Unmanned aerial vehicle autonomous inspection system and method based on AI identification

Abstract

The invention discloses an unmanned aerial vehicle autonomous inspection system and method based on AI identification, the system comprises an unmanned aerial vehicle body carrying a high-power optical zoom lens, an edge computing device and a function module, and the autonomous inspection of a power distribution tower is realized by fusing front-end AI identification and edge computing technologies. A Yolo structure is adopted to construct a lightweight tower recognition model, a dynamic route planning module is combined to realize single-point reference route generation and three obstacle crossing modes, and real-time coordinate correction in an RTK-free environment is supported. The edge computing device integrates a semi-supervised learning engine and a multi-sensor data fusion module, meets miniaturization design, and supports breakpoint continuous flight control and precise landing. The method covers automatic route generation, visual tracking, zoom cooperative control and self-adaptive task scheduling, solves the problems that a traditional unmanned aerial vehicle depends on manual operation, the data quality is poor and the efficiency is low, realizes whole-course automation of tower inspection in a complex environment, and is high in inspection efficiency, and the picture definition reaches the pin level.

Description

Technical Field

[0001] The present invention relates to the field of intelligent inspection technology, and in particular to an autonomous inspection system and method for unmanned aerial vehicles based on AI recognition. Background Art

[0002] With the rapid development of smart grid construction, the demand for distribution line inspections is growing. Traditional manual inspections suffer from low efficiency, high risk, and limited coverage. The application of drone technology offers a new technical path for power line inspections. Existing drones are widely used for transmission line inspections, using visible light / infrared equipment to perform basic inspections of tower components. However, the following technical bottlenecks remain in the field of distribution line inspections:

[0003] High reliance on manual labor: Current drone operations require professional pilots, and operator training cycles are long (usually 3-6 months). Manual operation can easily lead to inconsistent shooting angles and uneven photo quality, affecting the accuracy of subsequent defect analysis. Statistics show that the pass rate of standardized photos from manual inspections is less than 70%.

[0004] Poor environmental adaptability: Existing solutions mostly rely on RTK (real-time kinematic positioning) technology and pre-built 3D point cloud models, which are difficult to implement in areas without network coverage or in complex terrain. This is especially true in distribution network line scenarios, where towers are densely distributed and surrounded by many obstacles. Traditional route planning methods require on-site mapping and modeling 1-2 weeks in advance, which seriously restricts inspection efficiency.

[0005] Insufficient intelligence: Mainstream systems use a "front-end acquisition + back-end analysis" model, which cannot process image data in real time during flight. As a result, the flight path cannot be dynamically adjusted when the target is lost. Research data shows that the success rate of existing solutions in continuous target tracking in complex environments is less than 60%.

[0006] Conflict between safety and efficiency: To obtain clear images, drones need to fly close to the tower (usually less than 10 meters), but this operation can easily lead to collision accidents. At the same time, existing equipment often adopts a direct return strategy when battery power warnings are issued, resulting in repeated flights of unfinished routes and a single mission interruption rate of up to 35%.

[0007] Data processing lags: Naming and archiving inspection photos rely on manual processing, which creates the risk of data corruption. Statistics from a provincial power grid company show that duplicate inspections account for over 15% of inspections annually due to data management issues.

[0008] While recent research has attempted to integrate edge computing with drones, practical applications still face technical drawbacks such as excessive equipment load (typically >1kg), high recognition algorithm latency (>500ms), and insufficient dynamic zoom control accuracy (focus error rate >20%), making it difficult to meet the real-time requirements of precise inspections of distribution network towers. Therefore, developing an autonomous drone inspection system with front-end intelligent decision-making, lightweight deployment, and adaptive route planning capabilities has become a pressing technical challenge for the power industry. Summary of the Invention

[0009] The purpose of this invention is to provide an autonomous inspection system and method for drones based on AI recognition.

[0010] The technical solution of the present invention is achieved by the following methods:

[0011] The present invention discloses an autonomous inspection system for unmanned aerial vehicles based on AI recognition, comprising:

[0012] The drone body has a high-power optical zoom lens;

[0013] Edge computing equipment, mounted on the drone itself;

[0014] Task management module, including line management, single operation task generation and issuance, and inspection data management functions;

[0015] Tower recognition module, image recognition algorithm based on Yolo structure;

[0016] Dynamic route planning module, which supports the generation of standard routes based on a single tower reference point, including three flight mechanisms: direct flight, route-based flight, and increased altitude flight;

[0017] Breakpoint-resume flight control module, real-time power prediction and implementation of safe return strategy;

[0018] The landing control module supports precise positioning and landing on the spot.

[0019] As a further limitation of the present invention, the task management module includes:

[0020] Route construction unit supports one-click import of tower coordinates and generation of preset routes;

[0021] Task issuing unit, supporting inspection task parameter configuration and edge device communication;

[0022] Data synchronization unit, which realizes real-time synchronization of flight data and front-end software;

[0023] Intelligent renaming unit that automatically generates standardized photo naming rules based on flight records.

[0024] As a further limitation of the present invention, the tower identification module includes:

[0025] Feature reuse unit, building a multi-level feature fusion network based on the DenseNet principle;

[0026] Dynamic zoom unit, automatically adjusts optical zoom based on the target's 75% frame rate;

[0027] Real-time tracking unit, using a self-adversarial training model to maintain continuous target lock;

[0028] The focus optimization unit combines the Mosaic data enhancement strategy to improve image clarity.

[0029] As a further limitation of the present invention, the dynamic route planning module supports:

[0030] Three preset route types: single point, straight line, and cross route;

[0031] Custom route configuration interface;

[0032] Obstacle crossing strategy selection unit, providing three preset obstacle crossing modes;

[0033] Real-time coordinate correction unit, visual positioning compensation algorithm based on edge computing.

[0034] As a further limitation of the present invention, the breakpoint-resume flight control module includes:

[0035] Power prediction model, based on regression analysis algorithm of historical flight data;

[0036] The breakpoint recording unit uses time-space coding to store the task interruption location;

[0037] The safe return path planning unit generates the optimal return path based on the route topology.

[0038] As a further limitation of the present invention, the landing control module includes:

[0039] Landing point coordinate recognition, based on the improved real-time detection algorithm of YOLOv9;

[0040] 3D positioning unit, a hybrid positioning system integrating visual SLAM and GPS data;

[0041] The anti-interference unit uses multi-spectral fusion technology to improve recognition accuracy in complex environments.

[0042] As a further limitation of the present invention, the edge computing device includes:

[0043] Model optimization module, which uses semi-supervised learning algorithm to achieve online iterative update of neural network model;

[0044] Real-time inference module, which performs tower recognition and target tracking based on a lightweight convolutional neural network architecture;

[0045] A multi-sensor fusion module configured to process and fuse visual data from the drone's IMU, GPS module, and zoom lens;

[0046] The encrypted communication module uses a symmetric encryption algorithm to ensure the security of data transmission between the drone and the ground control terminal.

[0047] As a further limitation of the present invention, the edge computing device satisfies the following physical parameters:

[0048] The device size is ≤102mm×56.6mm×55mm and the weight is less than 500g;

[0049] Supports mobile cellular network communication and local data caching.

[0050] As a further limitation of the present invention, it also includes:

[0051] Grid inspection management unit supports collaborative operation planning of multiple drones;

[0052] Intelligent diagnosis unit, automatically generates defect analysis reports based on captured photos;

[0053] Emergency avoidance unit, an active obstacle avoidance system with integrated millimeter-wave radar.

[0054] The present invention discloses an autonomous inspection method for a drone, which is applied to a drone autonomous inspection system and includes the following steps:

[0055] S1. Automatically generate standard inspection routes based on single tower coordinates;

[0056] S2. Use visual servo control to achieve continuous tracking of the tower;

[0057] S3. Dynamic adjustment of flight altitude and coordinated control of zoom parameters;

[0058] S4. Adaptive task scheduling strategy based on remaining power;

[0059] S5. After completing the inspection, perform the precision landing verification procedure.

[0060] Beneficial effects of the drone autonomous inspection system and method based on AI recognition of the present invention:

[0061] 1. The entire process is intelligent, reducing manual reliance. Based on a lightweight convolutional neural network (MobileNet's Depthwise Conv structure) and a self-adversarial training model, it achieves real-time recognition and tracking of tower targets (recognition accuracy ≥ 95%, latency < 200ms), completely eliminating reliance on professional pilots. Automated photo naming and data synchronization functions increase the standardization rate of inspection data from the traditional 70% to over 98%, reducing errors caused by human intervention.

[0062] 2. A lightweight edge computing architecture integrates a semi-supervised learning engine and a multi-sensor fusion module in an edge device (measuring 102mm×56.6mm×55mm and weighing less than 500g). This device triples the computing energy efficiency and supports real-time image processing (30fps) for four hours of continuous operation. The model is optimized using the Mish / Relu6 activation function, increasing algorithm inference speed by 40-50% while maintaining 95% recognition accuracy.

[0063] 3. Safety and efficiency are collaboratively optimized. The breakpoint-resume flight control module reduces the mission interruption rate from 35% to below 8% through a power prediction model (error <5%) and topological path planning, shortening the flight time of unfinished routes by 70%. The dynamic zoom unit (with a 75% target lock mechanism) combined with millimeter-wave radar obstacle avoidance achieves "pin-level" clarity at a safe distance of 20-50 meters, reducing the collision accident rate to 0.1 times per thousand flights.

[0064] This invention provides a highly reliable, low-cost autonomous solution for the field of power inspection, and promotes the transformation and upgrading of the industry to an intelligent operation and maintenance model. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0066] Figure 1 This is a structural block diagram of an autonomous inspection system for drones based on AI recognition according to the present invention;

[0067] Figure 2 This is a functional block diagram of an edge computing device for an autonomous inspection system of a drone based on AI recognition according to the present invention;

[0068] Figure 3 This is a functional block diagram of a task management module of an autonomous inspection system for drones based on AI recognition according to the present invention;

[0069] Figure 4 This is a functional block diagram of a tower identification module of an autonomous inspection system of a drone based on AI identification according to the present invention;

[0070] Figure 5 This is a functional block diagram of a breakpoint-resume flight control module of an autonomous inspection system for unmanned aerial vehicles based on AI recognition according to the present invention;

[0071] Figure 6 This is a functional block diagram of a landing control module of an autonomous inspection system for unmanned aerial vehicles based on AI recognition according to the present invention;

[0072] Figure 7 This is a flow chart of an autonomous inspection method of a UAV based on AI recognition in the present invention. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 and Figure 2 As shown, the present invention provides an autonomous drone inspection system based on AI recognition, including: a drone body 10, which adopts a DJI M300RTK drone and is equipped with a 20x optical zoom lens. It can perform zoom shooting at a position 20-50 meters above the distribution network tower, ensuring the clarity of inspection photos while also ensuring flight safety.

[0076] The edge computing device 20, mounted on the drone body 10, measures 102mm*56.6mm*55mm and weighs less than 500g. It includes a model optimization module 201, which uses a semi-supervised learning algorithm to implement online iterative updates of the neural network model; a real-time inference module 202, which performs pole identification and target tracking based on a lightweight convolutional neural network architecture; a multi-sensor fusion module 203, configured to process and fuse visual data from the drone's IMU, GPS module, and zoom lens; and an encrypted communication module 204, which uses a symmetric encryption algorithm to ensure secure data transmission between the drone and the ground control terminal.

[0077] Model optimization module 201, whose hardware carrier can adopt the GPU core of NVIDIA Jetson Xavier NX (384 CUDA cores + 48 Tensor Cores). Semi-supervised learning process: Labeled dataset, pre-loaded with 10,000 labeled tower images (including labels of components such as insulators and crossarms); unlabeled data processing, real-time acquisition of unlabeled images during flight, and expansion of the training set using a pseudo-label generator (based on the prediction results of the pre-trained model with inference confidence greater than 0.85); online iterative update, initiating model fine-tuning every 500 new data, using an exponential moving average (EMA) update strategy (decay coefficient β = 0.999); optimization mechanism, setting the trigger condition for model update (forced retraining when the average recognition confidence of 50 consecutive frames is less than 0.7).

[0078] The real-time inference module 202 features a backbone network architecture that improves the MobileNetv3-Small structure (replacing ReLU activations with mixed Mish / Relu6 activation layers). Feature reuse mechanisms include the introduction of DenseNet's dense connections (with a growth rate of k = 32) in the Conv3_x layer. Deployment optimizations include model quantization using TensorRT 8.0 (FP16 precision), increasing inference speed to 45 fps (at 640×640 input resolution). Memory management includes a dedicated inference buffer (256 MB) for zero-copy data transfer.

[0079] The multi-sensor fusion module 203 has a hardware interface that receives IMU data (100Hz) via the SPI bus, visual data (30fps) via USB3.0, and GPS data (1Hz) via UART. Its timestamp alignment uses PTP (Precision Time Protocol) to achieve microsecond-level time synchronization. The fusion algorithm includes: primary fusion performs Kalman filtering on the IMU angular velocity (ω_x, ω_y, ω_z) and visual optical flow data, and outputs attitude corrections; advanced fusion fuses the GPS coordinates (WGS84), visual SLAM pose (local coordinate system), and barometer height into the drone pose in the ENU coordinate system (frequency 10Hz). Exception handling: When the GPS is out of lock for more than 5 seconds, it switches to pure visual inertial odometry (VIO) mode.

[0080] Encrypted communication module 204: Key generation: A dynamic session key (valid for 24 hours) is generated based on the device's unique ID (SN) and real-time clock (RTC) at each power-up. Data transmission is encrypted using SM4-CTR mode, with a 128-bit block length and a counter modulo 2^64 period. Integrity verification: An HMAC-SHA256 signature is appended (key derivation function PBKDF2, 1000 iterations).

[0081] like Figure 3 As shown, the task management module 30 has the functions of route management, single-operation task generation and issuance, and inspection data management. Specifically, the task management module 30 includes: a route construction unit 301, which supports one-click import of tower coordinates and preset route generation; a task issuance unit 302, which supports inspection task parameter configuration and edge device communication; a data synchronization unit 303, which enables real-time synchronization of flight data with front-end software; and an intelligent renaming unit 304, which automatically generates standardized photo naming rules based on flight records.

[0082] The route construction unit 301, whose hardware is deployed on the ground control terminal, features an Intel Core i7-1185G7 processor equipped with 16GB of DDR4 memory. Coordinate import supports batch import of tower coordinates in KML / KMZ, Shapefile, and CSV formats. The data parsing engine uses the GDAL library for geographic coordinate conversion. Route generation uses the Delaunay triangulation algorithm to construct tower topology relationships and automatically fill in missing coordinate points. A library of preset route templates includes: Single Point Mode, which generates a three-layer concentric circle route centered on the tower; Line Mode, which generates evenly spaced waypoints along the route; and Cross Mode, which generates east-west and south-north cross-scan paths.

[0083] Task dispatch unit 302 uses the MQTT protocol to communicate with edge computing devices; data encapsulation format: a custom binary protocol. Parameter configuration includes: flight parameters (altitude range 20-50m); speed setting: 3-15m / s. Shooting parameters include: zoom ratio: 8x-20x (automatically matched to altitude parameters), shooting interval adjustable from 0.5-5 seconds. The safety policy is geo-fencing: set no-fly zone coordinates, wind speed threshold > 10m / s, and emergency landing when triggered.

[0084] The data synchronization unit 303 includes the following synchronization mechanisms: incremental synchronization uses the rsync algorithm to compare data differences between edge devices and the cloud; breakpoint resuming uses the offset of transferred files to record; priority management uses the priority transmission of urgent data (such as fault photos) (occupying ≥ 50% of the bandwidth). Regular data uses time-slice round-robin scheduling.

[0085] The intelligent renaming unit 304 includes a naming rule engine and format settings.

[0086] Route creation: The operator drags a KML file into the interface, and the system automatically parses and generates a route consisting of 35 towers. The operator selects the "Cross Route" template, setting the flight altitude to 30m and the speed to 8m / s. Task dispatching: The task configuration file (XML format, approximately 50KB) is generated and encrypted and transmitted to the edge device via the 4G network (transmission takes less than 2 seconds). Data synchronization: In-flight photos are uploaded to a cloud storage bucket (AWS S3 protocol) in real time, and metadata is synchronized to the local database. Intelligent naming: 435 photos are renamed based on the GPS time in the flight log (PPS synchronization), and azimuth annotation errors in 12 photos are detected and corrected.

[0087] like Figure 4 As shown, the tower recognition module 40 is based on the image recognition algorithm of the Yolo structure.

[0088] The tower recognition module 40 includes: a feature reuse unit 401, which builds a multi-level feature fusion network based on the DenseNet principle; a dynamic zoom unit 402, which automatically adjusts the optical zoom according to the target's proportion of 75% in the picture; a real-time tracking unit 403, which uses a self-adversarial training model to keep the target continuously locked; and a focus optimization unit 404, which combines the Mosaic data enhancement strategy to improve image clarity.

[0089] Feature reuse unit 401, the network architecture includes: the backbone network is improved based on DenseNet-121, and a cross-stage feature reuse mechanism is introduced in the transition layer. Dense connection method: the input of the nth layer is the channel cascade of the output of the previous n-1 layers (growth rate k = 32). Feature reuse strategy: the output of the Conv2_x layer is weightedly fused with the output of the Conv4_x layer (weight coefficient α = 0.7). Multi-scale fusion: using the FPN (feature pyramid network) structure, the feature maps of the Conv3 (28×28), Conv4 (14×14), and Conv5 (7×7) layers are upsampled and fused. The output feature map size is 112×112×256, which contains multi-scale information of the tower components.

[0090] Dynamic zoom unit 402 calculates target ratio. Using the YOLOv9 detection box, the pixel area S of the tower region is obtained and the target ratio is calculated as: \(R = \frac{S_{box}}{S_{frame}}\times 100\%\) (\(S_{frame}\) is the total number of image pixels). The zoom adjustment strategy triggers zoom in when R < 70% and zoom out when R > 80%. The anti-shake mechanism combines IMU angular velocity data (ω_x, ω_y) to compensate for lens shake or uses Kalman filtering to predict the target's motion trajectory.

[0091] The real-time tracking unit 403 includes: adversarial sample generation, adding FGSM perturbation (ε=0.03) to the input image to generate adversarial samples; model online update, starting adversarial training every time the target loss is detected to be more than 10 frames (about 300ms).

[0092] The focus optimization unit 404 includes: an offline training phase, randomly selecting 4 training images for stitching, adding random rotation (-45° to +45°) and scaling (0.5 to 1.5 times), label processing, and adjusting the target frame coordinates to the new coordinate system after stitching. In the online inference phase, local mosaic processing is performed on the input video stream, and the robustness of the model to blurred images is improved through adversarial training. Clarity optimization, the focus evaluation function uses the Brenner gradient algorithm to calculate the image clarity score. Closed-loop control, when F < threshold (empirical value 1.5×10^6), autofocus is triggered, and the focus step strategy uses the golden section search method.

[0093] The Dynamic Route Planning Module 50 supports the generation of standard routes based on a single tower reference point, including direct flight, route-based flight, and increased altitude flight. It also includes three preset route types: single point, straight line, and cross-route. It also features a custom route configuration interface; an obstacle crossing strategy selection unit with three preset obstacle crossing modes; and a real-time coordinate correction unit using a visual positioning compensation algorithm based on edge computing.

[0094] like Figure 5 As shown, the breakpoint resume control module 60 predicts battery life in real time and implements a safe return strategy. It includes: a battery prediction model 601, based on a regression analysis algorithm for historical flight data; a breakpoint recording unit 602, which uses spatiotemporal encoding to store mission interruption locations; and a safe return path planning unit 603, which generates an optimal return path based on the route topology. Breakpoint resume capability: It predicts the flight time corresponding to the remaining battery life of the drone in real time. If the battery level is insufficient for a safe return, the mission is interrupted and safely returned along the route. The interruption point is recorded for use in breakpoint resume.

[0095] The battery prediction model 601 includes data modeling: historical flight data includes battery voltage curves, motor power consumption (P=IV), and ambient temperature (-20°C to +50°C). Real-time parameters are: current remaining battery power (%), flight speed (m / s), and wind speed vector (direction + magnitude). The regression model uses an XGBoost gradient boosting decision tree, with feature importance ranked as follows: battery internal resistance (40%), flight speed (25%), ambient temperature (20%), and wind speed (15%). The predicted output is the remaining flight time. Error control is achieved through cross-validation (5-fold) to ensure that the prediction error is ≤3%.

[0096] The breakpoint recording unit 602 includes: a data structure and a storage mechanism. Among them, the storage mechanism includes: local storage: FRAM non-volatile memory; cloud synchronization: incremental backup to the AWS DynamoDB database every 5 minutes.

[0097] The safe return path planning unit 603 includes: topological modeling, abstracting the inspection line as a directed graph \(G=(V, E)\), where the nodes are poles and towers, and the edge weights are flight distances (meters), and introducing virtual nodes to represent charging stations / landing points. Energy consumption calculation and decision-making strategies.

[0098] Specifically, in the real-time monitoring stage, the power prediction result is updated every 30 seconds, and the remaining time is displayed through the HMI interface (e.g., remaining 12 minutes 35 seconds ± 5%). When \(T_{remain}<T_{task}\times1.1\) is detected, an early warning is triggered (audible and visual alarm + vibration reminder). In the interruption handling stage, the breakpoint information is recorded in the FRAM memory (writing time < 10 ms), the return path planning is started, and the Dijkstra algorithm is used to calculate the optimal path, preferentially selecting the straight-line return mode and secondly the along-line return mode. In the return execution stage, the UAV is controlled to climb to a safe altitude (50 meters) and switch to the "increase altitude flight" mode, with real-time obstacle avoidance: integrating millimeter-wave radar (detection distance 30 meters) and visual data (YOLOv9 small target detection).

[0099] As Figure 6 shown, the landing control module 70 supports precise landing at the landing point. It includes: a landing recognition unit 701, a real-time detection algorithm improved based on YOLOv9; a three-dimensional positioning unit 702, a hybrid positioning system that integrates visual SLAM and GPS data; and an anti-interference unit 703, which uses multi-spectral fusion technology to improve the recognition accuracy in complex environments.

[0100] The landing point recognition unit 701 improves the YOLOv9 algorithm. Backbone network: Replace CSPDarknet53 with lightweight MobileNetv3 (reducing the number of parameters by 68%); Attention mechanism: Introduce the CBAM module (channel + spatial attention) in the Neck layer; Prior box optimization: Use K-means++ clustering to regenerate the landing point anchor sizes (3 groups: 0.8×0.8m, 1.0×1.0m, 1.2×1.2m).

[0101] 3D positioning unit 702, visual SLAM subsystem, improved ORB-SLAM3 framework, feature point extraction: extracts 1000 ORB features (FAST corner points + BRIEF descriptors) per frame, local mapping: constructs a sparse point cloud map (point spacing 0.1m), relocalization mechanism: based on the DBoW2 bag-of-words model (100,000 words vocabulary). GPS / INS fusion uses the extended Kalman filter (EKF).

[0102] The anti-interference unit 703 adopts multi-spectral fusion technology, with visible light band: 400-700nm (1920×1080 pixels, 30FPS) and infrared band: 8-14μm (640×512 pixels, 25FPS).

[0103] Specifically, during the coarse positioning phase (50m altitude), the wide field of view mode (8mm focal length) is activated to detect the landing area (taking less than 200ms) and generate the initial landing path (cubic spline curve fitting). During the fine alignment phase (10m altitude), the system switches to the narrow field of view mode (90mm focal length), identifies the cross mark at the center of the landing point, and adjusts the drone's posture through the PID controller. During the landing verification phase, touchdown detection is performed. When the pressure sensor threshold is >5kg (all four legs are triggered for success), visual verification is performed: the overlap between the landing point edge and the drone landing gear projection is ≥95%. Safety lock, motor stop delay: 500ms (to prevent bouncing), and a landing success signal is uploaded.

[0104] Example 2

[0105] The present invention provides an autonomous inspection method for drones based on AI recognition, comprising the following steps:

[0106] S1. Automatically generate standard inspection routes based on single tower coordinates.

[0107] Enter the coordinates of a single tower (longitude 113.5°, latitude 23.1°), and the system will execute. It will call a preset cross-shaped route template and generate four waypoints: the base point is 20 meters directly above the tower (pitch angle -90°, zoom 8x), and the extension points are offset by 5 meters in the east, south, west, and north directions (pitch angle -60°, zoom 12x). The AI algorithm will plan the flight path between the waypoints, with a total distance of 35 meters and an estimated flight time of 2 minutes and 15 seconds.

[0108] S2. Use visual servo control to achieve continuous tracking of the tower.

[0109] A feature point matching strategy is adopted to perform SIFT feature extraction on the tower insulators (≥50 feature points are extracted per frame), and the optical flow method (Lucas-Kanade algorithm) is used to track feature points between frames. The drone is controlled to keep the target centered (offset tolerance ±50 pixels). When the target is lost for five consecutive frames, the local search mode is activated: a spiral scan with a radius of 10 meters is performed with the last known position as the center.

[0110] S3. Dynamically adjust the flight altitude and coordinate the control of zoom parameters.

[0111] The flight altitude and zoom parameters are adjusted in tandem to establish an altitude-zoom mapping table:

[0112] Height (m) Zoom ratio Target proportion 20 8X 78％ 30 12X 75％ 50 20X 72％

[0113] The baseline zoom value is obtained by looking up the table based on the real-time altitude (barometer data), and fine-tuned (±2x) based on the image recognition results.

[0114] S4. Adaptive task scheduling strategy based on remaining power.

[0115] The power management strategy uses an LSTM prediction model built on historical data. Input parameters include: current battery level (%), flight speed (m / s), and ambient wind speed (m / s). When the predicted remaining flight time is less than the remaining mission time × 1.2, the mission compression mechanism is triggered, downgrading the cross-route flight mode to single-point shooting mode and disabling non-essential sensors (such as the infrared module).

[0116] S5. After completing the inspection, perform the precision landing verification procedure.

[0117] During the coarse positioning phase, YOLOv9 was used to identify the landing point (with a confidence threshold of 0.9); the pixel coordinates of the landing point's center were calculated (with an accuracy of ±15 pixels). During the fine alignment phase, the RGB-D camera was switched to generate a 3D point cloud (with an accuracy of ±2 cm). The yaw angle of the drone was adjusted to align the landing gear projection with the center of the landing point. For landing verification, after the pressure sensor detected the touchdown signal, a "landing successful" status code (0xAA) was transmitted to the control terminal. If no confirmation was received within 10 seconds, the system automatically switched to the backup communication channel for retransmission.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An autonomous inspection system for drones based on AI recognition, characterized by: include: The drone body has a high-power optical zoom lens; Edge computing equipment, mounted on the drone itself; Task management module, including line management, single operation task generation and issuance, and inspection data management functions; Tower recognition module, image recognition algorithm based on Yolo structure; Dynamic route planning module, which supports the generation of standard routes based on a single tower reference point, including three flight mechanisms: direct flight, route-based flight, and increased altitude flight; Breakpoint-resume flight control module, real-time power prediction and implementation of safe return strategy; The landing control module supports precise positioning and landing on the spot.

2. The system according to claim 1, wherein: The task management module includes: Route construction unit supports one-click import of tower coordinates and generation of preset routes; Task issuing unit, supporting inspection task parameter configuration and edge device communication; Data synchronization unit, which realizes real-time synchronization of flight data and front-end software; Intelligent renaming unit that automatically generates standardized photo naming rules based on flight records.

3. The system according to claim 1, wherein: The tower identification module includes: Feature reuse unit, building a multi-level feature fusion network based on the DenseNet principle; Dynamic zoom unit, automatically adjusts optical zoom based on the target's 75% frame rate; Real-time tracking unit, using a self-adversarial training model to maintain continuous target lock; The focus optimization unit combines the Mosaic data enhancement strategy to improve image clarity.

4. The system according to claim 1, wherein: The dynamic route planning module supports: Three preset route types: single point, straight line, and cross route; Custom route configuration interface; Obstacle crossing strategy selection unit, providing three preset obstacle crossing modes; Real-time coordinate correction unit, visual positioning compensation algorithm based on edge computing.

5. The system according to claim 1, wherein: The breakpoint-resume flight control module includes: Power prediction model, based on regression analysis algorithm of historical flight data; The breakpoint recording unit uses time-space coding to store the task interruption location; The safe return path planning unit generates the optimal return path based on the route topology.

6. The system according to claim 1, wherein: The landing control module includes: Landing point coordinate recognition, based on the improved real-time detection algorithm of YOLOv9; 3D positioning unit, a hybrid positioning system integrating visual SLAM and GPS data; The anti-interference unit uses multi-spectral fusion technology to improve recognition accuracy in complex environments.

7. The system according to claim 1, wherein: The edge computing device includes: Model optimization module, which uses semi-supervised learning algorithm to achieve online iterative update of neural network model; Real-time inference module, which performs tower recognition and target tracking based on a lightweight convolutional neural network architecture; A multi-sensor fusion module configured to process and fuse visual data from the drone's IMU, GPS module, and zoom lens; The encrypted communication module uses a symmetric encryption algorithm to ensure the security of data transmission between the drone and the ground control terminal.

8. The system according to claim 7, characterized in that The edge computing device meets the following physical parameters: The device size is ≤102mm×56.6mm×55mm and the weight is less than 500g; Supports mobile cellular network communication and local data caching.

9. The system according to claim 1, wherein: Also includes: Grid inspection management unit supports collaborative operation planning of multiple drones; Intelligent diagnosis unit, automatically generates defect analysis reports based on captured photos; Emergency avoidance unit, an active obstacle avoidance system with integrated millimeter-wave radar.

10. A drone autonomous inspection method, applied to the drone autonomous inspection system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Automatically generate standard inspection routes based on single tower coordinates; S2. Use visual servo control to achieve continuous tracking of the tower; S3. Dynamic adjustment of flight altitude and coordinated control of zoom parameters; S4. Adaptive task scheduling strategy based on remaining power; S5. After completing the inspection, perform the precision landing verification procedure.

Images