emergency response warning strategies and control strategies are generated based on the high-risk location type; the emergency response warning strategy includes at least the hazard type, level, location information and recommended disposal measures; the control strategy includes start-stop control commands for designated equipment or evacuation range planning.

[0017] Furthermore, the multi-source data collected from the inspection area forms an initial inspection dataset, including:

[0018] The first dataset is obtained by collecting environmental data of the inspection area using multiple sensors mounted on the drone;

[0019] The second dataset is obtained by acquiring geographic information data, meteorological data, and historical inspection records of the inspection area.

[0020] The first and second datasets are timestamped and located using coordinates to obtain the location results.

[0021] The multi-source data items in the marker positioning results are formatted and unified to form the initial inspection dataset;

[0022] Specifically, environmental data is collected synchronously using visible light sensors, infrared sensors, gas sensors, and sound sensors carried on the drone;

[0023] Acquire geographic information data, real-time meteorological data, and historical inspection record data of the inspection area;

[0024] All data sources are timestamped at the millisecond level using a unified time base, and coordinate positioning with centimeter-level accuracy is achieved using GPS or BeiDou positioning systems.

[0025] The labeled heterogeneous data is converted into a standardized structured format to form the initial inspection dataset.

[0026] Further, the step of extracting feature data for each data item in the initial inspection dataset to obtain a feature dataset includes:

[0027] Extract texture, color, and shape features from visible light image data;

[0028] Extract temperature distribution features, thermal anomaly region features, and temperature gradient features from infrared thermal imaging data;

[0029] Extract gas concentration characteristics, concentration change rate characteristics, and gas composition characteristics from gas sensor data;

[0030] Extract spectral features, volume features, and abnormal sound feature data from sound sensor data;

[0031] If the sensor malfunctions or the data quality does not meet the requirements, a prompt message will be returned to the user in a timely manner, and it will be suggested to replace the equipment or adjust the parameters until the data acquisition conditions meet the requirements before the data is transmitted to the preprocessing step.

[0032] The Z-score standardization method is used to convert the extracted feature data into a standard normal distribution with a mean of 0 and a variance of 1, forming a feature dataset with a unified dimension.

[0033] The process involves extracting key features from the transmitted multi-source data using a deep learning model, and then combining this with time series analysis algorithms to uncover the spatiotemporal correlation patterns between these features.

[0034] Furthermore, the preprocessing and spatiotemporal calibration of the feature dataset to generate a unified target feature dataset includes:

[0035] Data cleaning: Linear interpolation was used to handle missing values, wavelet thresholding was used to filter out noise, and the 3σ criterion was used to remove outliers.

[0036] Time series alignment: Using the time of the sensor data with the highest sampling frequency as a reference, other sensor data are aligned to the same time axis using linear interpolation;

[0037] Spatial location calibration: Establish a unified WGS-84 geographic coordinate system and map the spatial data of all sensors to this coordinate system through a coordinate transformation matrix;

[0038] Data normalization: The minimum-maximum scaling method is used for normalization, mapping the data to the [0,1] interval; the Z-score method is used for standardization to eliminate the influence of units.

[0039] Association analysis: Principal component analysis (PCA) is used for dimensionality reduction, and Pearson correlation coefficients are calculated to construct a feature association matrix, generating a fused target feature dataset.

[0040] Furthermore, the step of performing anomaly detection on the target feature dataset to obtain an anomaly inspection region dataset includes:

[0041] Establish an anomaly detection model based on machine learning and train a baseline pattern of feature data under normal conditions;

[0042] Set dynamic thresholds and judgment rules for anomaly detection with different feature parameters;

[0043] The target feature parameter set is input into the trained anomaly detection model for anomaly detection analysis.

[0044] If an abnormal area is found, the spatial topology modeling algorithm is used to assess the security level of the abnormal area, and the anomaly detection results are passed to the security level classification step.

[0045] Identify abnormal feature parameters whose deviations from the baseline pattern exceed the corresponding dynamic threshold in the abnormal detection results;

[0046] If additional data is needed or the boundaries of abnormal areas need to be confirmed, the inspection data should be collected again.

[0047] Record the spatiotemporal location information corresponding to the abnormal feature parameters to form an abnormal inspection area dataset;

[0048] Specifically, a hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM) is trained using feature data from a historical database under safe conditions to learn dynamic baseline patterns under normal operating conditions.

[0049] Set a dynamic anomaly threshold based on statistical distribution; for example, for temperature features, the threshold is set to ±2.5 times the standard deviation of the baseline value.

[0050] Input the real-time target feature dataset into the model and calculate its reconstruction error or deviation score compared with the baseline model;

[0051] When the deviation score exceeds the preset threshold, it is judged as abnormal, and the precise timestamp and geographic coordinates corresponding to the abnormal feature data are recorded.

[0052] Furthermore, the step of classifying the abnormal inspection areas into security levels based on the abnormal inspection area dataset to obtain the classification results includes:

[0053] Establish a safety level assessment system, setting four levels: low risk, medium risk, high risk, and extremely high risk;

[0054] Risk scores are calculated based on the type, intensity, and duration of anomalous data.

[0055] Risk weights are adjusted based on geographical environmental factors and historical accident records of the inspection area.

[0056] The safety level of abnormal inspection areas is classified according to the comprehensive risk score.

[0057] Generate partitioning results that include region coordinates, security level, and risk characteristic descriptions;

[0058] Specifically, the security level classification includes:

[0059] Set safety level scoring ranges: low risk (0-30 points), medium risk (31-60 points), high risk (61-90 points), and very high risk (91-100 points);

[0060] Example of basic risk score calculation formula: Basic score = (Type weight coefficient × Intensity normalized value × Duration decay factor); where, the type weight coefficient is determined by expert experience method or analytic hierarchy process, the intensity normalized value linearly transforms the original parameters to [0,1], and the duration decay factor considers the persistent effect of the anomaly; the type weight coefficient is determined by analytic hierarchy process (AHP), at least 5 domain experts are invited to compare the relative importance of various anomaly characteristics (such as gas leakage, temperature anomaly, structural deformation) pairwise, construct a judgment matrix and calculate the consistency ratio (CR<0.1) of each weight coefficient;

[0061] Comprehensive risk score = base score × (1 + geographical environment adjustment coefficient + historical accident adjustment coefficient);

[0062] The safety level is determined based on the range of the comprehensive risk score.

[0063] Furthermore, the safety risk areas in the classification results are screened for hazard levels to obtain multiple high-risk locations, including:

[0064] Areas classified as high-risk and extremely high-risk were selected from the classification results;

[0065] The selected safety risk areas are further subdivided into grids to establish refined monitoring points;

[0066] Calculate the risk index for each monitoring point, taking into account the intensity of abnormal parameters, diffusion trends, and scope of impact.

[0067] Set a risk screening threshold to identify monitoring points whose risk index exceeds the threshold;

[0068] The selected high-risk locations are located by coordinates and identified by hazard type to form a list of high-risk locations.

[0069] Specifically, the hazard screening includes: the grid size is dynamically adjusted according to the terrain complexity and sensor accuracy of the inspection area; flat areas can use 10m×10m, and complex areas can use 5m×5m or smaller.

[0070] Example of the hazard index calculation formula: Hazard index = Anomaly parameter intensity weight × Intensity normalized value + Diffusion trend weight × Trend score + Impact range weight × Range score; where the sum of each weight is 1, for example, it can be 0.4, 0.3, 0.3;

[0071] The risk index threshold is determined based on historical data statistics, for example, by selecting the top 10 percentile of the index value among all grid points as the threshold.

[0072] Furthermore, by controlling the drone to hover at different altitudes, orientations, and time points within the high-risk area, key monitoring and refined inspection of the high-risk area are conducted to obtain a secondary inspection dataset, including:

[0073] The transmitted anomaly detection results are used to call the path planning algorithm to generate a detailed inspection path for high-risk locations;

[0074] Based on the list of high-risk locations, a refined inspection path is planned for the drone, including different flight altitudes, observation directions, and hovering time points;

[0075] During the inspection, the flight controller monitors the drone's position and attitude in real time, and the cruise combined with obstacle avoidance algorithms obtains the drone's safe flight path;

[0076] Control the drone to fly to each high-risk location according to the planned path, and perform multi-angle fixed-point hovering;

[0077] If obstacles or external interference prevent the path from continuing, return to the anomaly detection step to reassess the abnormal area or adjust the path planning.

[0078] During hovering, a refined dataset including high-resolution images, precise temperature data, and gas concentration data is acquired using high-precision sensors.

[0079] Repeated observations were performed on key high-risk locations in the refined dataset to obtain time-series variation data.

[0080] After the inspection is completed, the detailed inspection data collected in the second step is transmitted to the preprocessing step.

[0081] Integrate refined data and time-series change data to form a secondary inspection dataset;

[0082] Furthermore, develop multi-dimensional inspection route plans for high-risk locations, setting different flight altitudes, observation azimuths, and hovering times;

[0083] Control the drone to fly to each high-risk location according to the planned path, and conduct multi-angle fixed-point hovering observation at each location;

[0084] During hovering, it continuously acquires high-resolution sensor data, including close-up images, precise temperature, and gas concentration data;

[0085] Repeated observations were conducted at the same high-risk location to obtain time-series variation data.

[0086] Acquire inspection data from multiple different time points and establish a data comparison benchmark library;

[0087] The collected detailed observation data is integrated and labeled to form a secondary inspection dataset for high-risk areas;

[0088] Specifically, the refined inspection includes: for each high-risk location, the inspection path planning sets at least three flight altitude gradients (such as 5m, 10m, 15m), four main observation azimuths (such as 0°, 90°, 180°, 270°), and multiple observation time points (such as 10 minutes apart).

[0089] At the hovering point, data is continuously collected using high-precision sensors (such as high-resolution infrared thermal imagers and laser gas analyzers);

[0090] Repeated observations of the same high-risk location at least three times are conducted to form a dataset that can be used for time series analysis.

[0091] Furthermore, the deep analysis of the secondary inspection dataset yields the following results:

[0092] Establish a deep learning model to intelligently identify and analyze secondary inspection data;

[0093] The data integration algorithm is invoked to clean and classify the inspection data, and the cleaning and classification results are obtained.

[0094] Identify the specific hazard source types and distribution characteristics of high-risk areas in the cleaning and classification process to obtain the identification results;

[0095] The identification results are comprehensively analyzed to assess the risk level of the high-risk areas currently being inspected.

[0096] Conduct hazard comparison analysis on multiple critical inspection areas to identify the hazard characteristics of similar anomalies;

[0097] Set thresholds for the hazard characteristics and duration of similar anomalies;

[0098] When the characteristics of similar anomalies exceed a preset threshold and the duration exceeds a preset time threshold, an alarm mechanism is triggered; and an emergency warning decision report is generated in conjunction with a report generation algorithm.

[0099] The report includes the distribution of abnormal areas, the results of the security level assessment, and recommendations for countermeasures;

[0100] The emergency early warning decision report is sent to the user terminal, and the inspection results are presented on the display terminal.

[0101] If there are missing or incomplete contents in the report, the user will be prompted to supplement the inspection information.

[0102] Based on the emergency early warning decision-making report, the development trend and potential diffusion path of the hazard source are predicted, and a comprehensive analysis result is finally generated, which includes the hazard source identification results, the degree assessment report, the anomaly point comparison analysis and the trend prediction information.

[0103] Specifically, the deep analysis includes: using a pre-trained convolutional neural network (CNN) model (such as YOLOv5) to identify hazards in visible light and infrared images;

[0104] The degree of hazard is assessed by fusing confidence levels of multi-sensor data (such as temperature and gas concentration) using the DS evidence theory.

[0105] The similarity between the current anomaly pattern and historical anomaly patterns is compared using the Dynamic Time Warping (DTW) algorithm.

[0106] Set a similarity threshold (e.g., 0.8) and a duration threshold (e.g., more than 30 minutes consecutively), and trigger an alarm when both are met.

[0107] Time series models such as ARIMA are used to predict the development trend and spread path of hazards over a future period (e.g., 60 minutes).

[0108] Secondly, a sensor-based intelligent drone inspection and monitoring system includes:

[0109] The data acquisition module is used to collect multi-source data from the inspection area to form an initial inspection dataset;

[0110] The feature extraction module is used to extract feature data for each data item in the initial inspection dataset to obtain a feature dataset;

[0111] The data processing module is used to preprocess and spatiotemporally calibrate the feature dataset to generate a unified target feature dataset.

[0112] The anomaly detection module is used to perform anomaly detection on the target feature dataset to obtain an anomaly inspection area dataset; based on the anomaly inspection area dataset, the anomaly inspection area is divided into security levels to obtain the division results.

[0113] The hazard screening module is used to screen the safety risk areas in the division results to obtain multiple high-risk points; obtain multiple high-risk point types and set corresponding hazard point change thresholds. When the change threshold of one of the high-risk points reaches the preset threshold, it is preliminarily determined that the current inspection area is in a high-risk occurrence zone.

[0114] The fine inspection module is used to control drones to fly to different altitudes, orientations, and time points in high-risk areas and hover at fixed points to conduct key monitoring and fine inspection of high-risk areas, and obtain secondary inspection datasets; deep analysis of the secondary inspection datasets is performed to obtain analytical results including hazard source identification, severity assessment, and trend prediction.

[0115] The early warning decision module is used to verify the preliminary judgment based on the analysis results: if the verification is successful, the current inspection area is confirmed as a high-risk area, and an emergency response early warning strategy and control strategy corresponding to the high-risk point type in the area are generated.

[0116] The beneficial effects of this invention are as follows: By collecting multi-source data from the inspection area to form an initial inspection dataset, this invention overcomes the limitations of traditional single-sensor data acquisition. The system simultaneously integrates multi-dimensional heterogeneous data such as visual images, infrared thermal imaging, gas concentration, ambient temperature and humidity, and sound vibration, covering comprehensive environmental perception through simultaneous acquisition by visible light sensors, infrared sensors, gas sensors, and sound sensors. This enables comprehensive monitoring of complex changes in the inspection environment, laying a solid data foundation for subsequent accurate analysis. Compared to traditional solutions, multi-source data fusion, through millisecond-level time synchronization marking technology, effectively reduces the risk of monitoring blind spots caused by the failure of a single sensor.

[0117] By extracting the feature data of each data item in the initial inspection dataset and performing preprocessing and spatiotemporal calibration, unified and standardized processing of multidimensional feature data is achieved. This technical solution employs algorithms such as missing value imputation, Kalman filter noise removal, and box plot outlier removal to effectively uncover the spatiotemporal correlation patterns between temperature gradients, gas concentration change rates, and acoustic signature spectral features. Alignment with GPS reference timestamps to millisecond-level accuracy ensures the comparability of spatial location information collected by different sensors. Compared to traditional simple numerical statistical methods, the intelligent feature extraction technology based on principal component analysis significantly improves the accuracy and reliability of abnormal pattern recognition and reduces the false alarm rate.

[0118] By establishing a machine learning-based anomaly detection model to detect anomalies in target feature datasets and classifying safety levels based on anomaly inspection area datasets, a four-level risk assessment system combining dynamic thresholds and statistical distribution thresholds was constructed. This mechanism employs a long short-term memory network or autoencoder architecture, combined with a dynamic range determined by a temperature parameter threshold of ±2.5℃ and the 3σ principle for gas concentration, and achieves quantitative anomaly assessment through Mahalanobis distance or reconstruction error indices. Compared to traditional fixed-threshold alarm mechanisms, intelligent anomaly detection technology can achieve dynamic threshold adjustment, automatically updating ±1.5℃ to adapt to changes in operating conditions when ambient temperature undergoes seasonal changes.

[0119] By subdividing safety risk areas into grids, screening for hazard levels, and setting change thresholds for monitoring, the system achieves precise location and dynamic monitoring of high-risk points. Employing a 1m×1m monitoring grid division method and calculating a hazard index, the system manages high-risk areas with precision. When the hazard index exceeds a preset threshold, a list of high-risk points is generated. When the gas concentration change rate at a high-risk point reaches 0.5% / second, a preliminary warning is triggered promptly, and a detailed inspection module is activated, effectively preventing sudden deterioration of abnormal conditions.

[0120] By controlling drones to conduct multi-angle, fixed-point, refined inspections at different heights (3m, 5m, and 8m) in high-risk areas, and repeatedly collecting data three times at 30-second intervals at the same location, time-series comparative data is generated. A deep learning model combining convolutional neural networks and recurrent neural networks is used to intelligently analyze the secondary inspection data. The YOLOv5 target detection algorithm is employed to identify hazard types, and multi-source data is fused using DS evidence theory to achieve integrated processing of hazard identification, severity assessment, and trend prediction. The resolution of the fixed-point inspection data is improved to an accuracy range of 0.5-0.8 meters, providing reliable data support for precise early warning decisions.

[0121] By analyzing the results, high-risk areas are reconfirmed and corresponding emergency response and early warning strategies are generated, establishing a complete closed-loop control process from detection to early warning. Through the coordinated operation of data acquisition, feature extraction, data processing, anomaly detection, hazard screening, detailed inspection, and early warning decision-making modules, the system can automatically match appropriate emergency response strategies based on different high-risk location types, shortening the emergency response time to less than 30 seconds and effectively reducing the false alarm rate for high-risk area identification. Compared to the traditional passive response mode, this technical solution achieves a paradigm shift from post-event handling to proactive prevention, thus reducing the false alarm rate of secondary confirmation.

[0122] This technical solution achieves intelligent upgrades and fully automated processing of drone inspection and monitoring through a modular architecture design that systematically integrates multi-source data fusion, intelligent feature extraction, precise anomaly detection, dynamic security assessment, high-risk location screening, fixed-point fine inspection, in-depth data analysis, and emergency early warning decision-making. Attached Figure Description

[0123] Figure 1 This is a schematic diagram of a sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles provided in an embodiment of the present invention;

[0124] Figure 2 This is a schematic diagram of a sensor-based unmanned aerial vehicle (UAV) intelligent inspection and monitoring system module provided in an embodiment of the present invention. Detailed Implementation

[0125] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0126] At least one embodiment of the present invention discloses a sensor-based intelligent inspection and monitoring method and system for unmanned aerial vehicles (UAVs), comprising:

[0127] like Figure 1 As shown, a sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles (UAVs) includes the following steps:

[0128] Step 1: Collect multi-source data from the inspection area to form an initial inspection dataset;

[0129] Step 2: Extract the feature data of each data item in the initial inspection dataset to obtain the feature dataset;

[0130] Step 3: Preprocess and spatiotemporally calibrate the feature dataset to generate a unified target feature dataset;

[0131] Step 4: Perform anomaly detection on the target feature dataset to obtain the anomaly inspection area dataset; based on the anomaly inspection area dataset, classify the anomaly inspection areas by security level to obtain the classification results;

[0132] Step 5: Screen the safety risk areas in the division results to obtain multiple high-risk points; obtain multiple high-risk point types and set corresponding risk point change thresholds. When the change threshold of one of the high-risk points reaches the preset threshold, it is preliminarily determined that the current inspection area is in a high-risk occurrence zone.

[0133] Step 6: By controlling the drone to fly to different altitudes, orientations, and time points in the high-risk area and hovering at fixed points, key monitoring and refined inspection of the high-risk area are carried out to obtain a secondary inspection dataset; the secondary inspection dataset is then deeply analyzed to obtain analytical results including hazard source identification, severity assessment, and trend prediction.

[0134] Step 7: Based on the analysis results, verify the preliminary judgment: if the verification is successful, confirm the current inspection area as a high-risk area, and generate an emergency response warning strategy and control strategy corresponding to the high-risk point type in the area.

[0135] In a preferred embodiment of the present invention, this application first considers how to construct a collaborative analysis framework for multi-source heterogeneous data, eliminating the data silo effect between sensors through spatiotemporal calibration. Addressing the insufficient intelligence in anomaly detection, a dynamic benchmark pattern learning mechanism is explored to enable the detection threshold to adapt to changes in characteristic parameters under different environmental conditions. To solve the accuracy problem of safety level classification, a dynamic risk scoring calculation model is designed, combining real-time parameters with historical data for weight adjustment. To address the limitations of high-risk site screening, a method combining grid subdivision and hazard index calculation is proposed to achieve dynamic location of potential hazards. To address the disconnect between the initial inspection and re-inspection processes, a threshold triggering mechanism and a multi-dimensional fixed-point re-inspection strategy are constructed to form a closed loop for anomaly verification. Finally, a comprehensive control system from data acquisition to early warning decision-making is established through a combination of multi-stage data processing and intelligent analysis.

[0136] To address this, this application proposes the following steps: collecting multi-source data from the inspection area, performing timestamp marking, coordinate positioning, and format standardization to form a structured initial inspection dataset; extracting features from various types of data in the initial inspection dataset and performing standardization to form a unified-dimensional feature dataset; performing data cleaning, time-series alignment, spatial location calibration, normalization, and standardization operations on the feature dataset, and establishing correlation analysis between features to generate a fused target feature dataset; inputting the target feature dataset into a pre-trained anomaly detection model to identify anomalous feature data deviating from the normal baseline pattern and recording their spatiotemporal location information to form an anomalous inspection area dataset; based on the anomalous inspection area dataset, calculating a comprehensive risk score for each anomalous area based on the type, intensity, and duration of the anomalous features, combined with geographical environmental sensitivity and historical accident data, dividing them into four safety levels—low risk, medium risk, high risk, and extremely high risk—according to the score range, and outputting the division results; selecting areas with high risk and extremely high risk from the division results; and gridding the selected areas. The system is divided into subdivided grid points, and the hazard index of each grid point is calculated. A hazard index threshold is set, and points exceeding the threshold are identified as high-risk points, generating a list of high-risk points. A dynamic change threshold is set for each type of high-risk point in the list. When the parameter change rate of a high-risk point reaches its corresponding change threshold, the area where the point is located is initially determined to be a high-risk area. In response to the initial determination, the UAV is controlled to fly to the high-risk area and hover at different flight altitudes, observation directions, and time nodes according to a preset multi-dimensional inspection path to collect high-resolution, refined data. Key points are repeatedly observed to obtain time-series data, which is then integrated to form a secondary inspection dataset. The secondary inspection dataset is deeply analyzed using deep learning models and data analysis algorithms to obtain analytical results including hazard source type identification, hazard level quantitative assessment, anomaly pattern comparison, and development trend prediction. The initial determination is verified based on the analytical results. If the verification is successful, the current area is confirmed as a high-risk area, and corresponding emergency response early warning strategies and control strategies are generated based on the high-risk point types.

[0137] Multi-source data acquisition refers to obtaining heterogeneous data from the inspection area through a combination of multiple sensors. Specifically, visible light sensors, infrared sensors, gas sensors, and sound sensors can be used to collaboratively collect environmental data. This data, combined with geographic information, meteorological data, and historical records, forms a structured dataset to address the insufficient monitoring dimensions caused by traditional single data sources, providing a multimodal data foundation for anomaly detection. Feature extraction involves filtering quantifiable indicators of abnormal states from the raw data. This can be achieved through image texture analysis, temperature gradient calculation, gas concentration change rate statistics, and sound spectrum analysis to establish a multi-dimensional feature space, addressing the problem that traditional methods cannot uncover the correlation patterns between data. Spatiotemporal calibration processing eliminates temporal and spatial deviations in multi-source data. This is achieved using dynamic time warping algorithms and spatial coordinate mapping techniques to ensure the spatiotemporal consistency of data from different sensors and improve the accuracy of anomaly area location. Anomaly detection identifies whether data patterns deviate from the normal baseline state. This is mainly achieved by analyzing abnormal fluctuation patterns of time-series feature parameters using a 1D-CNN+LSTM hybrid model to distinguish between normal and abnormal states, providing a foundation for subsequent comprehensive risk assessment. The safety level classification refers to introducing a multi-factor comprehensive risk assessment based on anomaly detection results. This is achieved through weighted calculations of anomaly parameter type, intensity, historical data, and geographical environmental factors, addressing the misjudgment problems caused by traditional static threshold classification. Hazard screening involves locating high-risk locations through grid subdivision and hazard index calculation. Specifically, it uses dynamically determined grid sizes based on the application scenario combined with a diffusion trend prediction model to achieve granular refinement from regional assessment to point-based monitoring. The preliminary judgment aims to quickly locate suspicious areas, optimize inspection resource allocation, and avoid unnecessary detailed inspections of normal areas. Fixed-point hovering refined inspection involves controlling drones to conduct repeated observations from multiple angles in three-dimensional space, achieved through multi-dimensional path planning and high-resolution data acquisition, to obtain multi-view verification data for anomaly areas. Deep analysis combines intelligent algorithms to predict trends in re-inspection data, specifically using time series comparison and deep learning models to identify anomaly evolution patterns, improving the reliability of early warning decisions. Secondary confirmation is based on more accurate secondary inspection data to determine the final hazard status, ensuring the accuracy of early warning decisions and forming a progressive verification relationship with the preliminary judgment.

[0138] The core innovation of this application lies in constructing a closed-loop control process of initial inspection anomaly identification, re-inspection data verification, and secondary early warning decision-making. It achieves accurate positioning of high-risk areas through multi-source data fusion and dynamic safety assessment, and forms a dual verification mechanism for anomaly confirmation by combining multi-dimensional fixed-point inspection and in-depth analysis. Finally, it generates an emergency strategy that matches the type of danger, significantly improving the proactive early warning capability of UAV inspection.

[0139] The working process and principle of this application are as follows: First, multi-source data of the inspection area is collected through various sensors, and time-stamped, coordinate located, and format-unified to form an initial inspection dataset containing environmental, geographical, and meteorological information. Then, features of various data types in this dataset, such as image texture, temperature distribution, and gas concentration, are extracted and standardized to form a unified-dimensional feature dataset. The feature dataset undergoes data cleaning, time-series alignment, spatial location calibration, normalization, and standardization, and correlation analysis between features is established to generate a fused target feature dataset. The target feature dataset is input into a pre-trained anomaly detection model to identify anomalous feature data deviating from the normal baseline pattern and record their spatiotemporal location information, forming an anomaly inspection area dataset. Based on the anomaly inspection area dataset, and considering the type, intensity, and duration of the anomalous features, combined with geographical environmental sensitivity and historical accident data, a comprehensive risk score is calculated for each anomaly area. Based on the score range, these areas are divided into four safety levels: low risk, medium risk, high risk, and extremely high risk. From the classification results, areas with high-risk and extremely high-risk safety levels are selected. These selected areas are then subdivided into grids, and a hazard index is calculated for each grid point. A hazard index threshold is set, and points exceeding this threshold are designated as high-risk points, generating a list of high-risk points. Dynamic change thresholds are set for each type of high-risk point in the list. When the parameter change rate of a high-risk point reaches its corresponding change threshold, the area where that point is located is preliminarily determined to be a high-risk zone. In response to this preliminary determination, a drone is controlled to fly to the high-risk zone and hover at different flight altitudes, observation azimuths, and time points according to a preset multi-dimensional inspection path. This hovering allows for the collection of high-resolution, refined data, and multiple repeated observations of key points to obtain time-series data, which is then integrated to form a secondary inspection dataset. Deep learning models and data analysis algorithms are used to perform deep analysis on the secondary inspection dataset, obtaining analytical results including hazard source type identification, hazard level quantitative assessment, anomaly pattern comparison, and development trend prediction. Finally, the preliminary judgment is verified based on the analysis results. If the verification is successful, the current area is confirmed as a high-risk zone, and corresponding emergency response early warning strategies and control strategies are generated based on the high-risk location types. This process forms a complete closed-loop verification from initial inspection to re-inspection to early warning, realizing dynamic monitoring from anomaly detection to accurate early warning.

[0140] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0141] During an inspection of a chemical plant area, environmental data was first collected simultaneously using visible light sensors, infrared sensors, gas sensors, and sound sensors mounted on a drone. At the same time, geographic information data, real-time meteorological data, and historical inspection records of the inspection area were acquired. All data sources were timestamped to milliseconds using a unified time base, and coordinate positioning with centimeter-level accuracy was achieved using GPS / BeiDou positioning systems. The tagged heterogeneous data was then converted into a standardized structured format to form the initial inspection dataset.

[0142] Feature data was extracted from the initial dataset. For visible light image data, texture, color, and shape features were extracted; for infrared thermal imaging data, temperature distribution, thermal anomaly region, and temperature gradient features were extracted; for gas sensor data, gas concentration and concentration change rate features were extracted; and for sound sensor data, spectral features, volume, and abnormal sound features were extracted. The Z-score normalization method was used to convert the extracted feature data into a standard normal distribution with a mean of 0 and a variance of 1, forming a feature dataset with uniform dimensions.

[0143] The feature dataset underwent data cleaning, time series alignment, spatial location calibration, normalization, and standardization, and correlation analysis was established between features. Data cleaning employed linear interpolation to handle missing values, wavelet thresholding to filter noise, and the 3σ criterion to remove outliers. Time series alignment used the sensor data with the highest sampling frequency as the benchmark, aligning other sensor data to the same time axis using linear interpolation. Spatial location calibration established a unified WGS-84 geographic coordinate system, mapping all sensor spatial data to this system through a coordinate transformation matrix. Data normalization used the min-max scaling method to normalize the data to the [0,1] interval; Z-score standardization was used to eliminate the influence of dimensions. Correlation analysis employed principal component analysis (PCA) for dimensionality reduction and calculated the Pearson correlation coefficient to construct a feature correlation matrix, generating the fused target feature dataset.

[0144] A machine learning-based anomaly detection model is established. A hybrid model combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM) is trained using feature data from a historical database under safe operating conditions to learn dynamic baseline patterns under normal operating conditions. A dynamic anomaly threshold based on statistical distribution is set; for example, for temperature features, the threshold is set to ±2.5 times the standard deviation of the baseline value. The real-time target feature dataset is input into the model, and its reconstruction error or deviation score compared to the baseline pattern is calculated. When the deviation score exceeds the preset threshold, an anomaly is identified, and the precise timestamp and geographic coordinates corresponding to the anomaly feature data are recorded, forming an anomaly inspection area dataset.

[0145] Safety levels are classified based on the abnormal inspection area dataset. Safety level scoring ranges are defined as follows: low risk (0-30 points), medium risk (31-60 points), high risk (61-90 points), and extremely high risk (91-100 points). The basic risk score is calculated using the formula: Basic Score = (Type Weight Coefficient × Intensity Normalized Value × Duration Decay Factor); where the type weight coefficient is determined by expert experience or the analytic hierarchy process, the intensity normalized value linearly transforms the original parameters to [0,1], and the duration decay factor considers the persistent effect of the anomaly. The comprehensive risk score is calculated as: Comprehensive Risk Score = Basic Score × (1 + Geographical Environment Adjustment Coefficient + Historical Accident Adjustment Coefficient). The safety level is determined based on the range of the comprehensive risk score.

[0146] From the classification results, areas classified as high-risk and extremely high-risk are selected, and these areas are further subdivided into grids. The grid size is dynamically adjusted based on the terrain complexity and sensor accuracy of the inspection area; flat areas can use 10m×10m, while complex areas use 5m×5m or smaller. The hazard index for each grid point is calculated using the formula: Hazard Index = Anomaly Parameter Intensity Weight × Intensity Normalized Value + Diffusion Trend Weight × Trend Score + Impact Range Weight × Range Score; where the sum of each weight is 1, for example, 0.4, 0.3, 0.3. The hazard index threshold is determined based on historical data statistics, for example, selecting the top 10% quantile of the index values ​​among all grid points as the threshold. Points exceeding the hazard index threshold are identified as high-risk points, and a list of high-risk points is generated. Dynamic change thresholds are set for each type of high-risk point in the list.

[0147] When the parameter change rate at a high-risk location reaches its corresponding change threshold, the area where that location is located is preliminarily identified as a high-risk zone. In response to this preliminary assessment, a drone is directed to fly to the high-risk zone and conducts a refined inspection according to a pre-set multi-dimensional inspection path. The inspection path planning for each high-risk location sets at least three flight altitude gradients (e.g., 5m, 10m, 15m), four main observation azimuths (e.g., 0°, 90°, 180°, 270°), and multiple observation time points (e.g., at 10-minute intervals). At the hovering point, high-precision sensors (e.g., high-resolution infrared thermal imagers, laser gas analyzers) are used to continuously collect data. The same high-risk location is observed at least three times to form a dataset suitable for time-series analysis, which is then integrated to form a secondary inspection dataset.

[0148] Deep learning models and data analysis algorithms are used to perform in-depth analysis of the secondary inspection dataset. Pre-trained convolutional neural network (CNN) models (such as YOLOv5) are used to identify hazards in visible light and infrared images. DS evidence theory is employed to fuse the confidence levels of multi-sensor data (such as temperature and gas concentration) for hazard assessment. The similarity between current and historical anomaly patterns is compared using the Dynamic Time Warping (DTW) algorithm. Similarity thresholds (e.g., 0.8) and duration thresholds (e.g., exceeding 30 minutes continuously) are set, and an alarm is triggered when both are met. Time series models such as ARIMA are used to predict the development trend and diffusion path of hazards over a future period (e.g., 60 minutes), obtaining analytical results that include hazard type identification, hazard level quantification, anomaly pattern comparison, and development trend prediction.

[0149] The preliminary judgment is verified based on the analysis results. If the verification is successful, the current area is confirmed as a high-risk zone. Taking a flammable gas leak as an example, the generated warning information includes: the coordinates of the leak point (118.2°E, 39.1°N), the methane concentration reaching 1500ppm, and the expected diffusion range being a circular area with a radius of 50 meters centered on the leak point. Based on the high-risk location type, a corresponding emergency response warning strategy is generated. The corresponding control strategies include: activating the southeast sprinkler system, evacuating all personnel within a 100-meter radius to a safe assembly point, closing the upwind valve, and notifying the fire department.

[0150] Through the above-described scheme, this application achieves dynamic and accurate monitoring of abnormal states in the inspection area. Spatiotemporal calibration and feature fusion of multi-source data improve the accuracy of anomaly detection. The machine learning-based anomaly detection model can adapt to environmental changes, reducing the false alarm rate. Dynamic safety level classification and hazard screening enhance the accuracy of high-risk area identification. A multi-dimensional fixed-point re-inspection mechanism enhances the reliability of anomaly verification. Deep analysis and trend prediction provide a comprehensive basis for early warning decisions. A closed-loop process from initial inspection to re-inspection to early warning is formed, significantly improving the timeliness and targeting of emergency response. This scheme effectively solves the problems of inaccurate anomaly detection, non-dynamic safety assessment, and delayed early warning decisions in traditional inspections, providing more reliable technical support for safety monitoring in industrial scenarios.

[0151] In some of the solutions mentioned above in this application, it is proposed to collect multi-source data from the inspection area to form an initial inspection dataset to support subsequent analysis. However, since the multi-source data includes heterogeneous data sources such as sensor data, geographic information data, meteorological data and historical data, these data differ in terms of collection methods, time synchronization and spatial positioning accuracy, making it difficult to directly fuse and analyze the original data, and failing to guarantee spatiotemporal consistency, which in turn affects the accuracy of subsequent feature extraction and anomaly detection.

[0152] To address this, this application further proposes collecting multi-source data from the inspection area, including: simultaneously collecting environmental data using visible light sensors, infrared sensors, gas sensors, and sound sensors mounted on a drone; acquiring geographic information data, real-time meteorological data, and historical inspection records of the inspection area; using a unified time base to timestamp all data sources at the millisecond level, and using the GPS / BeiDou positioning system for centimeter-level coordinate positioning; and converting the marked heterogeneous data into a standardized structured format to form an initial inspection dataset.

[0153] The visible light sensor, infrared sensor, gas sensor, and sound sensor can be configured in different locations on the drone's fuselage. For example, the visible light sensor can be mounted below the gimbal, while the infrared and gas sensors are integrated onto a rotatable bracket. Geographic information data is acquired in real time via a GPS module, meteorological data is received synchronously from a weather station via wireless transmission, and historical inspection records are retrieved from a local database. Timestamps use a unified time reference source, such as GPS timing signals, with marking accuracy controlled to the millisecond level. Coordinate positioning associates sensor data with geographic coordinates through a geographic information system, for example, mapping gas concentration data to a three-dimensional geographic coordinate system. Format standardization processing includes converting image data into geographically labeled raster data and converting numerical sensor data into timestamped tabular data, such as integrating infrared thermal imaging data and gas concentration data into a spatiotemporally correlated multidimensional data matrix.

[0154] Specifically, when the drone flies over the inspection area, it simultaneously collects environmental data through multiple sensors. For example, visible light sensors acquire surface images, infrared sensors capture thermal radiation distribution, gas sensors detect the concentration of specific chemical components, and sound sensors record the environmental noise spectrum. Geographic information data provides a spatial reference for the sensor data; for example, it associates thermal anomaly areas with terrain features using geographic coordinates. Meteorological data is dynamically linked to sensor data through timestamps; for example, it combines wind speed data with gas diffusion trends for analysis. Historical inspection records, after time alignment, provide a longitudinal comparison benchmark for the current data. Timestamp marking adopts a hierarchical marking method; for example, the original timestamp is marked at the data acquisition end, and time axis alignment is performed at the data processing end. Coordinate positioning maps discrete sensor data to continuous geographic space through spatial interpolation algorithms; for example, it uses Kriging interpolation to generate a spatial distribution surface of gas concentration. Format unification processing integrates heterogeneous data through data conversion interfaces; for example, it matches the EXIF ​​information of image files with the spatiotemporal labels of sensor data to form a structured dataset with triple indexes of time, space, and data type. Thus, multi-source data achieves uniformity in spatiotemporal benchmarks and data structures, providing highly consistent data input for subsequent anomaly detection.

[0155] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0156] Multi-source data from the inspection area is collected to form an initial inspection dataset. Specifically, environmental data of the inspection area is collected using visible light sensors, infrared sensors, gas sensors, and sound sensors mounted on the drone. The visible light sensors collect high-resolution image data, the infrared sensors collect temperature distribution data, the gas sensors collect various gas concentration data, and the sound sensors collect environmental noise data. Furthermore, geographic information data, meteorological data, and historical inspection records of the inspection area are acquired. Geographic information data includes spatial information such as topography and building distribution; meteorological data includes real-time meteorological parameters such as temperature, humidity, and wind speed; and historical inspection records include past anomalies and their handling.

[0157] Next, the collected sensor data, geographic information data, meteorological data, and historical inspection records will be timestamped and their coordinates determined. The timestamps are accurate to the millisecond level, and the coordinates are determined using a GPS / BeiDou dual-mode positioning system with centimeter-level accuracy. This achieves spatiotemporal alignment of multi-source data.

[0158] Finally, the labeled multi-source data underwent format standardization processing to form a structured initial inspection dataset. Specifically, image data was converted into a standardized pixel matrix format, sensor numerical data was converted into a unified time series format, geographic information data was converted into a standard GIS format, and text data was converted into a structured JSON format. Through format standardization processing, a standardized and structured initial inspection dataset was constructed, laying the foundation for subsequent analysis.

[0159] Through the above technical solutions, this application achieves efficient acquisition and fusion of multi-source heterogeneous data. By combining multiple sensors, it comprehensively covers key physical quantities for environmental monitoring. Timestamp marking and coordinate positioning ensure the spatiotemporal consistency of the data, eliminating spatiotemporal deviations between different data sources. Format standardization transforms heterogeneous data into a standardized, structured form, facilitating subsequent analysis and processing. Thus, a spatiotemporally unified and structurally standardized initial dataset is constructed, providing a high-quality data foundation for subsequent feature extraction and anomaly detection, and improving the accuracy and reliability of data analysis.

[0160] In some of the above-mentioned solutions in this application, feature parameters of each data item in the initial inspection dataset are extracted to form a feature parameter set. However, due to the heterogeneity and multidimensionality of multi-source data, the feature dimensions and physical meanings of different sensor data are significantly different. The directly extracted feature parameters are difficult to achieve an effective expression of a unified dimension, which leads to problems such as data matching difficulties and weak feature correlation in subsequent spatiotemporal calibration and anomaly detection processes.

[0161] In response, this application further proposes to extract texture features, color features, and shape feature parameters from visible light image data; extract temperature distribution features, thermal anomaly region features, and temperature gradient feature parameters from infrared thermal imaging data; extract gas concentration features and concentration change rate feature parameters from gas sensor data; and extract spectral features, volume features, and abnormal sound feature parameters from sound sensor data. The extracted feature parameters are then standardized to form a feature parameter set with a unified dimension.

[0162] The texture features of visible light image data are calculated using the gray-level co-occurrence matrix to determine energy values ​​and contrast parameters, for example, calculating a 14-dimensional texture descriptor within the 0-255 grayscale range. Color features are achieved using HSV color space histogram statistics, with the 8 bins of the H channel extracted in practice. Shape features are calculated using Canny edge detection to determine the ratio of contour area to perimeter, serving as a deformation quantification indicator. The temperature distribution features of infrared thermal imaging data are characterized by the mean and variance of temperature values ​​across all pixels in the thermal image, for example, calculating the temperature standard deviation within a 100×100 pixel area. Thermal anomaly region features are identified using an adaptive threshold segmentation method to identify regions where the temperature exceeds twice the standard deviation of the environmental mean. Temperature gradient features are calculated using the Sobel operator to determine the gradient magnitude in the X / Y directions, for example, using a gradient threshold of 0.5℃ / pixel. The concentration change rate features of gas sensor data are calculated using time-series differencing, for example, calculating the ppm / s change rate with a 10-second time window. The spectral features of sound sensor data are extracted using Fast Fourier Transform to determine the energy distribution in the 0-20kHz frequency band, for example, dividing it into 32 frequency bands for energy integration. The standardization process uses the Z-score method to map each feature parameter to the [-1,1] interval. For example, the temperature feature is transformed by (measured value - μ) / σ.

[0163] Specifically, for visible light image data, texture features are extracted by analyzing the pixel grayscale spatial distribution patterns. Energy values ​​reflect the uniformity of the image's grayscale distribution, while contrast parameters characterize texture clarity. Color features are quantified in the HSV space, and H-channel statistics effectively distinguish between vegetation cover changes and traces of chemical pollution. Shape features detect equipment deformation through contour analysis; for example, a shape anomaly is triggered when the tilt angle of a power transmission tower exceeds 5°. For infrared thermal imaging data, the mean value in the temperature distribution features reflects the overall thermal equilibrium state, and a variance exceeding 50℃² indicates a thermal anomaly. Thermal anomaly region features locate localized overheating through dynamic threshold segmentation; an alarm is triggered when the detected area accounts for more than 0.1% of the total area. Temperature gradient features reveal the direction of heat conduction, and abrupt changes in gradient direction can be detected in the early stages of electrical equipment failure. The concentration change rate feature of gas sensor data distinguishes between continuous leaks and transient interference; a leak event is identified when the change rate exceeds 5 ppm / s for three consecutive sampling periods. Abnormal sound features of sound sensor data are achieved by matching a pre-stored fault soundprint library; for example, an early warning is triggered when the similarity of the bearing damage soundprint reaches 85%. During the standardization process, each feature parameter undergoes dimensionless transformation to form a unified 128-dimensional feature vector, enabling the temperature, decibel, and concentration values ​​from different sensors to be correlated and analyzed within the same mathematical space. Through this technical solution, the feature dimension of data from four types of heterogeneous sensors was reduced from the original 256 dimensions to the standardized 128 dimensions, compressing the data volume by 50% while retaining over 95% of the feature information, thus establishing a calculable data foundation for subsequent spatiotemporal calibration.

[0164] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0165] Texture, color, and shape features are extracted from visible light image data. Texture features can be extracted using the gray-level co-occurrence matrix method, calculating statistics such as energy, contrast, and correlation; color features can be extracted using the HSV color space, calculating the mean and variance of hue, saturation, and brightness; and shape features can be extracted using contour detection algorithms, calculating geometric parameters such as area, perimeter, and roundness.

[0166] This method extracts temperature distribution features, thermal anomaly region features, and temperature gradient feature parameters from infrared thermal imaging data. It can calculate statistical quantities such as the mean, standard deviation, skewness, and kurtosis of the temperature field to characterize the temperature distribution; identify thermal anomaly regions by setting temperature thresholds and calculate their area proportions and highest temperatures; and use the Sobel operator to calculate the temperature gradient field and extract gradient magnitude and direction information.

[0167] Gas concentration characteristics and concentration change rate characteristics are extracted from gas sensor data. The system can calculate the average concentration, peak concentration, and duration of exceedance for various gases; and uses the sliding window method to calculate the concentration change rate over time.

[0168] Spectral features, volume features, and abnormal sound feature parameters are extracted from sound sensor data. Power spectral density can be calculated using Fast Fourier Transform to extract features such as dominant frequency and frequency band energy; sound pressure level can be calculated to characterize volume; and abnormal sound templates can be set up, with pattern matching methods used to identify abnormal sounds.

[0169] The extracted feature parameters are standardized to form a feature parameter set with a unified dimension. The Z-score standardization method can be used to convert feature parameters of different dimensions into a standard normal distribution with a mean of 0 and a variance of 1, thereby achieving a unified expression of the feature space.

[0170] Through the above technical solution, this application achieves feature extraction and unified representation of multi-source heterogeneous data. Differentiated feature extraction rules are designed for data from different sensors, capturing multi-dimensional anomalous features such as visible light, infrared, gas, and sound. Standardization eliminates the influence of different physical dimensions, mapping heterogeneous features to a unified mathematical space. This provides a computable data foundation for subsequent spatiotemporal correlation analysis and anomaly detection, effectively solving the problem of feature dimension mismatch in multi-source data and improving the comprehensiveness and relevance of feature representation.

[0171] In some of the solutions described above in this application, the original feature parameter sets suffer from data quality defects, spatiotemporal inconsistencies, and insufficient correlation among multi-source features. Specifically, the multi-source data collected by sensors contains missing values, noise interference, and outliers, leading to reduced reliability of the feature parameters; misaligned timestamps from different sensor data cause biases in time-series analysis; uncalibrated spatial coordinates result in broken spatial correlations; differences in feature parameter dimensions hinder unified analysis; and a lack of correlation mining leads to the failure of multi-source data fusion. These problems directly affect the accuracy of subsequent anomaly detection and the reliability of security assessment.

[0172] To address this, this application further proposes to process the feature parameter set by handling missing values, filtering noise, and removing outliers; performing time series alignment based on the timestamps of the sensor data; performing spatial position calibration based on the spatial coordinates of the sensor data; normalizing and standardizing the processed feature parameters; and establishing correlation analysis between feature parameters to generate a fused target feature parameter set.

[0173] Missing value handling can be achieved through linear interpolation or substitution using the mean of adjacent data. For example, cubic spline interpolation is used when more than three consecutive sampling points are missing. Noise filtering can be achieved using wavelet threshold denoising algorithm, with a decomposition level of 5 layers and a soft threshold function to handle high-frequency coefficients. Outlier removal is achieved by calculating the Z-score statistic and identifying data points with an absolute value exceeding 3 as outliers. Time series alignment uses a dynamic time warping algorithm, using the main sensor's acquisition frequency as a reference to resample and align other sensor data to the same time axis. Spatial location calibration is achieved through coordinate transformation matrices, such as registering lidar point cloud data with the visible light image coordinate system using affine transformation. Normalization uses the minimax method to map data to the [0,1] interval, and standardization uses the Z-score method to eliminate dimensional differences. Correlation analysis calculates the linear correlation between features using the Pearson correlation coefficient and analyzes nonlinear correlations using mutual information. For example, a strong correlation is established when the correlation coefficient between temperature gradient features and gas concentration change rate exceeds 0.7.

[0174] Specifically, in the data preprocessing stage, missing value handling ensures data integrity by imputing or deleting invalid data. For example, when gas sensor data is missing, linear interpolation is used to supplement it using the concentration gradient change trend of adjacent time points. Noise filtering uses the Kalman filter algorithm to eliminate random interference in infrared sensor thermal imaging data, achieving optimal filtering by setting the process noise covariance to 0.01 and the observation noise covariance to 0.1. Outlier removal uses box plots to identify outliers in voiceprint features, automatically removing data points exceeding 1.5 times the interquartile range. In the spatiotemporal calibration stage, time series alignment is based on the UTC timestamp of the GPS module, and a timestamp matching algorithm aligns the sampling times of different sensors to millisecond-level accuracy, eliminating time series misalignment caused by differences in acquisition frequency. Spatial location calibration unifies multi-source data to the WGS-84 geographic coordinate system through coordinate system transformation. For example, the local coordinate system data of UAV-borne lidar is converted to geodetic coordinates, ensuring the comparability of spatial location information collected by different sensors. In the data standardization stage, normalization linearly transforms the temperature characteristics from [-20, 100]℃ to the [0, 1] interval, and standardization adjusts the gas concentration characteristics to a standard normal distribution with a mean of 0 and a variance of 1. In the correlation analysis stage, principal component analysis extracts common variation patterns of temperature gradient, gas concentration change rate, and acoustic signature spectral characteristics, establishing a coupling relationship matrix between multi-source features. For example, when the temperature gradient rise rate exceeds 0.5℃ / s and the gas concentration change rate exceeds 10%, key monitoring of the acoustic signature spectral characteristics is triggered. This scheme, through four progressive processing steps—data cleaning, spatiotemporal calibration, standardization, and correlation mining—forms a high-precision, strongly correlated set of target feature parameters, providing reliable input for subsequent anomaly detection.

[0175] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0176] The feature parameter set undergoes missing value processing, noise filtering, and outlier removal. Missing value processing employs mean interpolation, replacing missing values ​​for continuous feature parameters with their mean, and for discrete feature parameters with their mode. Noise filtering utilizes wavelet transform to decompose the feature parameters at multiple scales, removing high-frequency noise components before reconstructing the signal. Outlier removal uses the 3σ criterion, considering data points exceeding the mean ± 3 standard deviations as outliers and removing them.

[0177] Time series alignment is performed based on the timestamps of the sensor data. First, the sensor data with the highest sampling frequency is selected as the baseline time axis, and other sensor data are aligned to the baseline time axis using linear interpolation. For sensors with significantly different sampling frequencies, a sliding window averaging method is used for downsampling to ensure all sensor data achieve a uniform time resolution.

[0178] Spatial position calibration is performed based on the spatial coordinates of the sensor data. A unified spatial coordinate system is established, and the spatial coordinates of different sensors are mapped to this unified coordinate system through coordinate transformation. For sensors with installation deviations, spatial position correction is performed using a calibration plate method to eliminate systematic errors.

[0179] The processed feature parameters are normalized and standardized. Normalization uses the min-max scaling method to linearly transform the feature parameters to the [0,1] interval. Standardization uses the Z-score standardization method to convert the feature parameters into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0180] A correlation analysis was conducted among the feature parameters to generate a fused target feature parameter set. Principal component analysis (PCA) was used to reduce feature dimensionality and extract key feature components. The correlation between feature parameters was calculated using the Pearson correlation coefficient to construct a feature correlation matrix. Based on the correlation matrix, a hierarchical clustering algorithm was used to group highly correlated features, selecting the feature parameter with the highest information content in each group as a representative. Finally, the selected representative feature parameters were combined to form the fused target feature parameter set.

[0181] Through the above technical solution, this application improves the quality and reliability of multi-source data, achieves spatiotemporal consistency of data from different sensors, eliminates dimensional differences in feature parameters, and uncovers the correlations between multi-source features. The resulting target feature parameter set possesses high information fusion and representativeness, providing a reliable data foundation for subsequent anomaly detection and security assessment. Furthermore, this solution, through a systematic data preprocessing and fusion process, effectively addresses issues such as data quality defects, spatiotemporal inconsistencies, and insufficient correlations among multi-source features in the original feature parameter set, thereby improving the accuracy of anomaly detection and the reliability of security assessment.

[0182] In some of the solutions mentioned above in this application, traditional methods lack intelligent anomaly detection mechanisms and cannot dynamically learn the feature parameter baseline pattern under normal conditions, resulting in misjudgment due to reliance on fixed thresholds for anomaly determination. At the same time, the failure to combine spatiotemporal location information for accurate marking during anomaly detection leads to inaccurate location of anomaly areas, making it difficult to support the refined requirements of subsequent security level classification.

[0183] To address this, this application further proposes establishing a machine learning-based anomaly detection model, training a baseline pattern of feature parameters under normal conditions, setting anomaly detection thresholds and anomaly judgment rules for various feature parameters, inputting the target feature parameter set into the anomaly detection model for real-time comparison and analysis, identifying anomaly feature parameters that deviate from the baseline pattern by more than a preset threshold, and marking the spatiotemporal location information corresponding to the anomaly feature parameters to form an anomaly inspection area dataset.

[0184] The machine learning model can employ a Long Short-Term Memory (LSTM) network or an autoencoder architecture, using unsupervised training to extract time-series patterns from historical normal data, for example, using time-series data with a sliding window length of 60 seconds as input samples. Anomaly detection thresholds are set as a combination of dynamic range and statistical distribution thresholds, with the temperature parameter threshold set to a baseline value ±2.5℃ and the gas concentration threshold determined using the 3σ principle. Spatiotemporal location information is marked using GPS coordinates and a GIS geocoding system, with timestamp accuracy controlled at the millisecond level. Correlation analysis between multi-source feature parameters is achieved using a Pearson correlation coefficient matrix; feature groups with correlation coefficients exceeding 0.8 are designated as associated detection objects. The anomaly determination rule employs a multi-condition logical judgment mechanism; when more than two associated features simultaneously trigger the threshold, an anomaly is confirmed.

[0185] Specifically, the baseline pattern of feature parameters under normal conditions is constructed through a combination of offline training and online updates. Historical data, after being sampled through a sliding window, is input into the model for feature extraction, generating a dynamic baseline template containing mean, variance, and trend terms. During the real-time detection phase, the preprocessed target feature parameter set is input into the model for pattern matching, and anomaly quantification is achieved by calculating Mahalanobis distance or reconstruction error indices. When a deviation of temperature distribution features from the baseline pattern exceeds 2.3℃ and lasts for more than 15 seconds, the corresponding geographic coordinates and detection time are automatically recorded. The spatiotemporal labeling of anomalous feature parameters adopts a hierarchical storage structure, with geographic coordinate information, sensor number, and detection time forming a multi-dimensional index, supporting rapid retrieval by the subsequent safety level assessment module. By integrating a dynamic threshold adjustment mechanism, the temperature baseline value can be automatically updated by ±1.5℃ to adapt to changes in operating conditions when the ambient temperature undergoes seasonal changes. The binding processing of anomaly detection results with spatiotemporal information improves the positioning accuracy of high-risk areas to within 0.5 meters, providing a precise spatial reference benchmark for subsequent emergency response.

[0186] As a preferred embodiment, the solution of this application is implemented as follows: An anomaly detection model based on machine learning is established, and a deep learning algorithm is used to train a baseline pattern of feature parameters under normal conditions. Historical normal operation data is used as the training set, including multi-dimensional feature parameters such as temperature, pressure, and vibration. The model adopts a multi-layer neural network structure, and the parameters are optimized through a backpropagation algorithm to learn the complex relationships between features.

[0187] Furthermore, anomaly detection thresholds and anomaly judgment rules are set for various characteristic parameters. For example, upper and lower limit thresholds are set for temperature parameters, frequency and amplitude thresholds are set for vibration parameters, and change rate thresholds are set for gas concentration parameters. The anomaly judgment rules adopt a multi-index comprehensive scoring mechanism, considering factors such as the degree of parameter deviation, duration, and correlation.

[0188] Therefore, the target feature parameter set is input into the anomaly detection model for real-time comparative analysis. The model calculates the degree of deviation between the input data and the baseline pattern, generating anomaly scores. Specifically, a sliding time window technique is used to dynamically analyze continuous time series data, capturing instantaneous and gradual anomalies.

[0189] The system identifies abnormal feature parameters that deviate from the baseline pattern by more than a preset threshold. When the anomaly score exceeds the set threshold, an anomaly detection alarm is triggered. As a preferred implementation, a multi-level threshold strategy is adopted, classifying anomalies into three levels: minor, moderate, and severe.

[0190] Finally, the spatiotemporal location information corresponding to the anomaly feature parameters is labeled to form an anomaly inspection area dataset. The precise timestamps and geographic coordinates of the anomalies are recorded, and information such as anomaly type, severity, and duration are associated. The anomaly inspection area dataset is stored in a structured format for easy subsequent analysis and visualization.

[0191] Through the above technical solutions, this application realizes an intelligent anomaly detection mechanism based on machine learning, overcoming the limitations of traditional fixed threshold detection methods. Dynamically learning the baseline pattern of feature parameters under normal conditions improves the accuracy and adaptability of anomaly detection. The setting of multi-dimensional anomaly detection thresholds and judgment rules effectively reduces false alarms and missed alarms. Real-time comparison analysis and dynamic threshold strategies enable timely capture and identification of anomalies. Precise spatiotemporal labeling of anomaly feature parameters provides a reliable basis for subsequent security level classification and risk assessment. Overall, this solution significantly improves the anomaly detection capability and early warning accuracy of the UAV intelligent inspection system, providing strong support for timely discovery of potential safety hazards and prevention of major accidents.

[0192] In some of the solutions mentioned above in this application, the traditional safety assessment process relies solely on a single abnormal parameter for static threshold judgment, which cannot dynamically integrate multi-dimensional risk factors and lacks quantitative assessment of the differences in abnormal characteristic parameter types, intensity fluctuations, and durations, resulting in deviations between the safety level classification results and the actual situation.

[0193] In response, this application further proposes a method for classifying safety levels based on anomaly inspection area datasets. This method includes establishing a safety level assessment system and setting four levels: low risk, medium risk, high risk, and extremely high risk. It also involves calculating risk scores based on the type, intensity, and duration of anomaly characteristic parameters, adjusting risk weights by combining geographical environmental factors and historical accident records, classifying the data based on the comprehensive risk score, and generating classification results that include regional coordinates, safety level, and risk characteristic descriptions.

[0194] The safety level assessment system comprises four levels, which can be achieved by setting scoring ranges. For example, 0-30 points is considered low risk, 31-60 points medium risk, 61-90 points high risk, and 91 points and above extremely high risk. During risk scoring calculation, the type weights of abnormal characteristic parameters can be set based on expert experience; for example, the weight coefficient for gas leaks is 0.6, and for temperature anomalies it is 0.4. Intensity parameters are normalized to a 0-1 range, and duration parameters are calculated using a logarithmic function for decay. Adjustments to geographical environmental factors can include terrain slope coefficient correction, increasing the risk weight by 20% when the slope exceeds 15 degrees; historical accident record adjustments can use a sliding time window statistical method, increasing the weight by 4-5% for each additional accident in the past three years. During classification, the comprehensive score is input into a decision tree model for level determination. Historical accident data and expert annotation results are used as supervision signals during model training. When generating the classification results, regional coordinates are projected and transformed using a geographic information system, and risk characteristic descriptions are automatically filled with abnormal parameter types and intensity ranges using a natural language generation template.

[0195] Specifically, the dataset of abnormal inspection areas is first input into the safety level assessment system. Four risk levels are quantified into classification boundaries through pre-defined scoring intervals, which can be dynamically adjusted based on historical accident data. During the risk scoring calculation phase, type weight coefficients are determined using the analytic hierarchy process (AHP), for example, the weight for structural deformation anomalies is 0.55, and for environmental parameters it is 0.45. Intensity parameters are converted to standard values ​​through maximum-minimum value normalization, and duration parameters are processed using an exponential decay function, triggering a persistent impact factor when a set threshold is exceeded. During the adjustment of geographical environmental factors, slope sensor data is compared in real-time with the geological database, automatically increasing the weight coefficient by 20% when a landslide risk area is detected. Historical accident records are used to calculate the impact weight of recent accidents using a time decay model, with the weight coefficient for accident records in the last three months reaching 0.8. After the comprehensive score calculation is completed, a fuzzy logic-based classifier is used for level determination. The classifier's rule base contains 12 judgment conditions derived from expert experience. The final classification results are stored in a spatial database, with regional coordinates labeled using the WGS84 coordinate system. The risk feature description field includes the peak value of abnormal parameters, duration, and weight adjustment records. This process, through dynamic quantitative assessment and multi-source data fusion, ensures that the safety level classification results are consistent with the actual risk distribution patterns, providing an accurate spatial positioning benchmark for subsequent screening of high-risk locations.

[0196] As a preferred embodiment, the specific implementation of this application's solution is as follows: A safety level assessment system is established, setting four levels: low risk, medium risk, high risk, and extremely high risk. Risk scores are calculated based on the type, intensity, and duration of abnormal characteristic parameters. For example, for gas concentration anomalies, the concentration values ​​can be divided into four intervals, corresponding to four risk levels; for temperature anomalies, different temperature thresholds can be set to classify risk levels. Intensity parameters can be normalized to unify different types of abnormal parameters into the range of 0-1. Duration can be set to multiple time periods, such as 0-10 minutes, 10-30 minutes, 30-60 minutes, and over 60 minutes, each assigned a different weight. Risk weights are adjusted by combining the geographical environmental factors and historical accident records of the inspection area. Geographical environmental factors can include terrain, climate conditions, surrounding facilities, etc., while historical accident records can statistically analyze the frequency and severity of accidents over a certain period. The safety level of the abnormal inspection area is classified based on the comprehensive risk score. Weighted average or machine learning algorithms can be used to synthesize various indicators to obtain the final risk score. A classification result containing area coordinates, safety level, and risk characteristic descriptions is generated. Coordinates can be represented by latitude and longitude, safety levels are indicated by color codes, and risk characteristic descriptions include key information such as the main abnormal parameter types, intensity, and duration.

[0197] Through the aforementioned technical solution, this application achieves precise safety level classification of abnormal inspection areas. By establishing a multi-dimensional safety level assessment system, combining the type, intensity, and duration of abnormal characteristic parameters, as well as geographical environmental factors and historical accident records, a comprehensive quantitative assessment of risk is conducted. This method overcomes the limitations of traditional single-threshold judgments, improving the accuracy and dynamic adaptability of safety level classification. Therefore, it provides reliable data support for the subsequent precise screening of high-risk locations and the formulation of emergency response strategies, effectively enhancing the targeting and effectiveness of risk management.

[0198] In some of the solutions mentioned above in this application, a technical solution for classifying the safety level of abnormal inspection areas was proposed. Therefore, this application further proposes to screen out areas with high-risk and extremely high-risk safety levels from the classification results; to further subdivide the screened low-safety areas into grids to establish refined monitoring points; to calculate the hazard index of each monitoring point, comprehensively considering the intensity of abnormal parameters, diffusion trends, and impact range; to set a hazard screening threshold and screen out monitoring points whose hazard index exceeds the threshold; and to locate the selected high-risk points by coordinates and identify their hazard types to form a list of high-risk points.

[0199] The screening of high-risk areas can be achieved by setting safety level thresholds, for example, classifying areas with a risk score higher than 80 as extremely high-risk. Grid subdivision can employ a dynamic grid division algorithm, automatically adjusting the grid size based on the area and terrain complexity; for example, using a 10m x 10m grid in flat areas and a 5m x 5m grid in complex terrain areas. The calculation of the hazard index can combine a weighted algorithm, such as assigning 40% weight to the intensity of abnormal parameters, 30% to the diffusion trend, and 30% to the area of ​​influence, generating an index value of 0-100 through normalization. Screening thresholds can be set based on historical data statistics; for example, selecting the top 10% of monitoring points by hazard index as high-risk points. Coordinate positioning can be achieved using a geographic information system for latitude and longitude coordinate conversion, and hazard type identification can be matched according to a pre-set hazard source classification coding table.

[0200] Specifically, the process begins by identifying high-risk and extremely high-risk areas based on safety level classification results, ensuring that subsequent processing focuses on critical areas. The selected areas are then subdivided into grids; for example, each high-risk area is divided into multiple 5m x 5m monitoring units, with the center point of each unit serving as a monitoring point. The hazard index for each point is generated using a multi-dimensional calculation model, where the intensity of abnormal parameters is quantified using sensor data, the spread trend is predicted through time series analysis, and the impact range is calculated using geographic information data. When the hazard index exceeds a preset threshold, such as a score of 75 or higher, the point is marked as a high-risk point. Finally, a structured list is generated through coordinate transformation and type matching; for example, high-temperature anomalies are marked as Class A hazard sources, and gas leaks are marked as Class B hazard sources. This process, through a multi-level screening mechanism and dynamic evaluation model, achieves a technical closed loop from area division to point focus, effectively improving the identification accuracy and positioning efficiency of high-risk points.

[0201] As a preferred embodiment, the specific implementation of this application is as follows: Areas with high-risk and extremely high-risk safety levels are selected from the classification results. The selected low-risk areas are further subdivided into grids to establish refined monitoring points. Grid subdivision can be achieved using an equal-interval method, dividing the low-risk areas into 10m × 10m grid units, with the center point of each grid unit serving as a monitoring point. The hazard index for each monitoring point is calculated, comprehensively considering the intensity of abnormal parameters, diffusion trends, and impact range. The hazard index calculation formula is: Hazard Index = Abnormal Parameter Intensity × 0.4 + Diffusion Trend × 0.3 + Impact Range × 0.3. Wherein, the intensity of abnormal parameters, diffusion trends, and impact range are all standardized values ​​from 0 to 100. A hazard screening threshold is set, and monitoring points with hazard indices exceeding the threshold are selected. The hazard screening threshold can be set to 80 points. The selected high-risk points are located by coordinates and identified by hazard type, forming a list of high-risk points. Coordinates are located using latitude and longitude, and hazard type identification includes categories such as gas leaks, fire hazards, and equipment malfunctions.

[0202] Through the above technical solutions, this application achieves a closed-loop technology from large-scale safety level classification to refined high-risk location positioning. Multi-level screening and refined analysis improve the accuracy and spatial resolution of high-risk location positioning. Grid-based subdivision and multi-dimensional hazard assessment overcome the limitations of single-indicator assessment, making precise hazard source location possible. Setting screening thresholds avoids misjudgments or missed detections due to improper threshold settings. The final high-risk location list provides clear target locations and classification criteria for subsequent refined UAV inspections, improving overall monitoring efficiency and early warning accuracy.

[0203] In some of the solutions mentioned above in this application, after the high-risk area is initially identified, conventional inspection methods cannot perform dynamic and refined monitoring of the multidimensional characteristics of high-risk locations, resulting in a single data collection angle and insufficient time series coverage, making it difficult to capture the dynamic change characteristics of the hazard source; at the same time, the lack of a comparison benchmark for multiple observations of the same high-risk location makes it impossible to effectively verify the persistence and development trend of the abnormal state, affecting the accuracy of subsequent analysis results and the timeliness of early warning.

[0204] To address this, this application further proposes a method of controlling drones to hover at different altitudes, orientations, and time points within high-risk areas for focused monitoring and refined inspection of these areas, thereby acquiring a secondary inspection dataset. This includes: developing multi-dimensional inspection path plans for high-risk locations, setting different flight altitudes, observation orientations, and hovering times; controlling drones to fly to each high-risk location along the planned path, performing multi-angle fixed-point hovering observations at each location; continuously collecting high-resolution sensor data during hovering, including close-range images, precise temperature, and gas concentration data; repeatedly observing the same high-risk location to obtain time-series variation data; based on the time-series variation data, acquiring inspection data at multiple different time points and establishing a data comparison benchmark set; and integrating and labeling the collected refined observation data to form a secondary inspection dataset for high-risk areas.

[0205] In the multi-dimensional inspection path planning, the flight altitude can be set from 5 to 50 meters. The observation azimuth is divided into eight azimuth angles at 45-degree intervals, centered on the high-risk point. The hovering time is set from 30 seconds to 5 minutes, depending on the sensor type. For example, in a gas leak monitoring scenario, the flight altitude is adjusted to below 10 meters to improve the accuracy of gas concentration detection, and the hovering time is extended to 3 minutes to obtain stable concentration gradient data. During multi-angle fixed-point hovering observation, the UAV maintains gimbal stability through the attitude adjustment module, enabling the visible light sensor to collect image data with a resolution higher than 0.5 mm / pixel within a 2-meter range of the target. The acquisition frequency of time-series variation data can be once per minute, with the continuous observation period covering the diurnal temperature variation period. The resulting comparison benchmark set contains at least three sets of standardized data templates at different time points.

[0206] Specifically, after high-risk locations are initially identified, the path planning algorithm generates a three-dimensional waypoint sequence containing altitude, azimuth, and dwell time based on the location coordinates. After the UAV arrives at the first high-risk location along the planned path, it hovers at a height of 8 meters above the ground and rotates to observe at four azimuths: 0 degrees, 90 degrees, 180 degrees, and 270 degrees, staying at each azimuth for 90 seconds to collect infrared thermal imaging data. In the second observation cycle, the flight altitude is adjusted to 15 meters, the azimuth interval is increased by 45 degrees, and the dwell time is shortened to 60 seconds to capture rapidly changing heat source characteristics. Two complete observation cycles are performed at the same location during the morning, noon, and evening periods, generating a time series containing 12 sets of heat distribution data. These data, after being timestamped, are correlated with volatile organic compound (VOC) concentration data collected by gas sensors at the same location, forming a traceable comparison benchmark. By analyzing the covariant relationship between thermal radiation intensity and gas concentration over three consecutive cycles, different risk modes of instantaneous overheating and continuous leakage on equipment surfaces can be distinguished, providing a quantitative basis for early warning strategies.

[0207] As a preferred embodiment, the specific implementation of this application's solution is as follows: For the selected high-risk locations, a three-dimensional inspection path is first generated according to the location type. The flight altitude is set to three gradients: 5 meters, 10 meters, and 15 meters. The observation azimuth is set to a combination of three directions: due north, southeast, and southwest. The hovering time is configured to three modes: 2 minutes, 5 minutes, and 8 minutes. After the UAV accurately arrives at the target coordinate point via the RTK positioning system, the three-axis gimbal stabilization device is activated, maintaining a hovering state at a height of 10 meters in the southeast direction for 5 minutes. Simultaneously, the high-resolution infrared thermal imager and laser gas detection module are activated to continuously collect temperature field distribution data and methane concentration fluctuation values ​​at a rate of 3 frames per second. The same high-risk location is repeatedly observed in three time periods within 24 hours, collecting three sets of time-series data at 08:00, 14:00, and 20:00 respectively. Each set of data contains 2000 temperature sampling points and 1500 gas concentration readings. After the observation data is double-marked with timestamps and spatial coordinates, a set of comparison benchmarks containing 72 sets of multidimensional data is formed, and each set of data is associated with a corresponding height-azimuth-time combination code.

[0208] Through the above technical solution, this application effectively overcomes the technical shortcomings of traditional inspections in terms of single data acquisition dimensions. By covering three-dimensional space and repeatedly observing over multiple time periods, it comprehensively captures the three-dimensional feature changes of high-risk locations. The established multi-dimensional data comparison benchmark set can accurately identify the correlation between the distribution of abnormal temperature gradients and the diffusion trend of gas concentrations, solving the problem of difficulty in verifying the persistence of abnormal states in existing technologies. The dynamically acquired high-frequency time-series data provides reliable data support for identifying instantaneous anomalies and persistent risks, improving the accuracy of predicting the development trend of hazardous sources to an engineering level where early warning can be implemented.

[0209] In some of the solutions mentioned above in this application, a deep analysis of the secondary inspection dataset is proposed to identify high-risk areas. However, in this process, there are problems such as the inability to effectively identify specific hazard types, the lack of multi-dimensional assessment of the degree of danger, the difficulty in discovering similar anomaly point representation patterns, and the inability to predict the development trend of hazard sources, resulting in insufficient accuracy of secondary confirmation and insufficient targeting of early warning strategies.

[0210] To address this, this application further proposes a deep analysis of the secondary inspection dataset to obtain the analysis results, including: establishing a deep learning model for intelligent identification and analysis of the secondary inspection data; identifying specific hazard types and distribution characteristics in high-risk areas using image recognition technology; comprehensively analyzing multi-sensor data using data fusion algorithms to assess the hazard level; comparing and analyzing multiple inspection data to identify similar anomaly representation patterns; setting preset thresholds for similar anomaly representations and their durations; triggering an alarm mechanism when similar anomaly representations exceed the preset thresholds and their durations exceed the preset thresholds; predicting the development trend and possible diffusion paths of hazard sources based on time series analysis methods; and generating a comprehensive analysis result containing hazard identification results, severity assessment reports, anomaly comparison analysis, and trend prediction information.

[0211] The deep learning model can be constructed using a hybrid architecture of convolutional neural networks and recurrent neural networks. For example, ResNet-50 can be used for feature extraction followed by LSTM network processing of time-series data. Image recognition technology can be combined with target detection algorithms, such as the YOLOv5 model, to identify hazard source types, while semantic segmentation technology can be used to extract heat maps of hazard source distribution. The data fusion algorithm can employ the DS evidence theory to fuse infrared temperature data and gas concentration data with confidence, where the weight for temperature anomalies is set to 0.6 and the weight for gas concentration anomalies is set to 0.4. The preset threshold for representing similar anomalies can be set to an 80% similarity matching rate, and the duration threshold can be set to more than 30 consecutive minutes. Time series analysis can use the ARIMA model to predict the hazard diffusion trend over the next 60 minutes, where the historical data window is set to the past 2 hours of data.

[0212] Specifically, after the drone completes a detailed inspection of high-risk areas, the secondary inspection data is input into a trained deep learning model for feature extraction, such as automatic identification of temperature gradient features in thermal imaging data. Through the image recognition module, high-temperature hotspots and gas leak areas can be accurately distinguished; high-temperature hotspots are marked as red polygonal areas, and gas leak areas are marked as yellow diffusion outlines. During multi-sensor data fusion, the 80°C anomaly data collected by the temperature sensor and the 500ppm methane concentration data detected by the gas sensor are weighted and calculated to generate a level-three hazard assessment result. In the comparative analysis stage, the similarity of anomaly features collected at different time points is calculated. When the similarity exceeds 80% for three consecutive detection cycles, it is determined to be a persistent anomaly pattern. After the alarm is triggered, the time-series analysis module predicts, based on gas diffusion rate data from the past two hours, that the danger zone will expand 5 meters southeast within the next hour. The final comprehensive analysis results include a hazard type marking map, a 3D risk assessment heatmap, an anomaly pattern matching report, and a diffusion path prediction animation. The hazard type marking map uses different color codes to distinguish five types of hazards, including electrical fires and chemical leaks. By associating and matching the feature vectors output by the deep learning model with the extracted standardized feature parameters, the reliability of anomaly pattern recognition is enhanced. For example, comparing the current temperature gradient features with a historical normal pattern database triggers a comparison analysis process if the deviation exceeds 15%. The weighting parameters introduced in the data fusion stage are linked to the safety level classification scoring mechanism; when the hazard level reaches level three, the decision weight of the gas concentration parameter is automatically increased to 60%. The duration threshold setting is synchronized with the baseline pattern update frequency of the anomaly detection model, and the threshold is dynamically calibrated every 10 minutes. This closed-loop analysis system improves the accuracy of hazard identification.

[0213] As a preferred embodiment, the specific implementation of this application is as follows: Visible light and infrared thermal imaging data of high-risk areas are acquired using a multispectral imaging device deployed on a drone. The visible light images are used for target detection with a YOLOv5 model to identify hazard sources such as corroded pipes and insulation damage. Infrared thermal imaging data is used to extract temperature anomaly area features through a ResNet50 network. Combined with time-series methane concentration data collected by a gas sensor, multi-source data fusion is performed using DS evidence theory to generate a comprehensive hazard index including temperature gradient and concentration change rate. The abnormal feature parameters obtained from three consecutive inspections are dynamically time-warped, and their Euclidean distance similarity is calculated. When the cumulative duration of similar anomaly points exceeds 30 minutes and the similarity threshold exceeds 0.85, an audible and visual alarm device is triggered. The temperature anomaly area is trend-predicted using an ARIMA model, calculating the probability distribution of heat diffusion range within the next two hours. Finally, a three-dimensional visualization analysis report is generated, including a hazard source location map, a hazard level heat map, and vectorized predicted path data.

[0214] Through the above technical solutions, this application achieves accurate identification of hazard source types in high-risk areas. A dynamic assessment system encompassing multiple dimensions, including temperature, concentration, and diffusion rate, is constructed by integrating multimodal sensor data. A dual-threshold judgment mechanism effectively distinguishes between sporadic equipment interference and persistent hazardous events. Combined with a time-series prediction model, the probability distribution of hazard diffusion paths is predicted two hours in advance, extending the emergency response time window from the traditional 15 minutes to 120 minutes. The final comprehensive analysis report overlays a three-dimensional visualization of the hazard source spatial distribution, quantitative assessment results, and predicted paths, enabling monitoring personnel to intuitively identify high-risk core areas and potential impact ranges, thus reducing the false alarm rate of secondary confirmation.

[0215] Through the above technical solutions, this application constructs a multi-module collaborative closed-loop control architecture, solves the problem of multi-source data fusion through spatiotemporal calibration processing, achieves accurate positioning of high-risk areas by using dynamic threshold mechanism and grid screening, and effectively avoids single detection errors by combining multi-dimensional secondary inspection data verification. Finally, it forms a fully automated processing flow from anomaly identification to emergency response, which significantly improves the accuracy and timeliness of hazard warning.

[0216] like Figure 2 As shown, a sensor-based unmanned aerial vehicle (UAV) intelligent inspection and monitoring system includes:

[0217] The data acquisition module is used to collect multi-source data from the inspection area to form an initial inspection dataset;

[0218] The feature extraction module is used to extract feature data for each data item in the initial inspection dataset to obtain a feature dataset;

[0219] The data processing module is used to preprocess and spatiotemporally calibrate the feature dataset to generate a unified target feature dataset.

[0220] The anomaly detection module is used to perform anomaly detection on the target feature dataset to obtain an anomaly inspection area dataset; based on the anomaly inspection area dataset, the anomaly inspection area is divided into security levels to obtain the division results.

[0221] The hazard screening module is used to screen the safety risk areas in the division results to obtain multiple high-risk points; obtain multiple high-risk point types and set corresponding hazard point change thresholds. When the change threshold of one of the high-risk points reaches the preset threshold, it is preliminarily determined that the current inspection area is in a high-risk occurrence zone.

[0222] The fine inspection module is used to control drones to fly to different altitudes, orientations, and time points in high-risk areas and hover at fixed points to conduct key monitoring and fine inspection of high-risk areas, and obtain secondary inspection datasets; deep analysis of the secondary inspection datasets is performed to obtain analytical results including hazard source identification, severity assessment, and trend prediction.

[0223] The early warning decision module is used to verify the preliminary judgment based on the analysis results: if the verification is successful, the current inspection area is confirmed as a high-risk area, and an emergency response early warning strategy and control strategy corresponding to the high-risk point type in the area are generated.

[0224] This application further proposes a system architecture including a data acquisition module, a feature extraction module, a data processing module, an anomaly detection module, a hazard screening module, a detailed inspection module, and an early warning decision-making module. The data acquisition module collects environmental, geographical, meteorological, and historical inspection data through multi-sensor collaboration to form a structured initial dataset. For example, it uses visible light sensors, infrared sensors, gas sensors, and sound sensors to simultaneously collect data, and achieves millisecond-level time synchronization through timestamps. The feature extraction module extracts standardized feature parameters for different data types. For example, it extracts texture feature parameters from visible light images, temperature gradient feature parameters from infrared data, and concentration change rate feature parameters from gas data, eliminating data heterogeneity. The data processing module unifies multi-source data to the same spatiotemporal reference through spatiotemporal calibration processing. For example, it uses a geographic coordinate system transformation algorithm to spatially align sensor data with geographic information data and applies a normalization algorithm to eliminate dimensional differences.

[0225] The anomaly detection module uses machine learning models to identify abnormal feature parameters that deviate from a baseline pattern. For example, it uses a convolutional neural network to establish a baseline pattern of feature parameters for normal states. When the target feature parameter deviates from the baseline value by more than 15%, an anomaly marker is triggered. The hazard screening module subdivides low-safety areas into grids. For example, high-risk areas are divided into 1m×1m monitoring grids, and a hazard index is calculated for each grid. When the hazard index exceeds a preset threshold, a list of high-risk locations is generated. The fine-tuning inspection module uses multi-dimensional path planning to control drones for fixed-point hovering observation. For example, it collects data from multiple angles at different heights of 3m, 5m, and 8m, and repeats the data three times at 30-second intervals at the same location to form time-series comparison data. The early warning decision module reconfirms high-risk states based on deep analysis results. For example, when both image recognition and gas concentration data exceed thresholds, an emergency response strategy is triggered, generating control instructions including evacuation routes.

[0226] Specifically, the data acquisition module obtains multi-source data, which is then processed through feature extraction and spatiotemporal calibration to form a unified target feature set. The anomaly detection module identifies abnormal areas and classifies safety levels through a dynamic threshold mechanism. The hazard screening module performs grid subdivision and hazard calculation on high-risk areas, screening out high-risk locations and setting change thresholds. When the gas concentration change rate at a certain high-risk location reaches 0.5% / second, a preliminary warning is triggered, and the fine inspection module is activated. The UAV hovers at different altitudes along a preset path to conduct multi-angle observations and acquire high-resolution secondary data. The warning decision module confirms the dangerous state by comparing time-series data. For example, when the temperature gradient exceeds 10℃ / m in three observations, a warning message containing the coordinates of the isolation area and the emergency response level is generated. Through closed-loop interaction of data flow and command flow, the modules achieve end-to-end control from anomaly identification to precise response. The spatiotemporal calibration error is controlled within ±0.3 meters. In the hazard index calculation, the anomaly parameter intensity weight accounts for 60%, the diffusion trend weight accounts for 30%, and the historical data correlation weight accounts for 10%. Through the collaborative operation of the modular architecture, the system reduces the false alarm rate of high-risk area identification to below 5% and shortens the emergency response time to within 30 seconds.

[0227] As a preferred embodiment, the solution of this application is implemented as follows: The data acquisition module synchronously collects environmental data through visible light sensors, infrared sensors, gas sensors, and sound sensors, and integrates terrain elevation data from the geographic information system with real-time wind speed and direction data transmitted from the meteorological station to form an initial inspection dataset containing spatial coordinates and timestamps. The feature extraction module performs gray-level co-occurrence matrix texture analysis on visible light images, extracts temperature gradient histogram features from infrared thermal imaging data, calculates standard deviation statistics for gas concentration data, and converts multi-source features into a standardized parameter matrix of unified dimensions. The data processing module uses a dynamic time warping algorithm to align multi-sensor time series, achieves spatial benchmark unification based on a geographic coordinate transformation model, and eliminates dimensional differences through Z-score standardization to generate a spatiotemporally correlated target feature parameter set. The anomaly detection module uses the isolated forest algorithm to establish a normal pattern benchmark, identifies abnormal areas where the temperature gradient standard deviation exceeds 3σ, calculates regional risk scores by weighting historical accident data, and outputs safety level classification results. The hazard screening module divides high-risk areas into 10m x 10m grids, calculates a weighted hazard index of temperature abrupt change rate and gas concentration change rate in each grid, and filters grids with an index exceeding 0.85 as high-risk locations. A preliminary warning is triggered when three consecutive frames of data for a location exceed the threshold. The fine-tuning inspection module controls a drone to conduct three-view hovering observations at heights of 5m, 10m, and 15m above the ground of high-risk locations, with a data acquisition time of no less than 120 seconds at each location. High-resolution thermal maps are acquired using a multispectral imager, generating a time-series comparison dataset. The early warning decision module uses a convolutional neural network to identify pipeline crack features in the thermal map, combines this with a gas diffusion model to predict the hazard range, and generates an emergency strategy that includes evacuation route planning and equipment control instructions.

[0228] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles (UAVs), characterized in that, include: Collect multi-source data from the inspection area to form an initial inspection dataset; Extract the feature data of each data item in the initial inspection dataset to obtain the feature dataset; The feature dataset is preprocessed and spatiotemporally calibrated to generate a unified target feature dataset; Perform anomaly detection on the target feature dataset to obtain an anomaly inspection area dataset; Based on the abnormal inspection area dataset, the abnormal inspection areas are classified into security levels to obtain the classification results; The safety risk areas in the above division results were screened for hazard levels, and multiple high-risk locations were obtained; Multiple high-risk location types are obtained, and corresponding risk point change thresholds are set. When the change threshold of one of the high-risk locations reaches the preset threshold, it is preliminarily determined that the current inspection area is in a high-risk occurrence zone. By controlling drones to fly to different altitudes, directions, and time points in high-risk areas and hovering at fixed points, key monitoring and refined inspection of high-risk areas can be carried out to obtain secondary inspection datasets. Deep analysis of the secondary inspection dataset yields analytical results including hazard identification, severity assessment, and trend prediction. Based on the analysis results, the preliminary judgment is verified: if the verification is successful, the current inspection area is confirmed as a high-risk area, and an emergency response warning strategy and control strategy corresponding to the high-risk point type in the area are generated.

2. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, Collect multi-source data from the inspection area to form an initial inspection dataset, including: The first dataset is obtained by collecting environmental data of the inspection area using multiple sensors mounted on the drone; The second dataset is obtained by acquiring geographic information data, meteorological data, and historical inspection records of the inspection area. The first and second datasets are timestamped and located using coordinates to obtain the location results. The multi-source data items in the marker positioning results are formatted and standardized to form the initial inspection dataset.

3. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, Extract the feature data of each data item in the initial inspection dataset to obtain the feature dataset, including: The system sends inspection commands through the control terminal. Upon receiving the inspection commands, it activates the multispectral camera, infrared thermal imager, and vibration sensor to collect multi-source data of the inspection area. Extract texture, color, and shape features from visible light image data; Extract temperature distribution features, thermal anomaly region features, and temperature gradient features from infrared thermal imaging data; Extract gas concentration characteristics, concentration change rate characteristics, and gas composition characteristics from gas sensor data; Extract spectral features, volume features, and abnormal sound feature data from sound sensor data; If the sensor malfunctions or the data quality does not meet the requirements, a prompt message will be returned to the user in a timely manner, and it will be suggested to replace the equipment or adjust the parameters until the data acquisition conditions meet the requirements before the data is transmitted to the preprocessing step. The extracted feature data are standardized to form a feature dataset with a unified dimension. The process involves extracting key features from the transmitted multi-source data using a deep learning model, and then combining this with time series analysis algorithms to uncover the spatiotemporal correlation patterns between these features.

4. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, The feature dataset is preprocessed and spatiotemporally calibrated to generate a unified target feature dataset, including: The feature dataset is processed by handling missing values, filtering noise, and removing outliers to obtain the preprocessed results. The timestamps of each data item in the preprocessing result are aligned to a time series to obtain the first processing result. The spatial coordinates of each data item in the first processing result are calibrated to obtain the second processing result; Normalize and standardize the feature datasets after the two processing steps to obtain the operation results; Establish correlation analysis among the feature data items in the operation results to generate a fused target feature dataset.

5. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, Anomaly detection is performed on the target feature parameter set to obtain an anomaly inspection region dataset, including: An anomaly detection model was established and trained using safe zone data from a historical database to obtain a baseline pattern of feature parameters. Set dynamic thresholds and judgment rules for anomaly detection with different feature parameters; The target feature parameter set is input into the trained anomaly detection model for anomaly detection analysis. If an abnormal area is found, the spatial topology modeling algorithm is used to assess the security level of the abnormal area, and the anomaly detection results are passed to the security level classification step. Identify abnormal feature parameters whose deviations from the baseline pattern exceed the corresponding dynamic threshold in the abnormal detection results; If additional data is needed or the boundaries of abnormal areas need to be confirmed, the inspection data should be collected again. Record the spatiotemporal location information corresponding to the abnormal feature parameters to form an abnormal inspection area dataset.

6. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, Based on the abnormal inspection area dataset, the abnormal inspection areas are classified into security levels, and the classification results are as follows: A basic risk score is calculated based on the type, intensity, and duration of the abnormal characteristic parameters; The basic risk score is weighted and adjusted by combining the geographical sensitivity of the inspection area and historical accident data to obtain a comprehensive risk score; Based on the range of the comprehensive risk score, the safety level of the abnormal inspection area is classified. The output includes the partitioning results, which contain information on area coordinates, security level, and risk description. Based on the classification results, a safety level assessment system was established, comprising four levels: low risk, medium risk, high risk, and extremely high risk.

7. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, A hazard screening was conducted on areas classified as high-risk and extremely high-risk in the risk assessment results, revealing several high-risk locations, including: Areas marked as medium risk, high risk, and extremely high risk were selected from the classification results; The selected area is divided into grids to generate multiple fine-grained monitoring points; Calculate the risk index for each monitoring point, which is based at least on the intensity of the abnormal parameters, the trend of change, and the potential range of impact. A risk index threshold is set, and monitoring points whose risk index exceeds the risk index threshold are identified as high-risk points; Each high-risk location is precisely located by coordinates and its hazard type is identified, generating a list of high-risk locations.

8. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, The method involves controlling a drone to hover at different altitudes, orientations, and time points within a high-risk area to conduct focused monitoring and refined inspections of the high-risk zone, acquiring a secondary inspection dataset, including: Based on the anomaly detection results, a path planning algorithm is invoked to generate a refined inspection path for high-risk locations; Based on the list of high-risk locations, a refined inspection path is planned for the drone, including different flight altitudes, observation directions, and hovering time points; During the inspection, the flight controller monitors the drone's position and attitude in real time, and the cruise combined with obstacle avoidance algorithms obtains the drone's safe flight path; Control the drone to fly to each high-risk location according to the planned path, and perform multi-angle fixed-point hovering; If obstacles or external interference prevent the path from continuing, return to the anomaly detection step to reassess the abnormal area or adjust the path planning. During hovering, a refined dataset including high-resolution images, precise temperature data, and gas concentration data is acquired using high-precision sensors. Repeated observations were performed on key high-risk locations in the refined dataset to obtain time-series variation data. After the inspection is completed, the detailed inspection data collected in the second step is transmitted to the preprocessing step. By integrating refined data and time-series change data, a secondary inspection dataset is formed.

9. The sensor-based intelligent inspection and monitoring method for unmanned aerial vehicles according to claim 1, characterized in that, The deep analysis of the secondary inspection dataset yields the following results: Establish a deep learning model to intelligently identify and analyze secondary inspection data; The data integration algorithm is invoked to clean and classify the inspection data, and the cleaning and classification results are obtained. Identify the specific hazard source types and distribution characteristics of high-risk areas in the cleaning and classification process to obtain the identification results; The identification results are comprehensively analyzed to assess the risk level of the high-risk areas currently being inspected. Conduct hazard comparison analysis on multiple critical inspection areas to identify the hazard characteristics of similar anomalies; Set thresholds for the hazard characteristics and duration of similar anomalies; When the characteristics of similar anomalies exceed a preset threshold and the duration exceeds a preset time threshold, an alarm mechanism is triggered; and an emergency warning decision report is generated in conjunction with a report generation algorithm. The report includes the distribution of abnormal areas, the results of the security level assessment, and recommendations for countermeasures; The emergency early warning decision report is sent to the user terminal, and the inspection results are presented on the display terminal. If there are missing or incomplete contents in the report, the user will be prompted to supplement the inspection information. Based on the emergency warning decision report, the development trend and potential diffusion path of the hazard source are predicted, and a comprehensive analysis result is finally generated, which includes the hazard source identification results, severity assessment report, anomaly point comparison analysis and trend prediction information.

10. A sensor-based intelligent inspection and monitoring system for unmanned aerial vehicles (UAVs), used to execute a sensor-based intelligent inspection and monitoring method for UAVs as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect multi-source data from the inspection area to form an initial inspection dataset; The feature extraction module is used to extract feature data for each data item in the initial inspection dataset to obtain a feature dataset; The data processing module is used to preprocess and spatiotemporally calibrate the feature dataset to generate a unified target feature dataset. The anomaly detection module is used to perform anomaly detection on the target feature dataset to obtain an anomaly inspection area dataset; based on the anomaly inspection area dataset, the anomaly inspection area is divided into security levels to obtain the division results. The hazard screening module is used to screen the safety risk areas in the division results to obtain multiple high-risk points; obtain multiple high-risk point types and set corresponding hazard point change thresholds. When the change threshold of one of the high-risk points reaches the preset threshold, it is preliminarily determined that the current inspection area is in a high-risk occurrence zone. The fine inspection module is used to control the drone to fly to different altitudes, directions and time points in high-risk areas and hover at fixed points to conduct key monitoring and fine inspection of high-risk areas and obtain secondary inspection datasets. Deep analysis of the secondary inspection dataset yields analytical results including hazard identification, severity assessment, and trend prediction. The early warning decision module is used to verify the preliminary judgment based on the analysis results: if the verification is successful, the current inspection area is confirmed as a high-risk area, and an emergency response early warning strategy and control strategy corresponding to the high-risk point type in the area are generated.

Images