Solid waste storage yard remote cruise and hidden danger identification system based on AI technology

Abstract

The invention relates to the technical field of intelligent inspection, in particular to a solid waste storage yard remote cruise and hidden danger recognition system based on the AI technology, which comprises a multi-source monitoring data acquisition module, a storage yard feature enhancement processing module, a multi-mode hidden danger collaborative recognition module and a dynamic cruise path generation and execution module. The unmanned aerial vehicle carries a multispectral camera and a laser radar to cruise according to an initial path, collects visible light images, infrared images and terrain point cloud data of a storage yard, and receives leachate monitoring data. Image enhancement is realized through optical dust scattering compensation and a pile body surface curvature mapping graph; the deformable convolution unit is combined with percolate temporal and spatial change rate anomaly judgment to trigger high-precision identification, and outputs a hidden danger type and a space coordinate set; according to the method, a closed loop of recognition, path optimization and re-acquisition is realized, and the hidden danger detection precision and the monitoring efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology, and in particular to a remote patrol and hazard identification system for solid waste storage yards based on AI technology. Background Technology

[0002] Solid waste disposal sites serve as crucial locations for the centralized storage, transfer, and temporary handling of urban and industrial solid waste. Their safe operation directly impacts the safety of surface and groundwater environments, the stability of surrounding ecosystems, and the structural stability of the disposal sites themselves. With the widespread application of sensing technologies such as drones, multispectral cameras, lidar, and leachate sensor networks in environmental monitoring, remote patrols and intelligent identification have become important directions for improving the efficiency of disposal site supervision.

[0003] Existing technologies still face several technical bottlenecks in practical applications: multi-source data acquisition is often conducted in time segments, lacking a strict spatiotemporal alignment mechanism, resulting in the inability to establish consistent spatial coordinate information and unified timestamps between image information and sensor data, affecting the accurate location of hazard features; dust in stockpile sites is non-uniformly distributed, and traditional visible light image enhancement methods mostly rely on fixed-parameter dehazing / descattering models, making it difficult to adaptively compensate for dust scattering in different areas, easily leading to loss of details or overcorrection; infrared images exhibit significant geometric deformation under the influence of the curvature of complex stockpile surfaces, and conventional correction methods do not combine three-dimensional surface reconstruction and differential geometric calculations, resulting in the offset of thermal anomaly area boundaries; leachate monitoring data usually only performs single-point threshold comparisons, lacking a systematic determination of spatiotemporal change rates and historical normal fluctuation ranges, making it impossible to trigger higher-precision identification at the moment of abnormal changes; drone patrol paths are mostly fixed presets or manually adjusted temporarily, lacking an adaptive dynamic patrol path generation mechanism based on regional hazard density distribution maps and density gradient information, resulting in uneven allocation of monitoring resources, overlapping coverage and blind spots, making it difficult to form a closed-loop optimization process of "identification—path optimization—re-acquisition—re-identification". Summary of the Invention

[0004] This invention provides a remote patrol and hazard identification system for solid waste storage yards based on AI technology.

[0005] A remote patrol and hazard identification system for solid waste storage yards based on AI technology includes: Multi-source monitoring data acquisition module: Controls the UAV equipped with a multispectral camera and lidar to cruise along the initial path, simultaneously acquiring visible light images, infrared images and terrain point cloud data of the stockpile, and receiving leachate monitoring data uploaded by the ground leachate sensor network to generate a multi-source monitoring dataset; The stockpile feature enhancement processing module performs optical dust scattering compensation on the visible light images in the multi-source monitoring dataset, and uses an adaptive transmittance correction matrix to eliminate non-uniform dust interference; it constructs a surface curvature mapping map of the stockpile based on the lidar point cloud data, and performs geometric deformation correction on the infrared images according to the curvature mapping map, and outputs an enhanced feature map set. Multimodal Hazard Collaborative Identification Module: The enhanced feature map set is input into a pre-trained hazard identification model, which includes deformable convolutional units, wherein the kernel parameters of the deformable convolutional units are dynamically adjusted according to the input curvature map; the leachate monitoring data is fused synchronously, and when an abnormal change in leachate concentration is detected in a certain area, the identification processing of the image of that area is triggered, and the hazard type and spatial coordinate set are output; Dynamic cruise path generation and execution module: Calculates the regional hazard density distribution map based on the hazard type and spatial coordinate set, generates a dynamic cruise path that prioritizes coverage of high-density hazard areas based on the gradient information of the hazard density distribution map, updates the UAV flight instructions, and controls the UAV to execute the next round of monitoring tasks.

[0006] Optionally, the multi-source monitoring data acquisition module includes: Control the drone to cruise along the initial path: Control the drone to fly according to the preset initial path; Acquiring visible light images of the stockpile: During the cruise, the UAV uses a rigidly mounted multispectral camera to acquire visible light images of the target area of ​​the stockpile. Infrared image and terrain point cloud data acquisition: The UAV is equipped with a rigid lidar to simultaneously acquire infrared images and terrain point cloud data of the target area during the cruise. Receive leachate monitoring data: Receive leachate monitoring data uploaded in real time from the ground leachate sensor network deployed in the stockyard; Generate a multi-source monitoring dataset: The collected visible light images, infrared images, and topographic point cloud data of the storage yard, as well as the received leachate monitoring data, are spatiotemporally aligned and correlated to generate a multi-source monitoring dataset containing timestamps and spatial coordinate information.

[0007] Optionally, the stockpile feature enhancement processing module includes: Acquire multi-source monitoring dataset: Receive multi-source monitoring dataset from the multi-source monitoring data acquisition module. This dataset includes visible light images, infrared images, topographic point cloud data, and leachate monitoring data of the stockyard. Establish a dust scattering model: Perform atmospheric optical analysis on the visible light images of the stockpile in the multi-source monitoring dataset to establish a dust scattering model characterizing dust concentration and particle size distribution; Perform optical dust scattering compensation: Using the established dust scattering model, generate an adaptive transmittance correction matrix, and apply the adaptive transmittance correction matrix to perform inversion calculation on the visible light image of the stockpile to eliminate non-uniform dust interference and obtain a dust-compensated visible light image. Constructing a curvature mapping map of the pile surface: Perform three-dimensional surface reconstruction and differential geometry calculation on the terrain point cloud data in the multi-source monitoring dataset to generate a curvature mapping map of the pile surface that characterizes the local concave and convex features of the pile surface; Perform infrared image geometric deformation correction: Based on the generated surface curvature mapping of the pile, calculate the pixel position offset of the infrared image in the multi-source monitoring dataset caused by the surface curvature of the pile, perform geometric transformation operation, complete geometric deformation correction, and obtain the deformed infrared image. Generate an enhanced feature map set: Integrate the obtained dust-compensated visible light image, deformation-corrected infrared image, and pile surface curvature map to form a unified enhanced feature map set.

[0008] Optionally, the multimodal hazard collaborative identification module includes: Acquire enhanced feature maps and leachate monitoring data: Receive enhanced feature maps from the stockpile feature enhancement processing module and leachate monitoring data from the multi-source monitoring data acquisition module; Configure deformable convolutional unit parameters: extract the curvature map of the pile surface from the enhanced feature map set, and input the curvature map of the pile surface into the deformable convolutional unit of the hazard identification model, dynamically adjust the convolutional kernel offset parameter of the deformable convolutional unit, so that the convolutional kernel adapts to the geometric features of the pile surface; Perform initial image hazard identification: Input the dust-compensated visible light image and the geometrically deformed infrared image from the enhanced feature map set into the configured hazard identification model for initial identification, and generate preliminary hazard identification results, including the preliminary identified potential hazard areas and their preliminary hazard types.

[0009] Optionally, the multimodal hazard collaborative identification module further includes: Monitoring abnormal changes in leachate concentration: Analyze the leachate monitoring data in real time, calculate the spatiotemporal gradient of leachate concentration at each monitoring point, and when an abnormal change in leachate concentration is detected in a certain area, mark the area as an abnormal leachate area and obtain its spatial coordinates. Trigger and execute high-precision identification: For marked leachate abnormal areas, extract high-resolution image data (visible light image and infrared image) of the corresponding spatial coordinates in the enhanced feature map set, input them into the hazard identification model for higher computational accuracy or higher resolution image analysis processing, and generate hazard identification results; The preliminary hazard identification results are fused and conflict resolution is performed with the obtained hazard identification results. Combined with the abnormal information of the leachate monitoring data, the hazard type and its precise spatial coordinate set of each hazard area are finally determined.

[0010] Optionally, the dynamic cruise path generation and execution module includes: Receive hazard identification results: Receive the hazard type and spatial coordinate set output by the multimodal hazard collaborative identification module; Calculate the regional hazard density distribution map: Based on the received set of spatial coordinates, the kernel density estimation algorithm is used to calculate the regional hazard density distribution map of the entire stockpile area. This map represents the number of hazards per unit area at different locations in grid form. Extracting density gradient information: Spatial gradient calculation is performed on the generated regional hazard density distribution map to obtain density gradient information that characterizes the rate and direction of change of hazard density.

[0011] Optionally, the dynamic cruise path generation and execution module further includes: Generate a dynamic cruise path: Based on the acquired density gradient information, a gradient ascent algorithm is used to plan a dynamic cruise path that starts from the current drone position, prioritizes passing through high-risk density areas, and covers preset key monitoring points. Update UAV flight commands: Convert the generated dynamic cruise path into a waypoint sequence and flight parameters that the UAV flight control system can recognize, and generate updated UAV flight commands; Execute the next round of monitoring: Send updated UAV flight commands to the UAV flight control system, control the UAV to execute the next round of monitoring tasks according to the dynamic cruise path, and trigger the multi-source monitoring data acquisition module to start a new round of data acquisition.

[0012] Optionally, the generation of spatiotemporal alignment and correlation in the multi-source monitoring dataset includes: Add spatial coordinate information generated by the UAV GNSS positioning module to the visible light image, infrared image, and terrain point cloud data of the stockyard; The leachate monitoring data is used to generate a leachate concentration distribution raster map that matches the image resolution using a spatial interpolation algorithm; Based on a high-precision clock source, a unified timestamp is added to the visible light image, infrared image, terrain point cloud data, and leachate concentration distribution raster map of the stockpile, and a four-dimensional index relationship is established.

[0013] Optionally, the determination of the abnormal mutation is performed in the following manner: Calculate the spatiotemporal rate of change of the leachate monitoring data; Compare the spatiotemporal rate of change with the historical concentration change pattern at that location; When the rate of change exceeds 30% of the upper limit of the historical normal fluctuation range, it is judged as an abnormal mutation.

[0014] The beneficial effects of this invention are: This invention utilizes a multi-source monitoring data acquisition module to simultaneously acquire visible light images, infrared images, topographic point cloud data, and leachate monitoring data of a storage yard within the same flight. It also establishes a "four-dimensional index relationship" by adding a unified timestamp using the spatial coordinate information from the UAV's GNSS positioning unit and a high-precision clock source. This mechanism achieves a strict correspondence between data from different sensors in both spatial and temporal dimensions, avoiding feature mismatches and misjudgments caused by asynchronous data acquisition and inconsistent coordinate systems in traditional methods. Simultaneously, spatial interpolation is performed on the leachate monitoring data to generate a leachate concentration distribution raster map, enabling this data to participate in subsequent analysis with resolution consistent with the images. This ensures the integrity and retrieval of input information during the hazard identification stage, improving the overall identification accuracy and stability of the system from the source.

[0015] This invention employs a stockpile feature enhancement processing module that generates an adaptive transmittance correction matrix through optical dust scattering compensation, resulting in a dust-compensated visible light image. It then utilizes topographic point cloud data to construct a surface curvature mapping map of the stockpile, and uses this map to perform geometric deformation correction on the infrared image, forming an enhanced feature atlas. A multimodal hazard collaborative identification module further dynamically adjusts the convolution kernel offset parameters of deformable convolutional units based on the stockpile surface curvature mapping map, ensuring the convolution kernel sampling shape adapts to the geometric features of the stockpile surface. Simultaneously, it introduces a spatiotemporal change rate determination based on leachate monitoring data; when the change rate exceeds the upper limit of the historical normal fluctuation range by 20%, it is identified as an abnormal mutation, triggering a high-precision identification process. This chain of "dust compensation—geometric correction—deformable convolution—abnormal trigger fine identification—result fusion" significantly improves the accuracy of hazard type and spatial coordinate set output, reducing false detections and missed detections caused by dust obstruction, surface curvature distortion, and insufficient single-modal information.

[0016] This invention utilizes a dynamic cruise path generation and execution module to calculate a regional hazard density distribution map based on hazard types and spatial coordinate sets. It extracts density gradient information and uses a gradient ascent algorithm to generate a dynamic cruise path, enabling the UAV to prioritize traversing high-hazard-density areas and covering preset key monitoring points from its current location. This path is then converted into updated UAV flight commands for the next round of monitoring, simultaneously triggering the multi-source monitoring data acquisition module again, achieving a closed-loop process of "identification result—path optimization—re-acquisition—re-identification." Compared to traditional methods of fixed flight paths or manual temporary path adjustments, this solution can rapidly replan when hazard distribution changes, improving monitoring efficiency and coverage effectiveness, reducing redundant tracks and invalid data collection, and enabling continuous tracking and dynamic, refined management of yard hazards. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the system flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the multimodal hazard collaborative identification module in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0020] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0021] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0022] like Figures 1-2 As shown, a remote patrol and hazard identification system for solid waste storage yards based on AI technology includes: Multi-source monitoring data acquisition module: Controls the UAV equipped with a multispectral camera and lidar to cruise along the initial path, simultaneously acquiring visible light images, infrared images and terrain point cloud data of the stockpile, and receiving leachate monitoring data uploaded by the ground leachate sensor network to generate a multi-source monitoring dataset; The stockpile feature enhancement processing module performs optical dust scattering compensation on the visible light images in the multi-source monitoring dataset, and uses an adaptive transmittance correction matrix to eliminate non-uniform dust interference; it constructs a surface curvature mapping map of the stockpile based on the lidar point cloud data, and performs geometric deformation correction on the infrared images according to the curvature mapping map, and outputs an enhanced feature map set. Multimodal Hazard Collaborative Identification Module: The enhanced feature map set is input into a pre-trained hazard identification model, which includes deformable convolutional units, wherein the kernel parameters of the deformable convolutional units are dynamically adjusted according to the input curvature map; the leachate monitoring data is fused synchronously, and when an abnormal change in leachate concentration is detected in a certain area, the identification processing of the image of that area is triggered, and the hazard type and spatial coordinate set are output; Dynamic cruise path generation and execution module: Calculates the regional hazard density distribution map based on the hazard type and spatial coordinate set, generates a dynamic cruise path that prioritizes coverage of high-density hazard areas based on the gradient information of the hazard density distribution map, updates the UAV flight instructions, and controls the UAV to execute the next round of monitoring tasks.

[0023] The multi-source monitoring data acquisition module includes: Controlling the UAV to cruise along the initial path: First, the UAV is controlled to fly according to the preset initial path. The initial path is generated in advance at the ground station based on the overall distribution of the storage yard, the location of no-fly zones, and line-of-sight obstructions, and is issued to the UAV in the form of a waypoint sequence. After takeoff, the UAV's autopilot controller performs the cruise mission point by point along the initial path. When encountering temporary obstacles or sudden weather changes, the autopilot controller only makes minor adjustments without deviating from the main line of the initial path, ensuring that the subsequently collected visible light images, infrared images, and terrain point cloud data of the storage yard remain consistent with the initial path information. During the cruise, the UAV's GNSS positioning unit continuously outputs spatial coordinate information, providing a basis for subsequent spatiotemporal alignment and correlation.

[0024] Acquiring Visible Light Images of the Stockyard: During cruise, the multispectral camera, rigidly mounted on the UAV, maintains a relatively stable attitude with the aircraft in a fixed installation manner. The multispectral camera images the target area at each waypoint or within a preset trigger interval between waypoints, acquiring visible light images of the stockyard. To reduce the impact of flight vibrations and changes in aircraft attitude on image clarity, the multispectral camera and the UAV flight control system are linked via a trigger synchronization line, ensuring that the image exposure time matches the moment when the UAV's attitude is relatively stable. Furthermore, the visible light image is embedded with the multispectral camera's internal time stamp during acquisition, and a field is reserved in the image metadata for writing the spatial coordinate information generated by the UAV's GNSS positioning unit.

[0025] Infrared image and terrain point cloud data acquisition: Synchronously with the acquisition of visible light images of the stockpile, the rigidly mounted lidar on the UAV scans the target area during its cruise, outputting terrain point cloud data in real time. The lidar's scanning frequency is coordinated with the UAV's flight speed to ensure that the generated terrain point cloud data spatially corresponds completely to the stockpile outline. Infrared image acquisition is achieved through synchronous triggering with the lidar: when the lidar begins a complete scan, the infrared imaging unit synchronously acquires infrared images, ensuring that the infrared images and terrain point cloud data are consistent in the temporal dimension. To ensure the accuracy of the infrared images in subsequent geometric deformation correction, the raw infrared images acquired by the infrared imaging unit are also pre-loaded with metadata structures that can write spatial coordinate information and a unified timestamp.

[0026] Receiving Leachate Monitoring Data: While the UAV conducts aerial patrols, the ground-based leachate sensor network deployed in the storage yard continues to operate. Multiple leachate sensors report leachate monitoring data at predetermined sampling intervals, including leachate concentration and the spatial location identifier of each node. The multi-source monitoring data acquisition module receives this leachate monitoring data in real time via a wireless communication link and temporarily caches it in the ground station or the UAV's onboard edge computing unit. During caching, each leachate monitoring data piece is appended with the time of reception for subsequent calibration using a unified timestamp based on a high-precision clock source.

[0027] Generating a multi-source monitoring dataset: After completing the above acquisition and reception, the multi-source monitoring data acquisition module performs spatiotemporal alignment and correlation on the visible light images, infrared images, topographic point cloud data, and leachate monitoring data of the stockpile, thereby generating a multi-source monitoring dataset containing timestamps and spatial coordinate information. This process specifically includes: 1. Adding Spatial Coordinate Information: Using the spatial coordinate information generated by the UAV's GNSS positioning unit, corresponding spatial coordinate information is written to the visible light image, infrared image, and terrain point cloud data of the storage yard. For image data (visible light image and infrared image of the storage yard), the spatial coordinate information can be recorded as the coordinates of the image's center point, while retaining the UAV's attitude information at the moment of image exposure. For terrain point cloud data, the spatial coordinate information is directly embedded in the attributes of each point cloud point. In this way, the three types of data achieve spatial reference unification.

[0028] 2. Generating a leachate concentration distribution raster map: Given the dispersed distribution of leachate sensors within the stockpile, a spatial interpolation algorithm is needed to generate a leachate concentration distribution raster map that matches the image resolution. In this implementation, the spatial interpolation algorithm is not limited to a specific mathematical form; a suitable interpolation strategy can be selected based on factors such as the stockpile's topographic relief and sensor density. After interpolation, each raster pixel corresponds to a pixel in the visible light or infrared image of the stockpile in geographic coordinates, ensuring the leachate concentration distribution raster map has the same spatial resolution as the image data.

[0029] 3. Adding a unified timestamp and establishing a four-dimensional index relationship: Based on a high-precision clock source, a unified timestamp is added to the visible light image, infrared image, topographic point cloud data, and leachate concentration distribution raster map of the storage yard. The unified timestamp is obtained by correcting the internal clock deviation of each data acquisition device, aligning different data sources on the time axis. Subsequently, a four-dimensional index relationship containing spatial coordinate information and the unified timestamp is constructed. Each data record is organized under the two-dimensional key value of "spatial coordinate information - unified timestamp", thus forming a centralized "space - time - type - content" four-dimensional retrieval structure for multi-source monitoring data. This four-dimensional index relationship ensures that subsequent hazard identification, dynamic patrol path generation, and execution modules can quickly locate the corresponding multi-source information using any time slice or spatial region as the search condition.

[0030] Example: Taking a single patrol mission as an example, the initial path covers the main storage area from the east to the west side of the storage yard. After takeoff, the UAV cruises along the initial path, with multispectral cameras acquiring visible light images of the storage yard between each adjacent waypoint, and lidar simultaneously outputting terrain point cloud data and triggering infrared image acquisition. Meanwhile, the ground-based leachate sensor network continuously uploads leachate monitoring data. After the mission, the multi-source monitoring data acquisition module writes all data into a unified multi-source monitoring dataset: each visible light image of the storage yard carries corresponding spatial coordinate information and a unified timestamp; the terrain point cloud data carries spatial coordinate information and a unified timestamp at the point level; and the leachate monitoring data, after being processed by a spatial interpolation algorithm to generate a leachate concentration distribution raster map, is matched to the image resolution and assigned a unified timestamp. The resulting multi-source monitoring dataset is organized using a four-dimensional index relationship, enabling the subsequent multi-modal hazard collaborative identification module to call corresponding images, point clouds, and leachate concentration information by region and time window, thus laying a data foundation for the accurate output of hazard types and spatial coordinate sets.

[0031] The yard feature enhancement processing module includes: Acquiring Multi-Source Monitoring Dataset: The stockpile feature enhancement processing module first receives the multi-source monitoring dataset from the multi-source monitoring data acquisition module. This dataset contains visible light images, infrared images, topographic point cloud data, and leachate monitoring data of the stockpile. Upon receipt, the data integrity is verified, including: 1. Check whether the timestamps of the visible light images and infrared images of the stockpile are continuous and whether the spatial coordinate information is complete; 2. Check whether the terrain point cloud data covers the boundary of the area corresponding to the image; 3. Check whether there are valid records of leachate monitoring data for the corresponding time period.

[0032] After verification, an internal data reference relationship is established based on the unified timestamp and spatial coordinate information in the multi-source monitoring dataset, so that any image frame can directly retrieve the topographic point cloud data and the corresponding leachate monitoring data at the same spatiotemporal location.

[0033] Establishing a dust scattering model: Atmospheric optical analysis was performed on the visible light images of the stockpile from the multi-source monitoring dataset to establish a dust scattering model. The specific process includes: 1. Optical background separation: Select a reference area with stable brightness characteristics in the visible light image of the stockpile, such as the sky background or the shadow area of ​​the stockpile, as the atmospheric optical path reference to distinguish between directly reflected light and light scattered by dust.

[0034] 2. Dust feature extraction: Statistical analysis of different brightness gradient regions in the visible light image of the stockpile was performed to analyze the brightness attenuation trend and hue shift characteristics, and optical response patterns related to dust concentration and particle size distribution were extracted.

[0035] 3. Construction of Dust Scattering Model: Based on the aforementioned optical response modes, a dust scattering model characterizing dust concentration and particle size distribution is constructed. This model uses pixels or pixel blocks as basic units and establishes a mapping relationship from observed brightness to actual reflected brightness by describing factors such as the proportion of scattered light and the proportion of background light intensity.

[0036] Example: During a typical cruise, a distinct dust dispersion band exists above the stack. In the visible light image of the stack site, a clear area surrounding the dispersion band is selected as a reference area. By comparing the brightness and hue differences between the clear area and the dust area, the key components in the dust scattering model used to describe non-uniform dust disturbance are obtained.

[0037] Optical dust scattering compensation is performed: An adaptive transmittance correction matrix is ​​generated using the established dust scattering model. This matrix provides a transmittance correction value for each pixel (or pixel block) in the visible light image of the stockpile. The compensation approach is as follows: 1. Transmittance estimation: Based on the dust scattering model, infer the degree of dust scattering influence on the current pixel and estimate the transmittance correction value of the pixel.

[0038] 2. Inversion calculation: The transmittance correction value is applied to the visible light image of the stockpile to perform inversion calculation, restore the true reflectance brightness weakened by dust scattering, and eliminate non-uniform dust interference.

[0039] 3. Edge consistency maintenance: To avoid blurring of target edges due to overcompensation, local smoothing and sharpening balance processing is performed on the edge areas of the image after transmittance correction, so that the compensation effect transitions smoothly at the texture boundary.

[0040] The final output is a dust-compensated visible light image, which serves as one of the inputs for subsequent integration.

[0041] Constructing a curvature mapping map of the pile surface: Three-dimensional surface reconstruction and differential geometric calculations are performed on the terrain point cloud data from the multi-source monitoring dataset to generate a curvature mapping map of the pile surface. The specific process includes: 1. Surface Reconstruction: First, noise filtering and sparse hole filling are performed on the terrain point cloud data, and then the surface of the pile is reconstructed by triangular mesh or other surface fitting methods.

[0042] 2. Differential geometry calculation: Calculate local concavity and convexity features on the reconstructed surface and extract curvature indices related to the surface curvature. Encode these curvature indices in raster or image form to obtain a curvature mapping map of the stockpile surface, so that each location corresponds to a value reflecting the local concavity and convexity of the stockpile surface.

[0043] 3. Spatial alignment: Align the surface curvature mapping of the stack with the visible light and infrared images of the stack site in terms of spatial coordinate information for subsequent unified integration.

[0044] Example: For the sharp protrusions formed at the top of the pile, the curvature map shows high curvature values; while in the gently transitioning areas at the edges of the pile, the curvature values ​​are relatively low. This difference will be directly used in the calculation of pixel position offsets in subsequent infrared image geometric deformation correction.

[0045] Perform infrared image geometric deformation correction: Based on the generated surface curvature map of the pile, calculate the pixel position offset of the infrared images in the multi-source monitoring dataset caused by the surface curvature of the pile, and perform geometric transformation operations, including: 1. Pixel position offset calculation: Based on the curvature mapping of the pile surface and combined with the projection geometry of the infrared image, the difference between the actual three-dimensional position of each pixel in the infrared image and the theoretical pixel coordinates that should fall on the image plane is inferred, forming a pixel position offset field.

[0046] 2. Geometric Transformation Operation: The pixel position offset field is applied to the infrared image to perform geometric deformation correction. The original radiance values ​​of the infrared image are preserved during correction to avoid introducing radiance intensity errors during pixel repositioning.

[0047] 3. Local consistency constraint: To avoid local stretching or compression distortion caused by deformation correction, a neighborhood consistency constraint is added to the geometric transformation operation to smooth the local area, so that the infrared image after deformation correction is continuous and has sharp edges.

[0048] After completing the above processing, the deformed infrared image is obtained.

[0049] Generate an enhanced feature map set: Integrate the obtained dust-compensated visible light image, deformation-corrected infrared image, and generated pile surface curvature map to form a unified enhanced feature map set. Specifically, this includes: 1. Resolution and coordinate unification: Check whether the three types of data are completely consistent in spatial coordinate information. If there is a resolution difference, use the visible light image of the stockpile as a reference and align the other data to the same spatial resolution through interpolation or resampling.

[0050] 2. Channel integration: The visible light image after dust compensation, the infrared image after deformation correction, and the curvature map of the pile surface are stacked according to the channel dimension, and the data type description corresponding to each channel is recorded in the enhanced feature map set.

[0051] 3. Metadata Improvement: To enhance the feature map set, unified timestamps, spatial coordinate information, and index relationships with the original multi-source monitoring dataset are added to ensure that the multimodal hazard collaborative identification module can quickly retrieve the required features according to the hazard type and spatial coordinate set.

[0052] Example: In a certain generation process, the enhanced feature map set contains three core channels: a dust-compensated visible light image channel to provide clear texture; a deformation-corrected infrared image channel to reflect internal thermal anomalies of the reactor body; and a reactor body surface curvature mapping channel to highlight locations of geometric abrupt changes in the reactor body. When the multimodal hazard collaborative identification module detects an abnormal abrupt change in leachate concentration in a certain area, it can call the corresponding enhanced feature map set for that area to achieve accurate output of the hazard type and spatial coordinate set.

[0053] The multimodal hazard collaborative identification module includes: Acquiring Enhanced Feature Atlas and Leachate Monitoring Data: First, the enhanced feature atlas is received from the stockpile feature enhancement processing module, and simultaneously, leachate monitoring data is received from the multi-source monitoring data acquisition module. To ensure the accuracy of subsequent collaborative processing, consistency checks and reference index construction are performed on both types of data, including: 1. Consistency check: Verify whether the timestamps and spatial coordinate information of the visible light image after dust compensation, the infrared image after deformation correction, and the surface curvature mapping map of the stack in the enhanced feature map set correspond to the leachate monitoring data.

[0054] 2. Reference Index Construction: Using "spatial coordinate information - unified timestamp" as the search key, a one-to-one or one-to-many reference relationship is established between the enhanced feature map set and the leachate monitoring data, ensuring that any subsequent spatial location can quickly locate multimodal data under the same time slice.

[0055] Example: If the timestamp corresponding to a deformation-corrected infrared image in the enhanced feature map set is... If the spatial coordinate information covers the northwest corner of the storage yard, then the timestamp should be matched with the leachate monitoring data that are located in the same area and have a sampling time close to that of the leachate monitoring data. An index is created on the records, which can then be directly accessed later.

[0056] Configure deformable convolutional unit parameters: Extract the curvature map of the pile surface from the enhanced feature map set, and input the curvature map of the pile surface into the deformable convolutional unit of the hazard identification model. This is used to dynamically adjust the kernel offset parameters of the deformable convolutional unit, so that the convolutional kernel adapts to the geometric features of the pile surface. The specific process includes: 1. Curvature-Offset Mapping Establishment: Based on the curvature values ​​at various locations in the curvature map of the heap surface, an offset reference map is generated to guide the offset direction and magnitude of the convolution kernel. Regions with high curvature guide the convolution kernel to focus more on local concave and convex details, while regions with low curvature maintain a smaller offset to maintain global consistency.

[0057] 2. Convolution kernel offset parameter update: Input the offset reference map into the deformable convolution unit to generate convolution kernel offset parameters, and call them in real time during model forward inference to make the convolution kernel present differentiated sampling shapes at different spatial locations.

[0058] 3. Parameter validity verification: Constrain the continuity of the convolution kernel offset parameters to prevent excessive offset from causing feature extraction misalignment.

[0059] Example: In locations where there are steep slopes on the surface of a pile, the curvature mapping of the pile surface shows obvious abrupt changes. Based on this, deformable convolutional units shift their convolution kernel sampling points towards the slope normal direction, thereby improving the ability to capture potential textures in this area.

[0060] Initial image hazard identification is performed: After configuring the deformable convolutional unit parameters, the dust-compensated visible light image and the deformation-corrected infrared image from the enhanced feature set are input into the hazard identification model for initial image hazard identification, generating preliminary hazard identification results, including: 1. Multi-channel input fusion: During the input stage, the model receives the visible light image after dust compensation and the infrared image after deformation correction, and processes them together with the curvature mapping map of the pile surface at the feature level, so that the hazard identification takes into account both visual texture and thermal anomalies and geometric features.

[0061] 2. Potential Hazard Area Extraction: The model outputs preliminarily identified potential hazard areas. These potential hazard areas are indexed by spatial coordinate information and can be marked as areas requiring further confirmation.

[0062] 3. Preliminary hazard type determination: The model also provides preliminary hazard types, such as signs of pile slippage, surface cracks, local thermal anomalies, etc. (These are just example types; specific types are defined based on the training data).

[0063] Example: In an initial image hazard identification, the model identified a potential hazard area on the southern slope of the stockpile, with the preliminary hazard type being "signs of stockpile slippage." This area will subsequently be the focus of monitoring and high-precision identification of abnormal leachate changes.

[0064] Monitoring for abrupt changes in leachate concentration: Real-time analysis of leachate monitoring data, calculation of the spatiotemporal gradient of leachate concentration at each monitoring point. When an abrupt change in leachate concentration is detected in a certain area, that area is marked as an abrupt leachate concentration region, and its spatial coordinates are obtained.

[0065] When analyzing leachate monitoring data in real time, a three-step judgment mechanism is introduced: "spatiotemporal change rate - comparison with historical normal fluctuation range - threshold determination." This ensures that the identification of abnormal changes in leachate has an objective benchmark and repeatability, including: 1. Calculate the spatiotemporal change rate of leachate monitoring data: For each monitoring point, calculate the concentration change rate between two adjacent sampling times, and combine this with the changes of surrounding monitoring points in the same time slice to form a comprehensive spatiotemporal change rate. To maintain consistency with the "spatiotemporal gradient" in the claims, this embodiment considers the spatiotemporal change rate as a quantitative expression of the "spatiotemporal gradient".

[0066] 2. Compare with historical concentration variation patterns at this location: The system automatically calculates the concentration change rate distribution at this location during its previous stable operating phases, forming the historical normal fluctuation range for this location. The historical normal fluctuation range can be obtained by fitting the mean and fluctuation range of long-term historical data (e.g., represented by the mean ± standard deviation range), and is stored as a baseline archive within the system for a long period.

[0067] 3. Threshold determination: When the spatiotemporal change rate calculated in real time exceeds 30% of the upper limit of the historical normal fluctuation range of the point, it is immediately determined to be an abnormal change, and the area where the monitoring point is located is marked as an abnormal leachate area. At the same time, its spatial coordinates are extracted for subsequent high-precision identification.

[0068] Triggering and executing high-precision identification: For marked leachate anomaly areas, high-resolution image data with corresponding spatial coordinates is extracted from the enhanced feature map set and input into the hazard identification model for higher computational accuracy or higher resolution image analysis processing to generate high-precision hazard identification results, including: 1. High-resolution image data extraction: Based on the spatial coordinates of the abnormal leachate area, the corresponding local area image is cropped from the enhanced feature map set, and the boundary range can be appropriately expanded during cropping to ensure the integrity of the context information.

[0069] 2. Upgraded identification strategy: In the process of high-precision hazard identification, deeper feature extraction units or denser deformable convolutional unit sampling strategies can be enabled in the hazard identification model to improve identification accuracy.

[0070] 3. High-precision hazard identification results output: Outputs high-precision hazard identification results, clearly identifying the hazard type in the area and providing a relatively accurate set of spatial coordinates. This result will be integrated with the preliminary hazard identification results for conflict resolution.

[0071] Example: In the area of ​​abnormal leachate, high-resolution image data is extracted and high-precision hazard identification is performed. Finally, the initial hazard type "pile slip signs" is refined into "pile slip signs accompanied by local crack expansion", and the precise set of spatial coordinates of the crack start and end positions is determined.

[0072] The system integrates and outputs the identification results: It fuses and resolves conflicts between the preliminary and high-precision hazard identification results, combines this with anomaly information from leachate monitoring data, and ultimately determines the hazard type and its precise spatial coordinate set for each hazard area. The results are then output as part of the multimodal hazard collaborative identification module, including: 1. Result Fusion Strategy: Based on spatial coordinate information, the preliminary hazard identification results are overlaid with the high-precision hazard identification results. If they match, the hazard identification is directly confirmed; if there are discrepancies, the high-precision hazard identification results take precedence, while retaining hazard information from the preliminary hazard identification results that was not covered but has independent regional characteristics.

[0073] 2. Conflict resolution mechanism: When two types of results conflict in terms of hazard type, the decision is made by considering the degree of anomaly in leachate monitoring data, the local features of the surface curvature mapping of the pile body, and the recognition credibility ranking within the model, and a unique hazard type is output.

[0074] 3. Spatial coordinate set refinement: During the fusion process, the spatial coordinate set is deredundant and the boundaries are smoothed to ensure that the output spatial coordinate set can not only represent the true range of the hidden danger area, but also facilitate direct calling by the subsequent dynamic cruise path generation and execution module.

[0075] 4. Results Organization and Output: The final results are classified according to the type of hazard and are accompanied by the corresponding set of spatial coordinates and source data index, forming a structured output for the dynamic cruise path generation and execution module to continue working based on the regional hazard density distribution map.

[0076] Example: During the fusion process, if the area marked by the initial hazard identification result is slightly larger than the area given by the high-precision hazard identification result, the boundaries of the two are aligned and cropped to output a spatial coordinate set primarily based on the high-precision hazard identification result. Simultaneously, by combining the abnormal area labels from leachate monitoring data, the importance of the hazard area is confirmed, providing a basis for subsequent route planning. The dynamic cruise path generation and execution module includes: Receiving Hazard Identification Results: First, receive the hazard type and spatial coordinate set output from the multimodal hazard collaborative identification module. To ensure the accuracy of subsequent calculations, perform the following processing: 1. Consistency verification: Check whether the timestamp and spatial coordinate information of each record in the hazard type and spatial coordinate set are complete, and compare them with the previous round of monitoring data to confirm that there is no duplication or omission.

[0077] 2. Structured organization: Aggregate the hazard types and spatial coordinate sets according to the spatial coordinate information to form a data structure that is easy to process by grid. At the same time, mark the location of the preset key monitoring points in the set so that they can be given priority in the path planning stage.

[0078] Example: If the set of hazard types and spatial coordinates contains multiple signs of stack slippage and local thermal anomalies, archive these hazard types and their corresponding spatial coordinate information separately, and mark which coordinate points are also located near the preset key monitoring points.

[0079] Calculation of regional hazard density distribution map: After data processing, based on the received spatial coordinate set, a kernel density estimation algorithm is used to calculate the regional hazard density distribution map of the entire stockpile area, including: 1. Grid division: The storage yard area is discretized according to a uniform grid form, and each grid unit corresponds to a fixed range of the actual location of the storage yard.

[0080] 2. Kernel density estimation algorithm execution: Project each hazard point in the spatial coordinate set onto the grid cell, and use the kernel density estimation algorithm to spread its influence range in the grid space to obtain a continuous distribution of the number of hazards per unit area.

[0081] 3. Result verification and smoothing: Perform boundary consistency checks on the regional hazard density distribution map to avoid abnormal density abrupt changes at the edge of the storage yard, and maintain a natural overall transition through appropriate smoothing.

[0082] Example: After processing by the kernel density estimation algorithm, a high-value area is formed on the northeast side of the storage yard, indicating that the number of potential hazards per unit area in this area is significantly higher than that in the surrounding areas, and subsequent path planning will focus on covering this area.

[0083] Extracting density gradient information: Spatial gradient calculation is performed on the regional hazard density distribution map to obtain density gradient information, which characterizes the rate and direction of change in hazard density, including: 1. Gradient field construction: Perform differential operation on the density values ​​of adjacent grid cells in the grid space to obtain the density gradient vector of each grid cell. The direction of the vector indicates the direction in which the density of the hidden danger increases the fastest.

[0084] 2. Noise suppression: To avoid local isolated outliers affecting the overall gradient trend, local consistency constraints are imposed on the gradient field to ensure that the gradient direction is relatively smooth in space and consistent with the actual distribution of potential hazards.

[0085] 3. Key area identification: Contiguous areas with high gradient intensity in the density gradient information are marked as key areas of interest, serving as an important reference for subsequent dynamic cruise path planning.

[0086] Example: In the regional hazard density distribution map, a continuous transition zone with density increasing from low to high appears in the southwest corner of the storage yard, with the gradient direction pointing towards the center of the density peak. This transition zone should be marked as an area that needs to be traversed along the gradient direction in order to promptly capture the potential hazard spread trend.

[0087] Generate a dynamic cruise path: Based on density gradient information, a gradient ascent algorithm is used to plan a dynamic cruise path, ensuring that from the current drone position, it prioritizes traversing areas with high hazard density and covers preset key monitoring points, including: 1. Starting point determination: The current location of the drone is used as the starting point of the dynamic cruise path to ensure that the path planning is consistent with the real-time status of the drone.

[0088] 2. Gradient Ascent Algorithm Execution: Iteratively search for the next path along the direction indicated by the density gradient information, so that the path passes through the grid cells with higher regional hazard density as much as possible, while avoiding repeated coverage of areas that have been fully monitored.

[0089] 3. Preset key monitoring point coverage: During the path search process, priority constraints are set for preset key monitoring points to ensure that the dynamic cruise path covers all preset key monitoring points within a reasonable path length.

[0090] 4. Path Feasibility Verification: The generated dynamic cruise path is checked for feasibility, including track continuity, turning smoothness, and avoidance of no-fly zones in storage yards, to ensure that the path can be directly converted into flight commands at the execution level.

[0091] Example: If the initial path obtained by the gradient ascent algorithm does not cross a certain preset key monitoring point, a transition path is inserted to include the preset key monitoring point without disrupting the main gradient direction; for areas that are already densely covered, the path will automatically detour to save cruising time.

[0092] Update UAV flight commands: Convert the generated dynamic cruise path into a sequence of waypoints and flight parameters recognizable by the UAV flight control system, forming updated UAV flight commands, including: 1. Waypoint sequence generation: Discretize the continuous spatial location nodes on the dynamic cruise path into a waypoint sequence to ensure that the distance between adjacent waypoints matches the UAV's attitude adjustment capability.

[0093] 2. Flight parameter settings: Assign flight parameters that match the characteristics of the path segment to each waypoint, such as flight altitude and cruise speed (the specific values ​​can be determined by the flight control strategy), and keep them consistent with the format that the UAV flight control system can recognize.

[0094] 3. Command Packaging and Verification: Before outputting updated UAV flight commands, the consistency and integrity of the waypoint sequence and flight parameters are verified to generate command packages that can be directly issued.

[0095] Example: If a certain route requires the UAV to stably acquire high-resolution images at low speed, a lower speed will be set in the flight parameters of that waypoint, and the flight altitude will be kept constant to facilitate the synchronous triggering of the multi-source monitoring data acquisition module.

[0096] Execute the next round of monitoring: Send the generated updated UAV flight commands to the UAV flight control system, control the UAV to execute the next round of monitoring tasks according to the dynamic cruise path, and trigger the multi-source monitoring data acquisition module to start a new round of data acquisition, including: 1. Command Issuance and Confirmation: The updated UAV flight commands are sent to the UAV flight control system via the communication link, and the command reception confirmation information is received from the flight control system.

[0097] 2. Path execution process monitoring: During the process of the UAV executing the dynamic cruise path, continuously monitor the UAV's status feedback to ensure that adjustments can be made in a timely manner if any unexpected situations occur during path execution.

[0098] 3. Data Acquisition Trigger: When the UAV enters the critical segment of the dynamic cruise path, the multi-source monitoring data acquisition module is triggered simultaneously to ensure that the acquisition of a new round of visible light images, infrared images, terrain point cloud data and leachate monitoring data of the stockpile is consistent with the dynamic cruise path in time and space.

[0099] Example: When the drone flies along the dynamic cruise path to the vicinity of the peak of the regional hazard density distribution map, it automatically triggers the multi-source monitoring data acquisition module to collect data at high frequency, so as to provide more intensive data support for the next round of multi-modal hazard collaborative identification module.

[0100] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A remote patrol and hazard identification system for solid waste storage yards based on AI technology, characterized in that, include: Multi-source monitoring data acquisition module: Controls the UAV equipped with a multispectral camera and lidar to cruise along the initial path, simultaneously acquiring visible light images, infrared images and terrain point cloud data of the stockpile, and receiving leachate monitoring data uploaded by the ground leachate sensor network to generate a multi-source monitoring dataset; The stockpile feature enhancement processing module performs optical dust scattering compensation on the visible light images in the multi-source monitoring dataset, and uses an adaptive transmittance correction matrix to eliminate non-uniform dust interference; it constructs a surface curvature mapping map of the stockpile based on the lidar point cloud data, and performs geometric deformation correction on the infrared images according to the curvature mapping map, and outputs an enhanced feature map set. Multimodal Hazard Collaborative Identification Module: The enhanced feature map set is input into a pre-trained hazard identification model, which includes deformable convolutional units, wherein the kernel parameters of the deformable convolutional units are dynamically adjusted according to the input curvature map; the leachate monitoring data is fused synchronously, and when an abnormal change in leachate concentration is detected in a certain area, the identification processing of the image of that area is triggered, and the hazard type and spatial coordinate set are output; Dynamic cruise path generation and execution module: Calculates the regional hazard density distribution map based on the hazard type and spatial coordinate set, generates a dynamic cruise path that prioritizes coverage of high-density hazard areas based on the gradient information of the hazard density distribution map, updates the UAV flight instructions, and controls the UAV to execute the next round of monitoring tasks.

2. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 1, characterized in that, The multi-source monitoring data acquisition module includes: Control the drone to cruise along the initial path: Control the drone to fly according to the preset initial path; Acquiring visible light images of the stockpile: During the cruise, the UAV uses a rigidly mounted multispectral camera to acquire visible light images of the target area of ​​the stockpile. Infrared image and terrain point cloud data acquisition: The UAV is equipped with a rigid lidar to simultaneously acquire infrared images and terrain point cloud data of the target area during the cruise. Receive leachate monitoring data: Receive leachate monitoring data uploaded in real time from the ground leachate sensor network deployed in the stockyard; Generate a multi-source monitoring dataset: The collected visible light images, infrared images, and topographic point cloud data of the storage yard, as well as the received leachate monitoring data, are spatiotemporally aligned and correlated to generate a multi-source monitoring dataset containing timestamps and spatial coordinate information.

3. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 2, characterized in that, The stockpile feature enhancement processing module includes: Acquire multi-source monitoring dataset: Receive multi-source monitoring dataset from the multi-source monitoring data acquisition module. This dataset includes visible light images, infrared images, topographic point cloud data, and leachate monitoring data of the stockyard. Establish a dust scattering model: Perform atmospheric optical analysis on the visible light images of the stockpile in the multi-source monitoring dataset to establish a dust scattering model characterizing dust concentration and particle size distribution; Perform optical dust scattering compensation: Using the established dust scattering model, generate an adaptive transmittance correction matrix, and apply the adaptive transmittance correction matrix to perform inversion calculation on the visible light image of the stockpile to eliminate non-uniform dust interference and obtain a dust-compensated visible light image. Constructing a curvature mapping map of the pile surface: Perform three-dimensional surface reconstruction and differential geometry calculation on the terrain point cloud data in the multi-source monitoring dataset to generate a curvature mapping map of the pile surface that characterizes the local concave and convex features of the pile surface; Perform infrared image geometric deformation correction: Based on the generated surface curvature mapping of the pile, calculate the pixel position offset of the infrared image in the multi-source monitoring dataset caused by the surface curvature of the pile, perform geometric transformation operation, complete geometric deformation correction, and obtain the deformed infrared image. Generate an enhanced feature map set: Integrate the obtained dust-compensated visible light image, deformation-corrected infrared image, and pile surface curvature map to form a unified enhanced feature map set.

4. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 3, characterized in that, The multimodal hazard collaborative identification module includes: Acquire enhanced feature maps and leachate monitoring data: Receive enhanced feature maps from the stockpile feature enhancement processing module and leachate monitoring data from the multi-source monitoring data acquisition module; Configure deformable convolutional unit parameters: extract the curvature map of the pile surface from the enhanced feature map set, and input the curvature map of the pile surface into the deformable convolutional unit of the hazard identification model, dynamically adjust the convolutional kernel offset parameter of the deformable convolutional unit, so that the convolutional kernel adapts to the geometric features of the pile surface; Perform initial image hazard identification: Input the dust-compensated visible light image and the geometrically deformed infrared image from the enhanced feature map set into the configured hazard identification model for initial identification, and generate preliminary hazard identification results, including the preliminary identified potential hazard areas and their preliminary hazard types.

5. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 4, characterized in that, The multimodal hazard collaborative identification module also includes: Monitoring abnormal changes in leachate concentration: Analyze the leachate monitoring data in real time, calculate the spatiotemporal gradient of leachate concentration at each monitoring point, and when an abnormal change in leachate concentration is detected in a certain area, mark the area as an abnormal leachate area and obtain its spatial coordinates. Trigger and execute hazard identification: For marked abnormal leachate areas, extract image data with corresponding spatial coordinates from the enhanced feature map set, input it into the hazard identification model for image analysis and processing, and generate hazard identification results; The preliminary hazard identification results are fused and conflict resolution is performed with the obtained hazard identification results. Combined with the abnormal information of the leachate monitoring data, the hazard type and its spatial coordinate set of each hazard area are finally determined.

6. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 5, characterized in that, The dynamic cruise path generation and execution module includes: Receive hazard identification results: Receive the hazard type and spatial coordinate set output by the multimodal hazard collaborative identification module; Calculate the regional hazard density distribution map: Based on the received set of spatial coordinates, the kernel density estimation algorithm is used to calculate the regional hazard density distribution map of the entire stockpile area. This map represents the number of hazards per unit area at different locations in grid form. Extracting density gradient information: Spatial gradient calculation is performed on the generated regional hazard density distribution map to obtain density gradient information that characterizes the rate and direction of change of hazard density.

7. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 6, characterized in that, The dynamic cruise path generation and execution module also includes: Generate a dynamic cruise path: Based on the acquired density gradient information, a gradient ascent algorithm is used to plan a dynamic cruise path that starts from the current drone position, prioritizes passing through high-risk density areas, and covers preset key monitoring points. Update UAV flight commands: Convert the generated dynamic cruise path into a waypoint sequence and flight parameters that the UAV flight control system can recognize, and generate updated UAV flight commands; Execute the next round of monitoring: Send updated UAV flight commands to the UAV flight control system, control the UAV to execute the next round of monitoring tasks according to the dynamic cruise path, and trigger the multi-source monitoring data acquisition module to start a new round of data acquisition.

8. The AI-based remote patrol and hazard identification system for solid waste storage yards according to claim 2, characterized in that, The spatiotemporal alignment and correlation of the generated multi-source monitoring dataset includes: Add spatial coordinate information generated by the UAV GNSS positioning module to the visible light image, infrared image, and terrain point cloud data of the stockyard; The leachate monitoring data is used to generate a leachate concentration distribution raster map that matches the image resolution using a spatial interpolation algorithm; Based on the clock source, a unified timestamp is added to the visible light image, infrared image, terrain point cloud data, and leachate concentration distribution raster map of the stockpile, and a four-dimensional index relationship is established.

9. A remote patrol and hazard identification system for solid waste storage yards based on AI technology as described in claim 5, characterized in that, The determination of the abnormal mutation is carried out in the following manner: Calculate the spatiotemporal rate of change of the leachate monitoring data; Compare the spatiotemporal rate of change with the historical concentration change pattern at that location; When the rate of change exceeds 30% of the upper limit of the historical normal fluctuation range, it is judged as an abnormal mutation.

Images