Unmanned aerial vehicle inspection method and system based on AI visual control

Abstract

The invention discloses an unmanned aerial vehicle inspection method and system based on AI visual control, and belongs to the technical field of unmanned aerial vehicle automatic control, and the method comprises the steps: carrying out the feature component extraction and compression according to discontinuous frame image data collected by an airborne camera and preset inertial navigation data, and obtaining a compressed feature flow; performing multi-mode obstacle safety level parameter analysis, generating an obstacle avoidance decision priority map, constructing a safety corridor, and generating a real-time self-correction route; semantic analysis is carried out on the deviation between the real-time flight trajectory and the real-time self-correction route, and target area abnormal mark data is generated; and dynamically adjusting the bandwidth allocation of the dual-channel communication network, and transmitting the abnormal mark data of the target area to a ground control terminal. According to the method, a non-continuous frame image and inertial navigation fusion modeling mechanism is adopted, and a multi-mode obstacle safety grading and double-channel dynamic bandwidth allocation strategy is combined, so that real-time response of dynamic obstacle avoidance, accurate judgment of anomaly detection and efficient transmission of key data can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) automatic control, and more particularly to a UAV inspection method and system based on AI vision control. Background Art

[0002] UAV inspection technology is widely used in scenarios such as power detection and facility monitoring, and its core relies on flight path planning and dynamic environment perception technology. Existing technologies usually use GPS and inertial navigation systems to construct flight trajectories, combined with visible light or infrared sensors for obstacle detection.

[0003] Current mainstream solutions collect environmental information through continuous video streams and use SLAM algorithms to generate 3D maps. Some systems use lidar point cloud data to compare with pre-set flight paths to achieve obstacle avoidance, and transmit detection data to ground terminals through a single communication link.

[0004] However, continuous frame data processing leads to excessive computing power load on edge devices, and significant model update delays in complex scenarios; lidar point cloud analysis has a lag in response to dynamic obstacles and is difficult to cope with the threat of sudden moving targets; single-channel communication is vulnerable to environmental interference, and the reliability of high-priority data transmission is insufficient, and there is a risk of data loss in extreme conditions. Summary of the Invention

[0005] To solve the above problems, the present invention provides a UAV inspection method and system based on AI vision control, which adopts a fusion modeling mechanism of discontinuous frame images and inertial navigation, combined with a multi-modal obstacle safety classification and dual-channel dynamic bandwidth allocation strategy, and can achieve real-time response to dynamic obstacle avoidance, accurate discrimination of abnormal detection, and efficient transmission of key data.

[0006] The above objectives can be achieved through the following solutions: A UAV inspection method and system based on AI vision control, including generating a dynamic three-dimensional space vector model according to discontinuous frame image data collected by an on-board camera and pre-set inertial navigation data; extracting and compressing feature components of the dynamic three-dimensional space vector model to generate a compressed feature stream; performing multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map; constructing a safety corridor according to the obstacle avoidance decision priority map, and using the safety corridor to generate a real-time self-correcting flight path; performing semantic analysis on the deviation between the real-time flight trajectory and the real-time self-correcting flight path to generate target area anomaly marking data; dynamically adjusting the bandwidth allocation of a dual-channel communication network, and transmitting the target area anomaly marking data to a ground control terminal.

[0007] Optionally, generating a dynamic three-dimensional space vector model based on the discontinuous frame image data collected by the airborne camera and the preset inertial navigation data includes: obtaining a scene image sequence and real-time inertial attitude parameters under a preset discontinuous frame sampling period; performing key feature point clustering on the scene image sequence to generate spatially distributed vector nodes; aligning the spatially distributed vector nodes and the real-time inertial attitude parameters in space and time to form a dynamic three-dimensional space vector model.

[0008] Optionally, extracting and compressing feature components of the dynamic three-dimensional space vector model to generate a compressed feature stream includes: performing feature hierarchical encoding on the dynamic three-dimensional space vector model to generate an environmental multi-dimensional feature matrix; screening out the main components in the environmental multi-dimensional feature matrix through a preset feature entropy value threshold to generate a compressed feature stream.

[0009] Optionally, performing multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map includes: performing semantic segmentation on the compressed feature stream to generate an obstacle contour feature vector; combining a preset obstacle type database to evaluate the safety factor of the obstacle contour feature vector to generate an obstacle avoidance decision priority map.

[0010] Optionally, constructing a safety corridor according to the obstacle avoidance decision priority map and using the safety corridor to generate a real-time self-correcting flight path includes: establishing a safety corridor according to the obstacle avoidance decision priority map; searching for feasible paths using the safety corridor and screening out the feasible paths that meet the preset flight path continuity rules to obtain a candidate path set; setting path weight coefficients for each candidate path in the candidate path set to generate an emergency path library; screening out a target path from the emergency path library according to the magnitude of the path weight coefficient and performing secondary optimization on the target path to generate a real-time self-correcting flight path.

[0011] Optionally, the method further includes: obtaining inertial motion parameters of the UAV in the current flight state, calculating the future flight path collision probability through the inertial motion parameters to generate a risk distribution heat map; performing superposition analysis on the risk distribution heat map and the obstacle avoidance decision priority map to generate a path weight coefficient.

[0012] Optionally, performing semantic parsing on the deviation between the real-time flight path and the real-time self-correcting flight path to generate target area anomaly marking data includes: obtaining the geometric topology difference parameters between the preset standard inspection path and the real-time flight path; using the geometric topology difference parameters to generate multi-layer semantic anomaly marking data; using the multi-layer semantic anomaly marking data to perform anomaly type recognition to generate target area anomaly marking data.

[0013] Optionally, the generating of the multi-level semantic anomaly marking data by using the geometric topology difference parameter includes: performing joint analysis on the geometric topology difference parameter to generate a confidence score; judging the magnitude relationship between the confidence score and a preset anomaly region confidence threshold; and generating the multi-level semantic anomaly marking data according to the judgment result.

[0014] Optionally, the dynamically adjusting the bandwidth allocation of the dual-channel communication network and transmitting the target region anomaly marking data to the ground control terminal includes: real-time monitoring of the signal quality fluctuation parameter and the transmission load status parameter of the primary and standby channels; calculating a channel score by using the signal quality fluctuation parameter and the transmission load status parameter; allocating bandwidth resources according to the channel score to generate a channel switching decision instruction; and adjusting the bandwidth allocation of the dual-channel communication network according to the channel switching decision instruction to transmit the target region anomaly marking data to the ground control terminal.

[0015] Based on the same inventive concept, the present invention also provides an unmanned aerial vehicle inspection system based on AI vision control. The system includes: a dynamic three-dimensional modeling module for generating a dynamic three-dimensional space vector model according to discontinuous frame image data collected by an on-board camera and preset inertial navigation data; a feature compression module for extracting and compressing feature components of the dynamic three-dimensional space vector model to generate a compressed feature stream; an obstacle avoidance decision module for performing multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map; a flight path self-correction module for constructing a safety corridor according to the obstacle avoidance decision priority map and using the safety corridor to generate a real-time self-corrected flight path; an anomaly detection module for performing semantic analysis on the deviation between the real-time flight trajectory and the real-time self-corrected flight path to generate target region anomaly marking data; and a communication optimization module for dynamically adjusting the bandwidth allocation of the dual-channel communication network and transmitting the target region anomaly marking data to the ground control terminal.

[0016] Compared with the prior art, the present invention has the following advantages: 1. The present invention constructs a dynamic three-dimensional vector model by fusing discontinuous frame images and inertial navigation data, effectively reducing the redundancy of data processing; the discontinuous sampling mechanism reduces the computing load of the edge computing end, and combined with the spatio-temporal alignment technology, it improves the environmental modeling accuracy, significantly improving the modeling efficiency and anti-interference ability in complex scenarios; 2. By using the feature entropy value threshold to screen the main components and combining the multi-modal obstacle safety level analysis, an efficient obstacle avoidance decision is realized; through semantic segmentation and dynamic feature extraction, the type and threat level of obstacles are accurately identified, and combined with the safety corridor path optimization mechanism, it ensures the real-time performance and safety of the obstacle avoidance response of the unmanned aerial vehicle in a dynamic environment; 3. An anomaly detection method based on a risk distribution heat map and semantic parsing enhances the reliability of inspection target recognition; through geometric topology difference analysis and multi-layer semantic tagging, precise positioning and classification of anomaly regions are achieved, solving the problem of insufficient adaptability of traditional methods to sudden environmental changes; 4. A dynamic dual-channel bandwidth allocation strategy ensures the integrity of critical data transmission; through the fusion of channel quality assessment and priority matrix, the network resource allocation mechanism is optimized, effectively coping with communication fluctuations in complex electromagnetic environments, and improving the real-time performance and robustness of anomaly data backhaul.

[0017] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures pointed out in the specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following-described drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a schematic flow chart of an unmanned aerial vehicle inspection method based on AI vision control according to an embodiment of the present invention.

[0020] Figure 2 It is a three-dimensional coordinate diagram of the distribution of non-continuous frame feature points according to an embodiment of the present invention.

[0021] Figure 3 It is a schematic diagram of a three-dimensional space vector model according to an embodiment of the present invention.

[0022] Figure 4 It is a characteristic information entropy distribution diagram according to an embodiment of the present invention.

[0023] Figure 5 It is a self-correcting flight path comparison diagram according to an embodiment of the present invention.

[0024] Figure 6 It is a main and backup channel quality score change diagram according to an embodiment of the present invention.

[0025] Figure 7 It is a schematic structural diagram of an unmanned aerial vehicle inspection system based on AI vision control according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Referring to Figure 1 , an embodiment of the present invention proposes a method for unmanned aerial vehicle inspection based on AI vision control. By adopting a fusion modeling mechanism of discontinuous frame images and inertial navigation, and combining a multi-modal obstacle safety grading and dual-channel dynamic bandwidth allocation strategy, it can achieve real-time response for dynamic obstacle avoidance, accurate discrimination of abnormal detection, and efficient transmission of key data.

[0028] The specific steps of the method in this embodiment include: Generate a dynamic three-dimensional space vector model based on the discontinuous frame image data collected by the on-board camera and the preset inertial navigation data; Extract and compress the feature components of the dynamic three-dimensional space vector model to generate a compressed feature stream; Perform multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map; Construct a safety corridor according to the obstacle avoidance decision priority map, and use the safety corridor to generate a real-time self-correcting flight path; Perform semantic analysis on the deviation between the real-time flight trajectory and the real-time self-correcting flight path to generate target area anomaly marking data; Dynamically adjust the bandwidth allocation of the dual-channel communication network and transmit the target area anomaly marking data to the ground control terminal.

[0029] Specifically, based on the fusion modeling principle of discontinuous frame image sampling and inertial navigation data, an image sequence is collected through a preset discontinuous frame interval, significantly reducing the computational amount of redundant data. At the same time, a spatio-temporal associated vector node network is constructed by combining real-time attitude parameters of inertial navigation. The spatial information of discrete image frames and continuous inertial data are dynamically bound by using the key feature point clustering and spatio-temporal alignment algorithm to generate a three-dimensional space vector model that can be updated in real time. This method breaks through the computing power bottleneck of the traditional continuous frame processing mode and realizes the balanced optimization of the dynamic environment representation ability. Through the heterogeneous sensor data fusion and dynamic modeling mechanism, while ensuring the environmental perception accuracy, the computational efficiency is greatly improved, effectively adapting to complex meteorological or mobile interference scenarios. The combined characteristics of discontinuous sampling and inertial compensation reduce the resource consumption at the edge, enabling the UAV to maintain high-frequency model iteration under limited computing power and supporting the basis for high-precision obstacle avoidance decision-making; the spatio-temporal alignment algorithm enhances the stability and scalability of the three-dimensional vector model in dynamic scenarios, providing anti-interference spatial information support for subsequent path planning.

[0030] Optionally, generating a dynamic three-dimensional space vector model according to the discontinuous frame image data collected by the airborne camera and the preset inertial navigation data includes: Obtaining a scene image sequence and real-time inertial attitude parameters under a preset discontinuous frame sampling period; Performing key feature point clustering on the scene image sequence to generate spatially distributed vector nodes; Performing spatio-temporal alignment on the spatially distributed vector nodes and the real-time inertial attitude parameters to form a dynamic three-dimensional space vector model.

[0031] Specifically, the distribution of key feature points in discontinuous frames is as Figure 2 shown. A scene image sequence is collected by a camera carried by a UAV according to a preset sampling period, and the preset period is a discontinuous frame capture mode with an interval of 5 - 30 seconds. At the same time, real-time inertial attitude parameters are obtained through an airborne IMU inertial measurement unit, and the real-time inertial attitude parameters include the pitch angle, roll angle, yaw angle, and three-axis acceleration components of the UAV. Key feature point clustering processing is performed on the scene image sequence. The SURF feature detection algorithm is used to extract a set of stable feature points for each image frame, and the feature points whose spatial positions appear more than a preset threshold in a continuous preset number of sampling frames are marked as key feature points. For the coordinates v of the spatially distributed vector nodes, there is: ; In the formula represents the frequency weight of the i-th key feature point, is the three-dimensional coordinate of the i-th key feature point in the image coordinate system. Finally, a three-dimensional point cloud structure composed of multiple vector nodes is generated. Align the timestamps corresponding to each spatially distributed vector node with the inertial attitude parameter time series. Use the Kalman filter algorithm to perform compensation calculations on the spatio-temporal displacement vector to eliminate the coordinate offset caused by the movement of the UAV. Map the vector nodes after displacement compensation to the IMU coordinate system to form a dynamic three-dimensional space vector model including a spatio-temporal alignment vector network, as Figure 3 shown.

[0032] Exemplarily, in the mountainous high-voltage line inspection scenario, the above steps are implemented. The UAV captures a set of discontinuous image sequences every 15 seconds and synchronously records the flight attitude parameters. It is found that the corner point features of the transmission tower appear stably in 3 consecutive samplings during the feature point clustering process, and high weights are assigned to them while the features of floating objects are filtered out because the appearance frequency is lower than the threshold of 2 times / period. In the vector model established after spatio-temporal alignment, the coordinate accuracy of the transmission tower structure is improved to ±0.3 meters, and the computing load is reduced by 60% compared with the traditional continuous frame modeling method. Discontinuous sampling reduces the amount of redundant data processing and improves the resource utilization efficiency of the edge computing module; the frequency weighting mechanism makes the modeling process have dynamic anti-interference ability, and the spatial position features of the infrastructure can still be accurately locked under complex working conditions.

[0033] Optionally, the extracting and compressing feature components of the dynamic three-dimensional space vector model to generate a compressed feature stream includes: Performing feature hierarchical encoding on the dynamic three-dimensional space vector model to generate an environmental multi-dimensional feature matrix; Screening the main components in the environmental multi-dimensional feature matrix through a preset feature entropy value threshold to generate a compressed feature stream.

[0034] Specifically, first perform feature hierarchical encoding on the dynamic three-dimensional space vector model. The dynamic three-dimensional space vector model is composed of several spatially distributed vector nodes, and each vector node contains three-dimensional coordinate information and associated acquisition timestamps. The edge computing terminal uses a multi-layer convolutional network to perform local feature extraction on the spatial distribution of vector nodes to generate a primary feature matrix containing geometric structure information, texture information, and dynamic change information. The cloud computing terminal receives the primary feature matrix and uses the global feature correlation algorithm to perform vertical fusion on the features of each layer to generate an environmental multi-dimensional feature matrix, where the rows of the matrix correspond to spatially distributed vector nodes and the columns represent different levels of feature dimensions. Screen the main components in the environmental multi-dimensional feature matrix through a preset feature entropy value threshold. The feature entropy value is calculated through the information entropy formula: , where is the information entropy of the feature dimension, is the normalized probability distribution value of the i-th eigenvalue among all spatial vector nodes. In specific calculations, the frequency distribution of eigenvalues in each column of the statistical environment multi-dimensional feature matrix is counted, and after normalization, we get . Set the entropy threshold as an empirically preset value of the system. When , it is determined that this feature dimension contains valid environmental information and is retained as a key component; when , it is determined as a redundant component and deleted. Finally, the filtered feature dimensions are reorganized in the order of spatial vector nodes to generate a compressed feature stream with reduced dimensions. In the above process, feature hierarchical coding realizes the efficient integration of local and global features, reducing the redundant data volume of the original three-dimensional model; feature entropy value screening is as Figure 4 shown, and the feature entropy threshold screening eliminates low-value noise features by quantifying information entropy and retains the main components that make significant contributions to obstacle recognition.

[0035] Exemplarily, in the power line inspection scenario, the edge computing terminal extracts the structural features of the insulator string of the transmission tower and the dynamic change features of the conductor sag, forming a primary feature matrix including geometric shape, surface rust grade, and motion trajectory. Each feature dimension corresponds to a 32-bit floating-point value. The cloud computing terminal performs longitudinal fusion on the feature matrices collected continuously for 5 minutes to generate an environmental multi-dimensional feature matrix of 1000 rows by 50 columns. By calculating the entropy value of each column, it is found that the wire vibration frequency dimension H = 2.3, which is higher than ; while the background cloud displacement dimension H = 0.5, which is lower than the threshold and is eliminated. The finally retained 30 columns of key features are reorganized to generate a compressed feature stream, and the data volume is reduced by 40%. Through the hierarchical coding mechanism, local detailed features and global context information are organically integrated, enabling the compressed feature stream to not only represent the microscopic state of the UAV inspection target but also reflect the macroscopic change law of the environment. The screening method based on information entropy effectively filters out invalid feature dimensions introduced by weather interference or sensor noise, reducing the computational complexity of the subsequent obstacle avoidance decision-making algorithm. The compressed feature stream retains the core features such as high-frequency vibration and structural deformation of the key components of the transmission line, ensuring the accuracy of safety level analysis, and at the same time significantly reducing the transmission bandwidth requirements between the edge and the cloud.

[0036] Optionally, the multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map includes: Performing semantic segmentation on the compressed feature stream to generate an obstacle contour feature vector; Combining with a preset obstacle type database to evaluate the safety factor of the obstacle contour feature vector and generate an obstacle avoidance decision priority map.

[0037] Specifically, first, semantic segmentation processing is performed on the compressed feature stream. The compressed feature stream is matrix-type data after dimensionality reduction, where each row corresponds to a spatial distribution vector node and each column represents the retained key feature dimensions. A deformable convolutional neural network is used to divide the compressed feature stream in the spatial domain, and a semantic segmentation mask is constructed based on the correlation between adjacent nodes in the feature matrix to identify potential obstacle blocks in the environmental scene. The dynamic change features of the obstacle blocks are extracted through a bidirectional LSTM network, including the motion speed vector and the shape change trend parameters, and encoded into an obstacle contour feature vector. Then, classification reference features are called from a preset obstacle type database. The database includes standard feature templates for fixed obstacle categories and moving obstacle categories. For example, the geometric shapes, reflectivities, and motion pattern features of typical obstacles such as insulator strings, tower structures, hanging objects, and bird flocks in a transmission line scene. Similarity matching calculations are performed between the obstacle contour feature vector and the classification reference features: , wherein, is the safety factor, with a value range of 0 - 1; represents the obstacle contour feature vector; represents the standard feature vector of a certain type of obstacle in the database; is the number of dimensions of the feature vector. By traversing the obstacle categories in the database, all values are calculated and the obstacle type corresponding to the maximum value is selected as the current obstacle classification result. Subsequently, safety level parameters are assigned according to the classification result, and the mapping rule of the safety level L is set: when identified as fixed infrastructure, L = 1, indicating no need to avoid; when identified as a low-dynamic obstacle, L = 2, indicating that path optimization is required; when identified as a high-risk moving obstacle, L = 3, indicating that emergency obstacle avoidance must be carried out. The spatial coordinates of each obstacle and its safety level parameters are integrated into a dynamic grid map to generate an obstacle avoidance decision priority map containing obstacle distribution density, motion direction, and safety level.

[0038] Exemplarily, in the inspection task of the power pipe gallery in the industrial park, the drone detects an abnormal moving target in a certain area through the compressed feature stream. Semantic segmentation processing separates the target block from the background, and the extraction of contour feature vectors shows that it has a horizontal displacement of 0.2 meters per second and an elliptical shape feature. After matching with the database, it is determined that the similarity of the target to the inspection robot is S = 0.8, and the similarity to the bird flock is S = 0.3. Accordingly, it is classified as a low-dynamic obstacle and L = 2 is set. Combining its position information, an orange warning area is generated in the priority map, triggering a smooth offset correction of the drone flight path. The combination of semantic segmentation and dynamic feature extraction can accurately identify the physical characteristics and motion laws of obstacles, avoiding misjudgment caused by environmental noise; the classification mechanism based on the safety factor enables the system to differentially process obstacles with different threat levels, taking into account both the inspection efficiency and flight safety; the priority map provides an intuitive decision-making basis through visual spatial annotation, especially significantly reducing the complexity of obstacle avoidance path planning in the mixed obstacle scenario, ensuring the drone's precise obstacle avoidance and continuous operation capabilities in the long and narrow pipe gallery environment.

[0039] Optionally, constructing a safety corridor according to the obstacle avoidance decision priority map and generating a real-time self-correcting flight path using the safety corridor includes: Establishing a safety corridor according to the obstacle avoidance decision priority map; Searching for feasible paths using the safety corridor and screening the feasible paths that meet the preset flight path continuity rules to obtain a set of candidate paths; Setting path weight coefficients for each candidate path in the set of candidate paths to generate an emergency path library; Screening the target path from the emergency path library according to the magnitude of the path weight coefficients and performing secondary optimization on the target path to generate a real-time self-correcting flight path.

[0040] Specifically, the self-correcting flight path is as Figure 5 shown. First, based on the spatial coordinates and safety level parameters of each obstacle marked in the obstacle avoidance decision priority map, a three-dimensional safety bounding box is established. For each high-risk moving obstacle with a safety level of L = 3, with the center point of the obstacle as the reference, a spherical safety boundary is generated by expanding according to a preset expansion coefficient, and the expansion coefficient is set according to 1.2 times the wingspan of the drone. For low-dynamic obstacles with a safety level of L = 2, a cylindrical safety boundary is extended along its moving direction, and the length of the cylinder is the product of the moving speed of the obstacle and the preset reaction time. Boolean union operations are performed on all safety boundaries to generate a three-dimensional safety corridor model containing multiple protection areas. The construction of the emergency path library searches for feasible paths in the safety corridor through the Rapidly-exploring Random Tree (RRT) algorithm, and a dynamic weight adjustment mechanism is introduced during the sampling process. The node selection probability is calculated by the formula: , Wherein, d represents the Euclidean distance between the sampling node and the target waypoint, k is the path continuity weight factor, and C is the normalization constant. The value of k is dynamically updated according to the current flight speed and attitude angle parameters of the UAV. When the UAV is in a high-speed flight state, the value of k is increased to preferentially select a path segment with a smaller route curvature as the feasible path. The feasible paths that meet the route continuity rule are marked as candidate paths and the path weight coefficients are set. An emergency path library is constructed using the candidate paths and the corresponding path weight coefficients. The route continuity rule requires that the change in the heading angle between adjacent path nodes does not exceed 5 degrees and the acceleration volatility is lower than the preset threshold. The candidate path with the largest path weight coefficient is selected from the emergency path library as the target path. The generation of the real-time self-correcting route uses a non-linear optimization algorithm to perform a secondary optimization on the target path. The objective function is established as follows: , Wherein, is the change value of the heading angle between adjacent path segments, is the acceleration requirement of the path point, , are the heading smoothing weight and the motion smoothness weight respectively, and their values are configured according to the model performance parameters of the UAV. The minimum value of the objective function is iteratively solved by the gradient descent method, and finally a self-correcting route that takes into account safety and flight stability is generated.

[0041] Optionally, the method further includes: Obtaining the inertial motion parameters of the UAV in the current flight state, calculating the future trajectory collision probability through the inertial motion parameters, and generating a risk distribution heat map; Overlaying and analyzing the risk distribution heat map and the obstacle avoidance decision priority map to generate a path weight coefficient.

[0042] Specifically, obtaining the inertial motion parameters of the UAV in the current flight state includes the real-time velocity vector, acceleration vector, and attitude angle change rate of the UAV in three-dimensional space. A future trajectory prediction time window is set in the on-board processor, and is taken as the reaction time corresponding to twice the braking distance of the current speed value of the UAV. A trajectory prediction model is established according to the inertial motion parameters. The model input is the current speed and the acceleration , and the output is the predicted position coordinates at the future time t: , Wherein , is the vector composed of the three axial velocity components, is the real-time acceleration vector. Based on the predicted trajectory and the obstacle distribution in the obstacle avoidance decision priority map, spatial correlation analysis is performed. For each predicted moment , calculate the shortest Euclidean distance between each obstacle bounding box and , and compare it with the preset dynamic safety threshold , which is set to 1.5 times the wingspan length according to the UAV type. When , it is counted as a potential collision event, and the collision probability is calculated by the formula: , where is the adjustment coefficient, calibrated by the UAV dynamic response parameters. Map the of each time slice to the three-dimensional space to generate a risk distribution heat map. The risk density value of each voxel in the map is the weighted sum of all within its range. Align the risk distribution heat map with the obstacle avoidance decision priority map in spatial grids, and perform multi-dimensional parameter superposition at the corresponding coordinate points to generate a comprehensive influence factor matrix. The path weight coefficient is calculated by the formula: , where is the obstacle avoidance priority weight coefficient, is the risk heat map weight coefficient, is the safety level parameter of all obstacles covered by the path segment, is the number of grids covered by the path segment. The weight coefficients , are dynamically adjusted according to the UAV operation mode, and increase the value when the speed exceeds the threshold to improve the risk avoidance priority.

[0043] By fusing real-time kinematic prediction and prior obstacle avoidance data, the path weight allocation mechanism can prospectively identify dynamic risks and effectively solve the threat of sudden moving obstacles. The direct calculation of inertial motion parameters combined with the distance attenuation function ensures that the risk probability model conforms to the real physical laws. The multi-dimensional superposition of the heat map and the priority map significantly enhances the decision-making robustness in complex scenarios, especially in areas with dense equipment, preventing route conflicts caused by inertial navigation errors or sudden changes in obstacles.

[0044] Optionally, the semantic parsing of the deviation between the real-time flight trajectory and the real-time self-correcting flight path generates target area anomaly marking data including: Obtain the geometric topology difference parameters between the preset standard inspection path and the real-time flight trajectory;​ Generate multi - layer semantic anomaly marking data by using the geometric topology difference parameter; Use the multi - layer semantic anomaly marking data to identify the anomaly type and generate the target area anomaly marking data.

[0045] Optionally, the generating multi - layer semantic anomaly marking data by using the geometric topology difference parameter includes: Conduct a joint analysis on the geometric topology difference parameter to generate a confidence score; Judge the magnitude relationship between the confidence score and the preset anomaly area confidence threshold; Generate multi - layer semantic anomaly marking data according to the judgment result.

[0046] Specifically, obtain the preset standard inspection path of the unmanned aerial vehicle and the real - time flight trajectory data. The standard inspection path includes a serialized GPS coordinate point set and corresponding flight altitude constraint parameters. The real - time flight trajectory consists of longitude, latitude, altitude, and timestamp output by the on - board integrated navigation system. Compare the spatial topological relationships of the path nodes synchronized at each timestamp, and calculate the three - dimensional Euclidean distance error and the path - heading angle deviation . Specifically, Calculate point - by - point using the following formula: , where , , represent the three - dimensional coordinates of the real - time flight trajectory nodes, , , are the three - dimensional coordinates of the standard inspection path nodes at the corresponding time. Obtained by calculating the angle between the real - time heading vector and the standard path vector, and the vector direction is determined by the connection line of adjacent path nodes. The three - dimensional Euclidean distance error , the path - heading angle deviation combined with the anomaly area parameter constitute the geometric topology difference parameter. Use the geometric topology difference parameter to calculate the joint feature vector : , where represents the joint eigenvalue of the j - th anomaly area; , , are preset weighting coefficients determined by training with historical anomaly data; is the cumulative sum of the distance errors of consecutive waypoints; represents the time change rate of the angle deviation; Calculated from the area of the minimum circumscribed rectangle of the projection of the abnormal area on the horizontal plane. The confidence score is calculated using the joint eigenvalue : , In the formula is the confidence level, is the attenuation coefficient. Compare with the preset confidence threshold of the abnormal area (taking values in the range of 0.7 - 0.9). When , it is determined as a valid detection target, otherwise it is regarded as false alarm noise. Hierarchical analysis is performed on the joint feature vector of the valid detection target. The error distribution pattern is extracted at the first level, and it is determined as the type of equipment body offset or environmental interference; the spatial expansion characteristics of the abnormal area are analyzed at the second level, and local point anomalies and continuous surface anomalies are classified; the anomaly persistence is verified by combining historical inspection data at the third level, and finally multi-layer semantic anomaly marking data including anomaly category, influence range, and confidence level is generated. Identify the anomaly type according to the information contained in the multi-layer semantic anomaly marking data, and summarize to form the anomaly marking data of the target area.

[0047] Optionally, the dynamic adjustment of the bandwidth allocation of the dual-channel communication network and the transmission of the anomaly marking data of the target area to the ground control terminal include: Real-time monitor the signal quality fluctuation parameters and transmission load status parameters of the primary and backup channels; Calculate the channel score using the signal quality fluctuation parameters and the transmission load status parameters; Allocate bandwidth resources according to the channel score and generate a channel switching decision instruction; Adjust the bandwidth allocation of the dual-channel communication network according to the channel switching decision instruction, and transmit the anomaly marking data of the target area to the ground control terminal.

[0048] Specifically, the quality scores of the primary and backup channels are as Figure 6 shown. The communication status parameters of the primary channel and the backup channel are collected in real time. The signal quality fluctuation parameters are obtained through the physical layer detection of the airborne communication module, including the received signal strength indication RSSI, the bit error rate BER, and the channel delay fluctuation value. RSSI is obtained by converting the received power of the wireless signal. BER statistically calculates the ratio of the number of error bits to the total number of transmitted bits per unit time. The channel delay fluctuation value measures the standard deviation of the round-trip time RTD of the data packet. The transmission load status parameters include the current channel bandwidth occupancy rate and the data queue backlog . The bandwidth occupancy rate is calculated by dividing the currently used bandwidth by the theoretical maximum bandwidth of the channel. The data queue backlog statistic is the total number of bytes of the data packets to be transmitted. A transmission data priority matrix is constructed, and the abnormal marked data in the target area is defined as the highest priority category P1, and other inspection data is classified into levels P2 to P4 according to the information type. The rows of the priority matrix correspond to the data types, and the columns include the priority weight, the minimum bandwidth requirement, and the maximum allowable delay. The priority weight is preset according to the urgency of the exception handling, and the weight of category P1 is the highest. The bandwidth allocation decision function calculates the comprehensive channel quality score of each channel: , where is the comprehensive score of the channel, which is used to quantify the transmission capacity of the channel. is the normalized received signal strength, BER is the current bit error rate value, is the bandwidth occupancy rate, is the normalized value of the data queue backlog. , , , are the weight coefficients, and their values are dynamically adjusted according to the UAV flight altitude and the environmental interference intensity. When in a low-altitude complex environment, the coefficient is increased to enhance the weight of the influence of the bit error rate. The real-time values of the primary and backup channels are calculated and compared independently. When the of the primary channel is lower than that of the backup channel and the difference exceeds the preset threshold , the channel switch is triggered. The bandwidth allocation algorithm performs two-phase operations according to the priority matrix: the first phase is the hard allocation to ensure that all P1-class data obtains its minimum bandwidth requirement ; the second phase is the dynamic allocation, and the bandwidth is allocated to other data types according to the proportion of the priority weights in the remaining bandwidth. When the available bandwidth of any channel cannot meet the of the P1-class data, the dual-channel parallel transmission mode is immediately started, so that the P1 data fragments are transmitted simultaneously through the primary and backup channels, and the data integrity is guaranteed through redundant verification.

[0049] Exemplarily, in the inspection task of an offshore oil and gas platform, the UAV detects pipeline corrosion anomalies and generates P1-level abnormal marked data. At this time, the RSSI of the primary channel decreases and the BER increases due to the reflection of the sea waves; the backup channel remains unchanged through the satellite link. It is calculated that the Score of the primary channel is 0.65, and the Score of the backup channel is 0.82, and the difference exceeds Threshold-triggered switching. The system preferentially allocates 2 Mbps of bandwidth to ensure the transmission of abnormal data, and the remaining 4 Mbps of bandwidth is allocated to the visible light inspection video stream. When a strong wind disturbance causes a short interruption in the satellite link, the algorithm automatically enables dual-channel parallel transmission, splitting the abnormal data into encrypted data blocks and transmitting them through the remaining available maritime radio channels. After the satellite resumes, the integrity check shows no packet loss. The dual-channel management mechanism based on the comprehensive channel quality score and priority matrix can autonomously select the optimal transmission path in a complex electromagnetic environment to ensure the real-time and reliable transmission of critical abnormal data. The dynamic bandwidth allocation strategy takes into account the transmission requirements of different data types, giving priority to high-priority services when the channel conditions deteriorate to avoid the loss of detection information caused by congestion in a single channel. The dual-channel parallel transmission and redundancy check mechanism effectively address the problem of link instability in extreme environments, significantly improving the success rate of transmitting inspection data back to the ground for UAVs in adverse weather and complex terrain areas.

[0050] Based on the same inventive concept, as Figure 7 shown, the present invention also provides an unmanned aerial vehicle (UAV) inspection system based on AI vision control, the system comprising: A dynamic three-dimensional modeling module, configured to generate a dynamic three-dimensional space vector model according to discontinuous frame image data collected by an on-board camera and preset inertial navigation data; A feature compression module, configured to extract and compress feature components of the dynamic three-dimensional space vector model to generate a compressed feature stream; An obstacle avoidance decision module, configured to perform multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map; A flight path self-correction module, configured to construct a safety corridor according to the obstacle avoidance decision priority map and use the safety corridor to generate a real-time self-corrected flight path; An anomaly detection module, configured to perform semantic analysis on the deviation between the real-time flight trajectory and the real-time self-corrected flight path to generate target area anomaly marking data; A communication optimization module, configured to dynamically adjust the bandwidth allocation of a dual-channel communication network and transmit the target area anomaly marking data to a ground control terminal.

[0051] It should be noted that the electrical connections between the above-mentioned various units do not necessarily represent direct connections of the lines. Indirect connection methods, as long as they can achieve the purpose of the present invention, are applicable to the embodiments of the present invention. The above are only exemplary embodiments of the present invention and cannot be used to limit the scope of the present invention.

[0052] That is, any equivalent changes and modifications made in accordance with the teachings of the present invention still fall within the scope covered by the present invention. Those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and the disclosure of the practical truth. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include the common general knowledge or conventional technical means in the technical field not recorded in the present invention.

Claims

1. An unmanned aerial vehicle inspection method based on AI vision control, characterized in that, The method includes: Generating a dynamic three-dimensional space vector model based on the discontinuous frame image data collected by an airborne camera and the preset inertial navigation data; Performing feature component extraction and compression on the dynamic three-dimensional space vector model to generate a compressed feature stream; Performing multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map; Constructing a safety corridor according to the obstacle avoidance decision priority map and using the safety corridor to generate a real-time self-correcting flight path; Performing semantic analysis on the deviation between the real-time flight trajectory and the real-time self-correcting flight path to generate target area anomaly marking data; Dynamically adjusting the bandwidth allocation of the dual-channel communication network and transmitting the target area anomaly marking data to the ground control terminal.

2. The method for drone inspection based on AI vision control according to claim 1, wherein The generating of a dynamic three-dimensional space vector model based on the discontinuous frame image data collected by an airborne camera and the preset inertial navigation data includes: Obtaining a scene image sequence and real-time inertial attitude parameters under a preset discontinuous frame sampling period; Performing key feature point clustering on the scene image sequence to generate spatially distributed vector nodes; Performing spatio-temporal alignment on the spatially distributed vector nodes and the real-time inertial attitude parameters to form a dynamic three-dimensional space vector model.

3. The method for drone inspection based on AI vision control according to claim 1, characterized in that, The performing of feature component extraction and compression on the dynamic three-dimensional space vector model to generate a compressed feature stream includes: Performing feature hierarchical coding on the dynamic three-dimensional space vector model to generate an environmental multi-dimensional feature matrix; Screening the principal components in the environmental multi-dimensional feature matrix through a preset feature entropy value threshold to generate a compressed feature stream.

4. The method for UAV inspection based on AI vision control according to claim 1, wherein, The performing of multi-modal obstacle safety level parameter analysis based on the compressed feature stream to generate an obstacle avoidance decision priority map includes: Performing semantic segmentation on the compressed feature stream to generate an obstacle contour feature vector; Combining with a preset obstacle type database to evaluate the safety factor of the obstacle contour feature vector to generate an obstacle avoidance decision priority map.

5. The method for drone inspection based on AI vision control according to claim 1, wherein The constructing of a safety corridor according to the obstacle avoidance decision priority map and using the safety corridor to generate a real-time self-correcting flight path includes: Establishing a safety corridor according to the obstacle avoidance decision priority map; Searching for feasible paths using the safety corridor and screening the feasible paths that meet the preset flight path continuity rules to obtain a candidate path set; Setting path weight coefficients for each candidate path in the candidate path set to generate an emergency path library; Screening a target path from the emergency path library according to the magnitude of the path weight coefficients and performing secondary optimization on the target path to generate a real-time self-correcting flight path.

6. The method for inspecting drones based on AI vision control according to claim 5, wherein The method further includes: Obtaining the inertial motion parameters of the UAV in the current flight state, calculating the future trajectory collision probability through the inertial motion parameters, and generating a risk distribution heat map; Performing overlay analysis on the risk distribution heat map and the obstacle avoidance decision priority map to generate path weight coefficients.

7. An unmanned aerial vehicle inspection method based on AI vision control according to claim 1, characterized in that, The performing of semantic analysis on the deviation between the real-time flight trajectory and the real-time self-correcting flight path to generate target area anomaly marking data includes: Obtaining the geometric topology difference parameters between the preset standard inspection path and the real-time flight trajectory; Using the geometric topology difference parameters to generate multi-layer semantic anomaly marking data; Use the multi-layer semantic anomaly marking data to identify the anomaly type and generate the anomaly marking data for the target area.

8. The UAV inspection method based on AI vision control according to claim 7, wherein, The generating of the multi-layer semantic anomaly marking data by using the geometric topology difference parameter includes: Perform joint analysis on the geometric topology difference parameter to generate a confidence score; Judge the magnitude of the confidence score and a preset confidence threshold for the anomaly area; Generate the multi-layer semantic anomaly marking data according to the judgment result.

9. An unmanned aerial vehicle inspection method based on AI vision control according to claim 7, characterized in that, The dynamically adjusting the bandwidth allocation of the dual-channel communication network and transmitting the anomaly marking data for the target area to the ground control terminal includes: Real-time monitor the signal quality fluctuation parameter and the transmission load status parameter of the primary and backup channels; Use the signal quality fluctuation parameter and the transmission load status parameter to calculate a channel score; Allocate bandwidth resources according to the channel score to generate a channel switching decision instruction; Adjust the bandwidth allocation of the dual-channel communication network according to the channel switching decision instruction and transmit the anomaly marking data for the target area to the ground control terminal.

10. A drone inspection system based on AI vision control, which is applied to the AI vision control-based drone inspection method described in any one of claims 1-9, and is characterized in that, The system includes: A dynamic 3D modeling module for generating a dynamic 3D space vector model according to the discontinuous frame image data collected by the airborne camera and the preset inertial navigation data; A feature compression module for extracting and compressing the feature components of the dynamic 3D space vector model to generate a compressed feature stream; An obstacle avoidance decision module for analyzing the multi-modal obstacle safety level parameters based on the compressed feature stream to generate an obstacle avoidance decision priority map; A flight path self-correction module for constructing a safety corridor according to the obstacle avoidance decision priority map and using the safety corridor to generate a real-time self-corrected flight path; An anomaly detection module for performing semantic analysis on the deviation between the real-time flight trajectory and the real-time self-corrected flight path to generate the anomaly marking data for the target area; A communication optimization module for dynamically adjusting the bandwidth allocation of the dual-channel communication network and transmitting the anomaly marking data for the target area to the ground control terminal.

Images