Low-altitude unmanned aerial vehicle intelligent inspection system and method based on artificial intelligence

Abstract

The invention discloses a low-altitude unmanned aerial vehicle intelligent inspection system and method based on artificial intelligence, and relates to the technical field of unmanned aerial vehicles. Based on an abnormal position library and an abnormal measure library, a space inspection priority thermodynamic diagram and a measure inspection priority list are generated, an inspection comprehensive score is calculated, and a strategy selection model is constructed; collecting real-time data information, and generating a real-time space inspection priority thermodynamic diagram and a real-time measure inspection priority list; according to a strategy selection model, a preliminary inspection strategy type is generated, a plurality of inspection routes are generated through a path planning algorithm, a comprehensive priority score of each inspection route is calculated, an optimal inspection scheme is determined, and the optimal inspection scheme is issued to the unmanned aerial vehicle executing the task, so that the automation level of unmanned aerial vehicle inspection is greatly improved, and energy waste and time waste are reduced.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to an intelligent inspection system and method for low-altitude UAVs based on artificial intelligence. Background Technology

[0002] Most existing agricultural drone inspection methods still rely on preset flight routes or manually planned paths. Manual planning often requires setting waypoints one by one according to the boundaries of the plots and the distribution of crops. This is not only labor-intensive, but also prone to unreasonable flight route planning due to differences in human experience, resulting in problems such as overlapping coverage or omissions. On the other hand, flight methods based on fixed flight route templates mostly assume that the plots are regular in shape and the crops are at the same height, which is difficult to adapt to the complex environment of farmland, such as irregular plots, differences in crop growth, and interference from obstacles.

[0003] Under the above conditions, drones are prone to problems such as incomplete flight path coverage, excessive flight energy consumption, and insufficient obstacle avoidance response, which in turn affects the complete collection of agricultural information and the efficiency of inspection. In addition, existing solutions usually lack real-time integration with dynamic environmental information, making it difficult to adjust flight paths in a timely manner to cope with changes in wind speed, differences in plant height, or temporary obstacles, resulting in insufficient flexibility and robustness in mission execution.

[0004] Therefore, how to dynamically generate and optimize drone flight paths in complex and ever-changing farmland scenarios, improve flight path coverage and efficiency, and balance obstacle avoidance safety with the ability of multiple drones to work together, has become a technical problem that needs to be solved in the field of intelligent inspection of agricultural drones. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent inspection system and method for low-altitude unmanned aerial vehicles based on artificial intelligence, so as to solve the problems raised in the prior art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: an intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence, the method comprising:

[0007] Step S100: Obtain image information from the inspection record through the drone inspection platform, and perform similarity calculation with standard images to determine abnormal image information;

[0008] Step S200: Collect basic information identifiers of historical inspection records, summarize the historical inspection record set for each basic information identifier, and construct an abnormal location database and an abnormal measure database based on the historical inspection records containing abnormal image information in the historical inspection record set.

[0009] Step S300: Set the training period, generate a spatial inspection priority heatmap and a measure inspection priority list based on the abnormal location library and abnormal measure library, calculate the comprehensive inspection score, and build a strategy selection model;

[0010] Step S400: Collect real-time data information and generate a real-time spatial inspection priority heat map and a real-time measure inspection priority list;

[0011] Step S500: Based on the strategy selection model, generate a preliminary inspection strategy type, generate multiple inspection routes based on the real-time spatial inspection priority heat map and the real-time measure inspection priority list, calculate the comprehensive priority score of each inspection route, determine the optimal inspection plan, and issue it to the UAVs that perform the task.

[0012] Furthermore, step S100 includes:

[0013] Step S101: In the UAV inspection platform, obtain the crops planted in the farmland to be inspected, and number the farmland and crops respectively. Record the field management measures implemented for the crops. Deploy environmental sensors in each farmland to collect the environmental information of the farmland. During the UAV inspection process, collect the flight information and image information of the UAV. Summarize the field management measures, environmental information, flight information and image information to generate an inspection record, and upload the inspection record to the UAV inspection platform.

[0014] Step S102: Obtain inspection records, extract image information from each inspection record, mark the crops in the image information using image recognition technology, and extract the outlines of the marked crops using image segmentation technology to obtain a complete image of the crops;

[0015] Step S103: Collect the time point of each inspection record, obtain the standard image of the crop corresponding to the time point, divide the complete image and the standard image into several standard squares, and calculate the similarity of the standard squares of the complete image and the standard image through a structural similarity algorithm to obtain the average similarity of the image information.

[0016] Step S104: Set an average similarity threshold, mark image information with a similarity value less than the average similarity threshold as abnormal image information, and write the abnormal image information back to the corresponding inspection record;

[0017] By simultaneously collecting field management measures, environmental information, flight information, and image information through the drone inspection platform, and generating a unified inspection record, it is possible to achieve a comprehensive and standardized record of farmland conditions, avoid omissions in manual inspections, and ensure the integrity and traceability of information.

[0018] By using image recognition and image segmentation techniques to mark and extract the contours of crops, complete images of crops can be extracted. Compared with traditional comparison methods based on the whole image, this method can more accurately distinguish crops from the background, improving the accuracy of recognition and subsequent analysis.

[0019] By introducing a structural similarity algorithm, the inspection images are compared with standard images in a grid pattern and the average similarity is calculated. This can quantitatively reflect the growth status and differences of crops, avoid subjective human judgment, and improve the objectivity and automation of anomaly identification.

[0020] Furthermore, step S200 includes:

[0021] Step S201: In the historical inspection record set, collect the farmland number and crop number of the historical inspection record, and construct a basic information identifier by combining the production cycle information of the crop in the farmland. The production cycle information is the complete cycle of the crop from sowing to harvest. Summarize the historical inspection records with the same basic information identifier and in the same production cycle to obtain the historical inspection record set for each basic information identifier.

[0022] Step S202: In the historical inspection record set of a certain basic information identifier, extract the historical inspection records with abnormal image information and set them as historical abnormal records. Collect the inspection time of each historical abnormal record and arrange the historical abnormal records in chronological order of inspection time. Obtain the abnormal location information of each historical abnormal record and construct the abnormal location library of the basic information identifier.

[0023] Step S203: In the historical inspection record set of a certain basic information identifier, improve the data utilization rate of each historical abnormal record: By collecting farmland number, crop number and production cycle information to construct basic information identifier, and summarizing historical inspection records within the same production cycle, the systematic collection of historical inspection data is realized, avoiding the scattered inspection data and improving the data utilization rate.

[0024] Extracting abnormal image information from historical inspection records and building an abnormal location database can intuitively reflect the distribution of abnormalities in time and space, realize the source and location of abnormal information, and facilitate targeted monitoring and management of specific areas of farmland.

[0025] By extracting the field management measures corresponding to historical anomaly records and building an anomaly measure database, subsequent inspections and management can combine historical handling experience to form a reference for measures to deal with anomalies, thereby improving the scientific nature and effectiveness of field management.

[0026] Time-sequencing and location-archiving of abnormal records within the same production cycle helps to discover patterns and regularities in the occurrence of abnormalities, providing data support for early warning of crop diseases and pests, and analysis of environmental stress.

[0027] Furthermore, step S202 involves constructing an anomaly location database, including:

[0028] Step S202-1: Obtain the farmland number from the basic information identifier, retrieve the satellite remote sensing image of the farmland number, perform three-dimensional modeling on the farmland using three-dimensional modeling technology to obtain a three-dimensional farmland model, divide the three-dimensional farmland model into several regions, and establish an independent spatial coordinate system in each region;

[0029] Step S202-2: Collect the geographic location information of the abnormal image information in the historical abnormal records, and combine it with the flight trajectory data of the UAV to determine the area corresponding to the abnormal image information, and calculate the precise spatial coordinates of the abnormal image in the area according to the spatial coordinate system of the area;

[0030] Step S202-3: Count the number of times historical abnormal records appear in a certain area, calculate the frequency of occurrence in the area, collect the anomaly type and average similarity of a certain historical abnormal record in the area, assign a value to each anomaly type, and normalize the frequency of occurrence and average similarity. Calculate the anomaly degree index of the area according to the following formula:

[0031] ;

[0032] Where A represents the regional anomaly index, B represents the normalized frequency of occurrence, and C... b D represents the numerical value of the anomaly type in the b-th historical anomaly record. b Let d represent the normalized average similarity value of the b-th historical anomaly record, d represent the total number of historical anomaly records, and a1 and a2 represent the weights of the frequency of occurrence and the average similarity, respectively.

[0033] Step S202-4: Divide the spatial coordinate system of a certain region into several grids, obtain the grid corresponding to each historical anomaly record, count the number of historical anomaly records in a certain grid as v1, calculate the distribution density of a certain grid Z=v1 / v2, where v2 represents the number of historical anomaly records in the region, normalize the distribution density and calculate it with the regional anomaly degree index to obtain the refined anomaly degree index of each grid, count the historical anomaly records of each grid in each growth stage, divide it by all historical anomaly records in the growth stage to calculate the occurrence frequency, and set it as the weight of the grid in the growth stage. Store the calculated regional anomaly degree index, refined anomaly degree index and weights in different growth stages into the anomaly location database.

[0034] By creating a 3D model of farmland and dividing it into independent spatial coordinate systems, and combining the flight trajectory data of UAVs, the specific areas and precise spatial coordinates of abnormal images are determined, thus realizing the mapping of abnormal information from a two-dimensional plane to a three-dimensional space, significantly improving the accuracy of abnormal location.

[0035] By constructing regional anomaly degree indicators based on multi-dimensional information such as frequency of occurrence, anomaly type and similarity, it is possible to quantitatively characterize anomalies, avoid relying solely on human experience or vague descriptions, and improve the scientificity and comparability of anomaly assessment.

[0036] By further dividing the regional spatial coordinate system into several spatial coordinate ranges and calculating their anomaly distribution density, and then weighting and integrating them with the regional anomaly degree index, a more granular and refined anomaly degree index can be formed, which can intuitively reflect the clustering and distribution characteristics of anomalies within the region.

[0037] Furthermore, step S203 involves constructing an exception handling library, including:

[0038] Step S203-1: Collect information on field management measures corresponding to historical anomaly records, including measure type, implementation time, implementation duration, and implementation amount; and bind the field management measures with historical anomaly records to form a correspondence between anomalies and measures;

[0039] Step S203-2: According to the implementation time sequence, using the inspection time of historical anomaly records as a reference point, extract the field management measures within the preset time window before the anomaly occurred to obtain the sequence of measures related to the anomaly;

[0040] Step S203-3: According to the measure sequence, count the number of times each measure appears in a preset time window, calculate the frequency of each measure, and set the frequency of occurrence as the abnormal correlation degree of the measure. Collect the time interval between the implementation time and the inspection time in each historical abnormal record, obtain the implementation amount of the implemented measure, and normalize the implementation amount. Use the normalized implementation amount and abnormal correlation degree as input and the time interval as output to construct a training dataset. Train the training dataset through a linear regression model to obtain the trained weights, update the linear regression model, and set it as the measure prediction model. Store it in the abnormal measure library.

[0041] By linking historical anomaly records with corresponding field management measures, a correspondence between measures and anomalies can be formed, which can quantify the impact of different measures on anomalies and provide data support for subsequent anomaly prediction and management optimization.

[0042] By extracting the sequence of measures before and after the occurrence of an anomaly, and statistically analyzing the frequency and intensity parameters of each measure, a predictive model for the correlation between anomalies and the intensity of measures is established. This enables subsequent farmland management decisions to be made based on historical data for precise management, avoiding blind implementation of policies.

[0043] By training and updating the training dataset based on a linear regression model, the predictive model for measures can be dynamically optimized, thereby improving the decision support capability in farmland management. Through multi-dimensional analysis of implementation time, intensity, and time interval, the optimal timing and intensity of future management measures can be predicted, further enhancing the scientific nature and timeliness of agricultural management.

[0044] Furthermore, step S300 includes:

[0045] Step S301: Select a consecutive number of days as the training period. Collect environmental information for each day within the training period. Using the anomaly location database, map the refined anomaly level index of a certain basic information identifier to the spatial coordinate system of each area of ​​farmland to generate a static anomaly probability heatmap. Obtain the time interval between each anomaly record and the current day. Calculate the dynamic score of the grid according to the following formula:

[0046] ;

[0047] Where E represents the dynamic score, F g Let t represent the refined anomaly level index corresponding to the g-th anomaly record, u represent the decay rate parameter, and t represent the decay rate parameter. gLet f represent the time interval of the g-th abnormal record, and let f represent the total number of abnormal records. Obtain the current crop production cycle information, extract the crop growth stage of the day, and multiply the weight of each grid in the growth stage with the dynamic score to calculate the priority score of each grid on the day. Map the priority score to the static anomaly probability heat map to obtain the spatial inspection priority heat map.

[0048] Step S302: Obtain the field management measures within the time window prior to the current day, collect the implementation amount of each measure, perform normalization calculations, input the data into the measure prediction model, calculate the prediction time interval, and calculate the urgency score according to the following formula:

[0049] ;

[0050] Where K represents the urgency score, j1 represents the time interval between the current day and the implemented measures, and j2 represents the predicted time interval. The implemented measures are sorted from high to low according to the urgency score to generate a list of measures inspection priorities.

[0051] Step S303: Simulate multiple predefined inspection strategies daily, assign a number to each strategy, acquire UAV flight information, extract the number of new anomalies and flight parameters during inspection, perform normalization calculations, and calculate the comprehensive inspection score according to the following formula:

[0052] ;

[0053] Where P represents the overall inspection score, S represents the number of new anomalies in the inspection after normalization, Q represents the normalized average flight power, N represents the normalized flight time, and Y represents the normalized flight time. n h3 represents the normalized duration of the nth flight phase. n Let h1 represent the weight of the nth flight phase, m represent the total number of flight phases, and h2 represent the weights of the number of newly discovered anomalies and the average flight power, respectively.

[0054] Step S304: Obtain the environmental information of the day and perform normalization calculation; collect the growth stage corresponding to the crop of the day and assign a value to each growth stage; extract the priority score of each grid in the spatial inspection priority heat map of the day; extract the urgency score of each measure in the measure inspection priority list; normalize the priority score and urgency score; fuse the normalized environmental information, priority score, urgency score and growth stage value of the day to construct the input feature vector of the day; select the strategy number of the comprehensive inspection score of the day as the output label to obtain the strategy training sample; use the gradient boosting tree model to train the strategy training dataset; update the weak learner along the negative gradient direction of the loss function in each iteration; select the cross-entropy loss function according to the classification nature of the task to obtain the strategy selection model.

[0055] Because the focus of field management differs significantly at different growth stages of crops, relying solely on anomaly records to calculate dynamic scores may lead to a discrepancy between the distribution of inspection priorities and actual production needs. By assigning differentiated weights to each grid at different stages, the model can be guided to better align with the laws of agricultural production, concentrating inspection resources in more critical areas, thereby improving inspection efficiency and the scientific nature of decision-making.

[0056] By calculating the dynamic score of the grid and combining it with the growth stage of the crop, different priority scores are assigned to each grid, so that the inspection process can automatically adjust the inspection focus according to the real-time needs of crop growth, thereby improving the efficiency of the allocation of inspection resources.

[0057] By using urgency scoring to prioritize field management measures, managers can intelligently schedule inspection tasks based on the urgency of different measures, avoiding resource waste and ensuring that critical issues are addressed in a timely manner.

[0058] By simulating and scoring various predefined inspection strategies, and combining parameters such as the number of newly discovered anomalies, flight power, and flight time, the comprehensive effect of different inspection strategies can be dynamically evaluated, providing a quantitative assessment for farmland inspection tasks and improving inspection efficiency and accuracy.

[0059] Based on the comprehensive inspection score and other relevant input data, a strategy training dataset is constructed, and the inspection strategy selection is optimized through a machine learning classification model. This enables continuous optimization of inspection decisions based on historical data, thereby improving the level of intelligence in farmland management.

[0060] Furthermore, step S400 includes:

[0061] Step S401: Receive real-time inspection tasks and obtain the farmland number, crop number and production cycle stage of the day corresponding to the task;

[0062] Step S402: Read the real-time environmental information of the day, call the abnormal location library, calculate the real-time dynamic score of the grid, and generate a real-time spatial inspection priority heat map;

[0063] Step S403: Call the abnormal measures library, input the field management measures records within the time window, calculate the real-time urgency score, and generate a real-time measures inspection priority list.

[0064] Furthermore, step S500 includes:

[0065] Step S501: Input the real-time environmental information, growth stage, grid priority score, and urgency score into the strategy selection model to obtain the recommended preliminary inspection strategy type;

[0066] Step S502: Preset priority scoring threshold, take the real-time spatial inspection priority heat map, preliminary inspection strategy type, environmental factor conditions and UAV flight constraints as input, extract the grids in the real-time spatial inspection priority heat map whose priority scores exceed the priority scoring threshold as candidate waypoints, and generate multiple inspection routes based on the candidate waypoints through the path planning algorithm;

[0067] Step S503: Collect the coverage area information, priority list of measures, and energy and time consumption data of each inspection route, and calculate the comprehensive priority score according to the following formula:

[0068] ;

[0069] Where R represents the overall priority score, U i Let K represent the normalized priority score for covering the i-th grid, where r represents the total number of grids covered, and K is the total number of grids covered. x Let represent the normalized urgency score of the x-th inspection measure, y represent the total number of inspection measures, w1 and w2 represent the weights of the coverage grid and the inspection measure, respectively, and w3 represent the weights of the inspection measures. n Y represents the weight of the nth flight phase in the overall priority score. n This represents the normalized duration of the nth flight phase, where m represents the total number of flight phases.

[0070] Step S504: Summarize the comprehensive priority scores of each inspection route, select the inspection route with the highest comprehensive priority score as the optimal inspection plan, and issue the optimal inspection plan to the UAV performing the task;

[0071] By integrating and analyzing real-time environmental information, crop growth stage, grid priority score, and urgency of measures through the strategy selection model, a preliminary inspection strategy that matches the current farmland condition can be automatically generated, reducing reliance on manual experience.

[0072] By setting priority scoring thresholds, high-value grids are extracted from the real-time spatial inspection priority heatmap as candidate waypoints, making inspection route planning more targeted and ensuring that limited resources are concentrated in the most critical areas.

[0073] After generating multiple inspection routes, factors such as coverage grid priority, urgency of measures, energy consumption and time consumption are comprehensively considered, and a unified comprehensive priority scoring formula is used to quantify the evaluation, so as to achieve a comprehensive comparison and scientific selection of inspection routes.

[0074] By balancing coverage area, measure priority, and flight cost, the energy consumption and time expenditure of drones can be effectively reduced, and the inspection efficiency and cost-effectiveness can be improved.

[0075] Ultimately, the system automatically selects the inspection route with the highest overall priority score and directly sends it to the execution drone, achieving full automation from data input to optimal solution execution, which greatly improves the real-time performance and execution efficiency of the inspection.

[0076] To better implement the above methods, an intelligent inspection system for low-altitude UAVs based on artificial intelligence is also proposed. The system includes an abnormal image information module, a feature library construction module, a training module, a real-time acquisition module, and a scheme determination module.

[0077] Abnormal Image Information Module: Acquires image information from inspection records through the drone inspection platform, calculates similarity with standard images, and determines abnormal image information;

[0078] Feature library construction module: Collect basic information identifiers of historical inspection records, summarize the historical inspection record set for each basic information identifier, and construct an abnormal location library and an abnormal measure library based on the historical inspection record sets containing abnormal image information.

[0079] Training module: Set the training period, generate a spatial inspection priority heatmap and a measure inspection priority list based on the abnormal location library and abnormal measure library, calculate the comprehensive inspection score, and build a strategy selection model;

[0080] Real-time data acquisition module: Collects real-time data and generates a real-time spatial inspection priority heatmap and a real-time measure inspection priority list;

[0081] The scheme determination module generates preliminary inspection strategy types based on the strategy selection model. Based on the real-time spatial inspection priority heatmap and the real-time measure inspection priority list, it generates multiple inspection routes through path planning algorithms, calculates the comprehensive priority score of each inspection route, determines the optimal inspection scheme, and distributes it to the UAVs that will execute the mission.

[0082] Furthermore, the abnormal image information module includes an inspection record unit and an abnormal image information determination unit:

[0083] Inspection Record Unit: In the UAV inspection platform, the crops planted in the farmland to be inspected are obtained, and the farmland and crops are numbered respectively. The field management measures implemented for the crops are recorded. Environmental sensors are deployed in each farmland to collect the environmental information of the farmland. During the UAV inspection, the flight information and image information of the UAV are collected. The field management measures, environmental information, flight information and image information are summarized to generate an inspection record, and the inspection record is uploaded to the UAV inspection platform.

[0084] Identifying Abnormal Image Information Units: This involves acquiring inspection records, extracting image information from each record, marking crops in the images using image recognition technology, extracting the outlines of the marked crops using image segmentation technology to obtain complete images of the crops, collecting the time point of each inspection record, obtaining the corresponding standard image of the crop, dividing the complete image and the standard image into several standard squares, calculating the similarity between the standard squares of the complete image and the standard image using a structural similarity algorithm to obtain the average similarity of the image information, setting a preset average similarity threshold, marking image information with similarity values ​​less than the average similarity threshold as abnormal image information, and writing the abnormal image information back to the corresponding inspection record.

[0085] Compared with the prior art, the beneficial effects of the present invention are: through artificial intelligence algorithms, it can automatically analyze and process images and data captured by drones, identify the health status of crops in real time, judge abnormal situations, reduce manual intervention, and improve inspection efficiency;

[0086] By constructing an abnormal location database and an abnormal measure database, and combining them with historical inspection data, we can accurately identify the location of abnormalities and prioritize these high-risk areas in subsequent inspections, thereby improving the accuracy and timeliness of inspections.

[0087] By collecting environmental data, crop growth information, and historical inspection records in real time, the system can dynamically adjust the inspection strategy according to the current environment and crop growth stage, and generate a real-time inspection priority heat map and urgency score to ensure that the inspection plan is more in line with the current needs.

[0088] By comprehensively considering factors such as the energy efficiency, timeliness, and inspection priority of flight paths, multiple inspection routes are intelligently generated, and the comprehensive priority score of each route is calculated to ensure the generation of the optimal inspection plan and reduce energy and time waste.

[0089] Based on real-time environmental information and production cycle stages, it can automatically assign tasks to drones, generate precise inspection strategies, and assign inspection routes through path planning algorithms, greatly improving the automation level of drone inspections. Attached Figure Description

[0090] Figure 1 This is a flowchart illustrating an intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence, according to the present invention.

[0091] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based low-altitude unmanned aerial vehicle (UAV) intelligent inspection system according to the present invention. Detailed Implementation

[0092] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] Please see Figure 1 and Figure 2 This invention provides a technical solution: an intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence, the method comprising:

[0094] Step S100: Obtain image information from the inspection record through the drone inspection platform, and perform similarity calculation with standard images to determine abnormal image information;

[0095] Step S100 includes:

[0096] Step S101: In the UAV inspection platform, obtain the crops planted in the farmland to be inspected, and number the farmland and crops respectively. Record the field management measures implemented for the crops. Deploy environmental sensors in each farmland to collect the environmental information of the farmland. During the UAV inspection process, collect the flight information and image information of the UAV. Summarize the field management measures, environmental information, flight information and image information to generate an inspection record, and upload the inspection record to the UAV inspection platform.

[0097] Step S102: Obtain inspection records, extract image information from each inspection record, mark the crops in the image information using image recognition technology, and extract the outlines of the marked crops using image segmentation technology to obtain a complete image of the crops;

[0098] Step S103: Collect the time point of each inspection record, obtain the standard image of the crop corresponding to the time point, divide the complete image and the standard image into several standard squares, and calculate the similarity of the standard squares of the complete image and the standard image through a structural similarity algorithm to obtain the average similarity of the image information.

[0099] Step S104: Set an average similarity threshold, mark image information with a similarity value less than the average similarity threshold as abnormal image information, and write the abnormal image information back to the corresponding inspection record;

[0100] For example, a drone is inspecting a wheat field. The platform has already obtained information about the field to be inspected and has determined that the field is planted with wheat.

[0101] The farmland is numbered "Farmland 001", and the crop is numbered "Wheat 001";

[0102] The management practices for this farmland may include the use of fertilization, irrigation, and pest and disease control measures.

[0103] Multiple environmental sensors are deployed in the farmland to collect information such as soil temperature and humidity, air temperature and humidity, and light intensity in real time.

[0104] The drone captured high-resolution images of farmland during its flight;

[0105] The drone will collect flight data, image information, environmental data, etc. to generate inspection records and upload them to the drone inspection platform;

[0106] Using image recognition technology, the system processes the uploaded images and marks the areas of wheat crops. For example, it uses a deep learning model to identify the leaves and stems of wheat.

[0107] Image segmentation techniques are used to extract the outlines of the marked crops, thereby obtaining a complete image of the wheat crop and removing background clutter, clouds, and other interfering information.

[0108] Step S200: Collect basic information identifiers of historical inspection records, summarize the historical inspection record set for each basic information identifier, and construct an abnormal location database and an abnormal measure database based on the historical inspection record set;

[0109] Step S200 includes:

[0110] Step S201: In the historical inspection record set, collect the farmland number and crop number of the historical inspection record, and construct a basic information identifier by combining the production cycle information of the crop in the farmland. The production cycle information is the complete cycle of the crop from sowing to harvest. Summarize the historical inspection records with the same basic information identifier and in the same production cycle to obtain the historical inspection record set for each basic information identifier.

[0111] Step S202: In the historical inspection record set of a certain basic information identifier, extract the historical inspection records with abnormal image information and set them as historical abnormal records. Collect the inspection time of each historical abnormal record and arrange the historical abnormal records in chronological order of inspection time. Obtain the abnormal location information of each historical abnormal record and construct the abnormal location library of the basic information identifier.

[0112] Step S202 includes:

[0113] Step S202-1: Obtain the farmland number from the basic information identifier, retrieve the satellite remote sensing image of the farmland number, perform three-dimensional modeling on the farmland using three-dimensional modeling technology to obtain a three-dimensional farmland model, divide the three-dimensional farmland model into several regions, and establish an independent spatial coordinate system in each region;

[0114] Step S202-2: Collect the geographic location information of the abnormal image information in the historical abnormal records, and combine it with the flight trajectory data of the UAV to determine the area corresponding to the abnormal image information, and calculate the precise spatial coordinates of the abnormal image in the area according to the spatial coordinate system of the area;

[0115] Step S202-3: Count the number of times historical abnormal records appear in a certain area, calculate the frequency of occurrence in the area, collect the anomaly type and average similarity of a certain historical abnormal record in the area, assign a value to each anomaly type, and normalize the frequency of occurrence and average similarity. Calculate the anomaly degree index of the area according to the following formula:

[0116] ;

[0117] Where A represents the regional anomaly index, B represents the normalized frequency of occurrence, and C... b D represents the numerical value of the anomaly type in the b-th historical anomaly record. b Let d represent the normalized average similarity value of the b-th historical anomaly record, d represent the total number of historical anomaly records, and a1 and a2 represent the weights of the frequency of occurrence and the average similarity, respectively.

[0118] Step S202-4: Divide the spatial coordinate system of a certain region into several grids, obtain the grid corresponding to each historical anomaly record, count the number of historical anomaly records in a certain grid as v1, calculate the distribution density of a certain grid Z=v1 / v2, where v2 represents the number of historical anomaly records in the region, normalize the distribution density and calculate it with the regional anomaly degree index to obtain the refined anomaly degree index of each grid, count the historical anomaly records of each grid in each growth stage, divide it by all historical anomaly records in the growth stage to calculate the occurrence frequency, and set it as the weight of the grid in the growth stage. Store the calculated regional anomaly degree index, refined anomaly degree index and weights in different growth stages into the anomaly location database.

[0119] For example, acquire satellite remote sensing images of farmland 001 and use 3D modeling technology to construct a 3D model of the farmland. Assume that the model includes 10 regions, such as A1, A2, A3...A10, each region representing a partition of the farmland.

[0120] Suppose that the anomalous image in the a-th historical anomaly record is a yellowed image, located in area A1 of farmland 001; through UAV flight trajectory data, we can determine the geographical location information of this anomalous image as "longitude: 116.5°, latitude: 39.9°";

[0121] By combining the spatial coordinate system of the 3D model of the farmland, we can convert the spatial coordinates of this abnormal image into precise coordinates within the A1 area of ​​the farmland. For example, the coordinates of this location within the A1 area are "X: 30, Y: 45".

[0122] Assume that in the first growth stage, G1 has 20 historical abnormal records, G2 has 15 historical abnormal records, and the total number of historical abnormal records is 100; in the second growth stage, G1 has 10 historical abnormal records, G2 has 25 historical abnormal records, and the total number of historical abnormal records is 150.

[0123] The weights of G1 and G2 were calculated to be 0.2 and 0.15 in the first growth stage, and 0.06 and 0.16 in the second growth stage.

[0124] Step S203: In the historical inspection record set of a certain basic information identifier, obtain the field management measures in each historical abnormal record, and construct the abnormal measures library of the basic information identifier;

[0125] Step S203 includes:

[0126] Step S203-1: Collect information on field management measures corresponding to historical anomaly records, including measure type, implementation time, implementation duration, and implementation amount; and bind the field management measures with historical anomaly records to form a correspondence between anomalies and measures;

[0127] Step S203-2: According to the implementation time sequence, using the inspection time of historical anomaly records as a reference point, extract the field management measures within the preset time window before the anomaly occurred to obtain the sequence of measures related to the anomaly;

[0128] Step S203-3: According to the measure sequence, count the number of times each measure appears in a preset time window, calculate the frequency of each measure, and set the frequency of occurrence as the abnormal correlation degree of the measure. Collect the time interval between the implementation time and the inspection time in each historical abnormal record, obtain the implementation amount of the implemented measure, and normalize the implementation amount. Use the normalized implementation amount and abnormal correlation degree as input and the time interval as output to construct a training dataset. Train the training dataset through a linear regression model to obtain the trained weights, update the linear regression model, and set it as the measure prediction model. Store it in the abnormal measure library.

[0129] For example, in the first historical anomaly record, the inspection time was March 10, the measures were fertilization and watering, the implementation time was March 7, and the implementation amount was 8 and 5;

[0130] The inspection time in the second historical anomaly record is March 18, the measures are fertilization and spraying, the implementation time is March 15, and the implementation amount is 10 and 3.

[0131] The third historical anomaly record shows that the inspection time was March 25, the measure was watering, the implementation time was March 23, and the implementation amount was 6.

[0132] The fourth historical anomaly record shows that the inspection time was April 2, the measure was fertilization, the implementation time was March 28, and the implementation amount was 9.

[0133] The calculated frequency of fertilization was 0.75, the frequency of watering was 0.5, and the frequency of pesticide spraying was 0.25.

[0134] Historical records of abnormal fertilization were compiled, and the time intervals were calculated to be 3 days, 3 days, and 5 days, respectively. The normalized application amounts were 0.1, 0.9, and 0.5, respectively.

[0135] The training dataset obtained is as follows: Sample 1: Input features are (implementation value = 0.1, anomaly correlation = 0.75), output time interval = 3;

[0136] Sample 2: Input features are (implementation quantity = 0.9, anomaly correlation degree = 0.75), output time interval = 3;

[0137] Sample 3: Input features are (implementation quantity = 0.5, anomaly correlation degree = 0.75), output time interval = 5;

[0138] Input the above training data into the linear regression model to obtain the weights of the linear regression model.

[0139] Step S300: Set the training period, generate a spatial inspection priority heatmap and a measure inspection priority list based on the abnormal location library and abnormal measure library, calculate the comprehensive inspection score, and build a strategy selection model;

[0140] Step S300 includes:

[0141] Step S301: Select a consecutive number of days as the training period. Collect environmental information for each day within the training period. Using the anomaly location database, map the refined anomaly level index of a certain basic information identifier to the spatial coordinate system of each area of ​​farmland to generate a static anomaly probability heatmap. Obtain the time interval between each anomaly record and the current day. Calculate the dynamic score of the grid according to the following formula:

[0142] ;

[0143] Where E represents the dynamic score, F g Let t represent the refined anomaly level index corresponding to the g-th anomaly record, u represent the decay rate parameter, and t represent the decay rate parameter. g Let f represent the time interval of the g-th abnormal record, and let f represent the total number of abnormal records. Obtain the current crop production cycle information, extract the crop growth stage of the day, and multiply the weight of each grid in the growth stage with the dynamic score to calculate the priority score of each grid on the day. Map the priority score to the static anomaly probability heat map to obtain the spatial inspection priority heat map.

[0144] Step S302: Obtain the field management measures within the time window prior to the current day, collect the implementation amount of each measure, perform normalization calculations, input the data into the measure prediction model, calculate the prediction time interval, and calculate the urgency score according to the following formula:

[0145] ;

[0146] Where K represents the urgency score, j1 represents the time interval between the current day and the implemented measures, and j2 represents the predicted time interval. The implemented measures are sorted from high to low according to the urgency score to generate a list of measures inspection priorities.

[0147] Step S303: Simulate multiple predefined inspection strategies daily, assign a number to each strategy, acquire UAV flight information, extract the number of new anomalies and flight parameters during inspection, perform normalization calculations, and calculate the comprehensive inspection score according to the following formula:

[0148] ;

[0149] Where P represents the overall inspection score, S represents the number of new anomalies in the inspection after normalization, Q represents the normalized average flight power, N represents the normalized flight time, and Y represents the normalized flight time. n h3 represents the normalized duration of the nth flight phase. n Let h1 represent the weight of the nth flight phase, m represent the total number of flight phases, and h2 represent the weights of the number of newly discovered anomalies and the average flight power, respectively.

[0150] Step S304: Obtain the environmental information of the day and perform normalization calculation; collect the growth stage corresponding to the crop of the day and assign a value to each growth stage; extract the priority score of each grid in the spatial inspection priority heat map of the day; extract the urgency score of each measure in the measure inspection priority list; normalize the priority score and urgency score; fuse the normalized environmental information, priority score, urgency score and growth stage value of the day to construct the input feature vector of the day; select the strategy number of the comprehensive inspection score of the day as the output label to obtain the strategy training sample; use the gradient boosting tree model to train the strategy training dataset; update the weak learner along the negative gradient direction of the loss function in each iteration; select the cross-entropy loss function according to the classification nature of the task to obtain the strategy selection model.

[0151] For example, Day 1: Temperature 25℃, Humidity 65%;

[0152] Day 2: Temperature 26℃, humidity 68%;

[0153] Day 3: Temperature 27℃, humidity 70%;

[0154] Day 4: Temperature 28℃, humidity 72%;

[0155] Day 5: Temperature 29℃, humidity 73%;

[0156] There are 3 anomalous records in grid G1, and the fine-grained anomalousness indexes for each anomalous record are 0.8, 0.6, and 0.4, respectively.

[0157] There are 2 anomalous records in grid G2, and the fine-grained anomalousness indexes for each anomalous record are 0.7 and 0.5, respectively.

[0158] The attenuation rate parameter is set to 0.1, and the time intervals in G1 are 1, 2, and 4; the time intervals in G2 are 1 and 3.

[0159] The calculated dynamic score for G1 is 1.483, and the dynamic score for G2 is 1.005.

[0160] Currently in the first growth stage, the priority score of G1 is calculated to be 0.2966, and the priority score of G2 is 0.15075.

[0161] Assume that the predefined inspection strategy includes the regular inspection strategy numbered 1;

[0162] The refined inspection strategy numbered 2;

[0163] Optimized inspection strategy number 3;

[0164] The input feature vector for the first training dataset is [0.625, 0.650, 1.0, 0.2966, 0.15075, 0.80, 0.60], and the output strategy number is 1.

[0165] The input feature vector for the second training dataset is [0.650, 0.680, 1.0, 0.3100, 0.16000, 0.83, 0.63], and the output strategy number is 3.

[0166] The input feature vector for the third training dataset is [0.675, 0.700, 1.0, 0.2966, 0.15075, 0.85, 0.65], and the output strategy number is 2.

[0167] The training dataset is input into the gradient boosting tree model. In each iteration, the gradient of the loss function for the current prediction is calculated.

[0168] Fit a new decision tree to approximate this gradient;

[0169] Overlay this tree onto the old prediction results;

[0170] After multiple iterations, the model's predicted probabilities across the three categories gradually converge, generating a policy selection model.

[0171] Step S400: Collect real-time data information and generate a real-time spatial inspection priority heat map and a real-time measure inspection priority list;

[0172] Step S400 includes:

[0173] Step S401: Receive real-time inspection tasks and obtain the farmland number, crop number and production cycle stage of the day corresponding to the task;

[0174] Step S402: Read the real-time environmental information of the day, call the abnormal location library, calculate the real-time dynamic score of the grid, and generate a real-time spatial inspection priority heat map;

[0175] Step S403: Call the abnormal measures library, input the field management measures records within the time window, calculate the real-time urgency score, and generate a real-time measures inspection priority list.

[0176] Step S500: Based on the strategy selection model, generate a preliminary inspection strategy type, generate multiple inspection routes through a path planning algorithm, calculate the comprehensive priority score of each inspection route, determine the optimal inspection plan, and send it to the UAVs performing the task.

[0177] Step S500 includes:

[0178] Step S501: Input the real-time environmental information, growth stage, grid priority score, and urgency score into the strategy selection model to obtain the recommended preliminary inspection strategy type;

[0179] Step S502: Preset priority scoring threshold, take the real-time spatial inspection priority heat map, preliminary inspection strategy type, environmental factor conditions and UAV flight constraints as input, extract the grids in the real-time spatial inspection priority heat map whose priority scores exceed the priority scoring threshold as candidate waypoints, and generate multiple inspection routes based on the candidate waypoints through the path planning algorithm;

[0180] Step S503: Collect the coverage area information, priority list of measures, and energy and time consumption data of each inspection route, and calculate the comprehensive priority score according to the following formula:

[0181] ;

[0182] Where R represents the overall priority score, U i Let K represent the normalized priority score for covering the i-th grid, where r represents the total number of grids covered, and K is the total number of grids covered. x Let represent the normalized urgency score of the x-th inspection measure, y represent the total number of inspection measures, w1 and w2 represent the weights of the coverage grid and the inspection measure, respectively, and w3 represent the weights of the inspection measures. n Y represents the weight of the nth flight phase in the overall priority score. n This represents the normalized duration of the nth flight phase, where m represents the total number of flight phases.

[0183] Step S504: Summarize the comprehensive priority scores of each inspection route, select the inspection route with the highest comprehensive priority score as the optimal inspection plan, and issue the optimal inspection plan to the UAV performing the task;

[0184] For example, real-time environmental information: temperature: 28℃, humidity: 72%;

[0185] Growth stage: Jointing stage;

[0186] Mesh priority scores: G1: 0.2966, G2: 0.15075;

[0187] Urgency rating: Spraying: 1.00, Fertilizing: 0.80, Irrigating: 0.25;

[0188] The initial inspection strategy number is 1;

[0189] Assuming the priority score threshold is 0.25, only grids with a priority score greater than 0.25 will be selected as candidate waypoints.

[0190] G1: The priority score is 0.2966, which exceeds the threshold, so it is selected as a candidate waypoint;

[0191] G2: Priority score is 0.15075, which is below the threshold and will not be selected as a candidate waypoint;

[0192] Based on the selected candidate waypoint G1, multiple routes are generated using a path planning algorithm. Assume two inspection routes are generated:

[0193] Route 1: From point A to G1, then to point B;

[0194] Route 2: From point C to G1, then to point D;

[0195] Coverage grid information: Route 1: Coverage grid G1, priority score 0.2860;

[0196] Route 2: Covers grid G1, priority score 0.2913;

[0197] Route energy consumption and time consumption data: Route 1: Energy consumption 50W, time consumption 30 minutes

[0198] Route 2: Energy consumption 55W, travel time 35 minutes;

[0199] The overall priority score for route 1 was calculated to be 0.2945, and the overall priority score for route 2 was 0.3013.

[0200] Route 2 is selected as the optimal inspection plan.

[0201] To better implement the above methods, an intelligent inspection system for low-altitude UAVs based on artificial intelligence is also proposed. The system includes an abnormal image information module, a feature library construction module, a training module, a real-time acquisition module, and a scheme determination module.

[0202] Abnormal Image Information Module: Acquires image information from inspection records through the drone inspection platform, calculates similarity with standard images, and determines abnormal image information;

[0203] The abnormal image information module includes an inspection record unit and an abnormal image information determination unit.

[0204] Inspection Record Unit: In the UAV inspection platform, the crops planted in the farmland to be inspected are obtained, and the farmland and crops are numbered respectively. The field management measures implemented for the crops are recorded. Environmental sensors are deployed in each farmland to collect the environmental information of the farmland. During the UAV inspection, the flight information and image information of the UAV are collected. The field management measures, environmental information, flight information and image information are summarized to generate an inspection record, and the inspection record is uploaded to the UAV inspection platform.

[0205] Identifying Abnormal Image Information Units: This involves acquiring inspection records, extracting image information from each record, marking crops in the images using image recognition technology, extracting the outlines of the marked crops using image segmentation technology to obtain complete images of the crops, collecting the time point of each inspection record, obtaining the corresponding standard image of the crop, dividing the complete image and the standard image into several standard squares, calculating the similarity between the standard squares of the complete image and the standard image using a structural similarity algorithm to obtain the average similarity of the image information, setting a preset average similarity threshold, marking image information with similarity values ​​less than the average similarity threshold as abnormal image information, and writing the abnormal image information back to the corresponding inspection record.

[0206] Feature library construction module: Collect basic information identifiers of historical inspection records, summarize the historical inspection record set for each basic information identifier, and construct an abnormal location library and an abnormal measure library based on the historical inspection record sets containing abnormal image information.

[0207] Training module: Set the training period, generate a spatial inspection priority heatmap and a measure inspection priority list based on the abnormal location library and abnormal measure library, calculate the comprehensive inspection score, and build a strategy selection model;

[0208] Real-time data acquisition module: Collects real-time data and generates a real-time spatial inspection priority heatmap and a real-time measure inspection priority list;

[0209] The scheme determination module generates preliminary inspection strategy types based on the strategy selection model. Based on the real-time spatial inspection priority heatmap and the real-time measure inspection priority list, it generates multiple inspection routes through path planning algorithms, calculates the comprehensive priority score of each inspection route, determines the optimal inspection scheme, and distributes it to the UAVs that will execute the mission.

[0210] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for intelligent inspection of low-altitude unmanned aerial vehicles based on artificial intelligence, characterized in that, The methods include: Step S100: Obtain image information from the inspection record through the drone inspection platform, and perform similarity calculation with standard images to determine abnormal image information; Step S200: Collect basic information identifiers of historical inspection records, summarize the historical inspection record set for each basic information identifier, and construct an abnormal location database and an abnormal measure database based on the historical inspection records containing abnormal image information in the historical inspection record set. Step S300: Set the training period, generate a spatial inspection priority heatmap and a measure inspection priority list based on the abnormal location library and abnormal measure library, calculate the comprehensive inspection score, and build a strategy selection model; Step S400: Collect real-time data information and generate a real-time spatial inspection priority heat map and a real-time measure inspection priority list; Step S500: Based on the strategy selection model, generate a preliminary inspection strategy type, generate multiple inspection routes based on the real-time spatial inspection priority heat map and the real-time measure inspection priority list, calculate the comprehensive priority score of each inspection route, determine the optimal inspection plan, and issue it to the UAVs that perform the task.

2. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, Step S100 includes the following steps: Step S101: In the UAV inspection platform, obtain the crops planted in the farmland to be inspected, and number the farmland and crops respectively. Record the field management measures implemented for the crops. Deploy environmental sensors in each farmland to collect the environmental information of the farmland. During the UAV inspection process, collect the flight information and image information of the UAV. Summarize the field management measures, environmental information, flight information and image information to generate an inspection record, and upload the inspection record to the UAV inspection platform. Step S102: Obtain inspection records, extract image information from each inspection record, mark the crops in the image information using image recognition technology, and extract the outlines of the marked crops using image segmentation technology to obtain a complete image of the crops; Step S103: Collect the time point of each inspection record, obtain the standard image of the crop corresponding to the time point, divide the complete image and the standard image into several standard squares, and calculate the similarity of the standard squares of the complete image and the standard image through a structural similarity algorithm to obtain the average similarity of the image information. Step S104: Set an average similarity threshold, mark image information with a similarity value less than the average similarity threshold as abnormal image information, and write the abnormal image information back to the corresponding inspection record.

3. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 2, characterized in that, Step S200 includes the following steps: Step S201: In the historical inspection record set, collect the farmland number and crop number of the historical inspection record, and construct a basic information identifier by combining the production cycle information of the crop in the farmland. The production cycle information is the complete cycle of the crop from sowing to harvest. Summarize the historical inspection records with the same basic information identifier and in the same production cycle to obtain the historical inspection record set for each basic information identifier. Step S202: In the historical inspection record set of a certain basic information identifier, extract the historical inspection records with abnormal image information and set them as historical abnormal records. Collect the inspection time of each historical abnormal record and arrange the historical abnormal records in chronological order of inspection time. Obtain the abnormal location information of each historical abnormal record and construct the abnormal location library of the basic information identifier. Step S203: In the historical inspection record set of a certain basic information identifier, obtain the field management measures in each historical abnormal record, and construct the abnormal measures library of the basic information identifier.

4. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 3, characterized in that, The step S202 of constructing the abnormal location database includes the following steps: Step S202-1: Obtain the farmland number from the basic information identifier, retrieve the satellite remote sensing image of the farmland number, perform three-dimensional modeling on the farmland using three-dimensional modeling technology to obtain a three-dimensional farmland model, divide the three-dimensional farmland model into several regions, and establish an independent spatial coordinate system in each region; Step S202-2: Collect the geographic location information of the abnormal image information in the historical abnormal records, and combine it with the flight trajectory data of the UAV to determine the area corresponding to the abnormal image information, and calculate the precise spatial coordinates of the abnormal image in the area according to the spatial coordinate system of the area; Step S202-3: Count the number of times historical abnormal records appear in a certain area, calculate the frequency of occurrence in the area, collect the anomaly type and average similarity of a certain historical abnormal record in the area, assign a value to each anomaly type, and normalize the frequency of occurrence and average similarity. Calculate the anomaly degree index of the area according to the following formula: ； Where A represents the regional anomaly index, B represents the normalized frequency of occurrence, and C... b D represents the numerical value of the anomaly type in the b-th historical anomaly record. b Let d represent the normalized average similarity value of the b-th historical anomaly record, d represent the total number of historical anomaly records, and a1 and a2 represent the weights of the frequency of occurrence and the average similarity, respectively. Step S202-4: Divide the spatial coordinate system of a certain region into several grids, obtain the grid corresponding to each historical anomaly record, count the number of historical anomaly records in a certain grid as v1, calculate the distribution density of a certain grid Z=v1 / v2, where v2 represents the number of historical anomaly records in the region, normalize the distribution density, and calculate it with the regional anomaly degree index to obtain the refined anomaly degree index of each grid, count the historical anomaly records of each grid in each growth stage, divide it by all historical anomaly records in the growth stage to calculate the occurrence frequency, and set it as the weight of the grid in the growth stage. Store the calculated regional anomaly degree index, refined anomaly degree index and weights in different growth stages into the anomaly location database.

5. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 3, characterized in that, The step S203 of constructing the anomaly response library includes the following steps: Step S203-1: Collect information on field management measures corresponding to historical anomaly records, including measure type, implementation time, implementation duration, and implementation amount; and bind the field management measures with historical anomaly records to form a correspondence between anomalies and measures; Step S203-2: According to the implementation time sequence, using the inspection time of historical anomaly records as a reference point, extract the field management measures within the preset time window before the anomaly occurred to obtain the sequence of measures related to the anomaly; Step S203-3: Based on the measure sequence, count the number of times each measure appears in a preset time window, calculate the frequency of each measure, and set the frequency of occurrence as the abnormal correlation degree of the measure. Collect the time interval between the implementation time and the inspection time in each historical abnormal record, obtain the implementation amount of the implemented measure, and normalize the implementation amount. Use the normalized implementation amount and abnormal correlation degree as input and the time interval as output to construct a training dataset. Train the training dataset using a linear regression model to obtain the trained weights, update the linear regression model, and set it as a measure prediction model. Store it in the abnormal measure library.

6. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 5, characterized in that, Step S300 includes the following steps: Step S301: Select a consecutive number of days as the training period. Collect environmental information for each day within the training period. Using the anomaly location database, map the refined anomaly level index of a certain basic information identifier to the spatial coordinate system of each area of ​​farmland to generate a static anomaly probability heatmap. Obtain the time interval between each anomaly record and the current day. Calculate the dynamic score of the grid according to the following formula: ； Where E represents the dynamic score, F g Let t represent the refined anomaly level index corresponding to the g-th anomaly record, u represent the decay rate parameter, and t represent the decay rate parameter. g Let f represent the time interval of the g-th abnormal record, and let f represent the total number of abnormal records. Obtain the current crop production cycle information, extract the crop growth stage of the day, and multiply the weight of each grid in the growth stage with the dynamic score to calculate the priority score of each grid on the day. Map the priority score to the static anomaly probability heat map to obtain the spatial inspection priority heat map. Step S302: Obtain the field management measures within the time window prior to the current day, collect the implementation amount of each measure, perform normalization calculations, input the data into the measure prediction model, calculate the prediction time interval, and calculate the urgency score according to the following formula: ； Where K represents the urgency score, j1 represents the time interval between the current day and the implemented measures, and j2 represents the predicted time interval. The implemented measures are sorted from high to low according to the urgency score to generate a list of measures inspection priorities. Step S303: Simulate multiple predefined inspection strategies daily, assign a number to each strategy, acquire the drone's flight information, extract the number of new anomalies and flight parameters during inspection, perform normalization calculations, and calculate the comprehensive inspection score according to the following formula: ； Where P represents the overall inspection score, S represents the number of new anomalies in the inspection after normalization, Q represents the normalized average flight power, N represents the normalized flight time, and Y represents the normalized flight time. n h3 represents the normalized duration of the nth flight phase. n Let h1 represent the weight of the nth flight phase, m represent the total number of flight phases, and h2 represent the weights of the number of newly discovered anomalies and the average flight power, respectively. Step S304: Obtain the environmental information for the day and perform normalization calculations. Collect the growth stage corresponding to the crops for the day and assign a value to each growth stage. Extract the priority score of each grid in the spatial inspection priority heatmap for the day. Extract the urgency score of each measure in the measure inspection priority list. Normalize the priority score and urgency score. Perform feature fusion on the normalized environmental information, priority score, urgency score and growth stage values ​​for the day to construct the input feature vector for the day. Select the strategy number of the comprehensive inspection score for the day as the output label to obtain the strategy training sample. Use the gradient boosting tree model to train the strategy training dataset. Update the weak learner along the negative gradient direction of the loss function in each iteration. Select the cross-entropy loss function according to the classification nature of the task to obtain the strategy selection model.

7. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 6, characterized in that, Step S400 includes the following steps: Step S401: Receive real-time inspection tasks and obtain the farmland number, crop number and production cycle stage of the day corresponding to the task; Step S402: Read the real-time environmental information of the day, call the abnormal location library, calculate the real-time dynamic score of the grid, and generate a real-time spatial inspection priority heat map; Step S403: Call the abnormal measures library, input the field management measures records within the time window, calculate the real-time urgency score, and generate a real-time measures inspection priority list.

8. The intelligent inspection method for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 7, characterized in that, Step S500 includes the following steps: Step S501: Input the real-time environmental information, growth stage, grid priority score, and urgency score into the strategy selection model to obtain the recommended preliminary inspection strategy type; Step S502: Preset priority scoring threshold, take the real-time spatial inspection priority heat map, preliminary inspection strategy type, environmental factor conditions and UAV flight constraints as input, extract the grids in the real-time spatial inspection priority heat map whose priority scores exceed the priority scoring threshold as candidate waypoints, and generate multiple inspection routes based on the candidate waypoints through the path planning algorithm; Step S503: Collect the coverage area information, priority list of measures, and energy and time consumption data of each inspection route, and calculate the comprehensive priority score according to the following formula: ； Where R represents the overall priority score, U i Let K represent the normalized priority score for covering the i-th grid, where r represents the total number of grids covered, and K is the total number of grids covered. x Let represent the normalized urgency score of the x-th inspection measure, y represent the total number of inspection measures, w1 and w2 represent the weights of the coverage grid and the inspection measure, respectively, and w3 represent the weights of the inspection measures. n Y represents the weight of the nth flight phase in the overall priority score. n This represents the normalized duration of the nth flight phase, where m represents the total number of flight phases. Step S504: Summarize the comprehensive priority scores of each inspection route, select the inspection route with the highest comprehensive priority score as the optimal inspection plan, and issue the optimal inspection plan to the UAV performing the task.

9. An artificial intelligence-based intelligent inspection system for low-altitude unmanned aerial vehicles (UAVs), used to implement the artificial intelligence-based intelligent inspection method for low-altitude UAVs as described in any one of claims 1-8, characterized in that, The system includes an abnormal image information module, a feature library construction module, a training module, a real-time acquisition module, and a scheme determination module; The abnormal image information module: acquires image information from inspection records through the drone inspection platform, calculates similarity with standard images, and determines abnormal image information; The feature library construction module: collects basic information identifiers of historical inspection records, summarizes the historical inspection record set for each basic information identifier, and constructs an abnormal location library and an abnormal measure library based on the historical inspection records containing abnormal image information in the historical inspection record set. The training module: sets a training period, generates a spatial inspection priority heatmap and a measure inspection priority list based on the abnormal location library and the abnormal measure library, calculates the comprehensive inspection score, and constructs a strategy selection model; The real-time acquisition module collects real-time data information and generates a real-time spatial inspection priority heatmap and a real-time measure inspection priority list. The scheme determination module generates a preliminary inspection strategy type based on the strategy selection model, generates multiple inspection routes based on the real-time spatial inspection priority heatmap and the real-time measure inspection priority list, calculates the comprehensive priority score of each inspection route, determines the optimal inspection scheme, and distributes it to the UAVs performing the task.

10. The intelligent inspection system for low-altitude unmanned aerial vehicles based on artificial intelligence according to claim 9, characterized in that, The abnormal image information module includes an inspection recording unit and an abnormal image information determination unit: The inspection record unit: In the UAV inspection platform, it acquires the crops planted in the farmland to be inspected, and numbers the farmland and crops respectively, records the field management measures implemented for the crops, deploys environmental sensors in each farmland to collect the environmental information of the farmland, collects the flight information and image information of the UAV during the UAV inspection, summarizes the field management measures, environmental information, flight information and image information, generates an inspection record, and uploads the inspection record to the UAV inspection platform; The abnormal image information determination unit: acquires inspection records, extracts image information from each inspection record, marks crops in the image information using image recognition technology, extracts the contours of the marked crops using image segmentation technology to obtain complete images of the crops, collects the time point of each inspection record, obtains the standard image of the crops corresponding to the time point, divides the complete image and the standard image into several standard squares, calculates the similarity between the standard squares of the complete image and the standard image using a structural similarity algorithm to obtain the average similarity of the image information, presets an average similarity threshold, marks image information with a similarity less than the average similarity threshold as abnormal image information, and writes the abnormal image information back to the corresponding inspection record.

Images