An intelligent inspection unmanned aerial vehicle device based on AI visual recognition

Abstract

This invention discloses an intelligent inspection drone device based on AI visual recognition, comprising: a visual image acquisition module, an edge visual lightweight inference module, a spatial visual feature alignment module, a multi-source data fusion processing module, a flight attitude collaborative control module, and an inspection result transmission module. The device captures images of the inspection area through the visual acquisition module, extracts target features through the edge lightweight inference module, calibrates features using adaptive coordinate mapping and cross-scale matching technology in the spatial visual feature alignment module, integrates visual and flight attitude data to generate a unified feature vector in the multi-source data fusion processing module, adjusts the drone's attitude parameters accordingly in the flight attitude collaborative control module, and transmits the processing results through an encrypted link in the inspection result transmission module. This device solves the problems of insufficient feature alignment accuracy and lack of coordination between data processing and transmission, achieving efficient processing, accurate identification, and stable transmission of inspection data.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection drone technology, and in particular to an intelligent inspection drone device based on AI visual recognition. Background Technology

[0002] In fields such as power line inspection, oil and gas pipeline monitoring, and forestry resource exploration, traditional manual inspection methods face prominent problems such as complex operating environments, limited coverage, and low inspection efficiency, making it difficult to meet the needs of large-scale, high-precision, and real-time inspections. With the deep integration of artificial intelligence, drone technology, and communication technology, intelligent inspection drones have become an important means of replacing traditional manual inspections. Equipped with visual acquisition devices and intelligent processing systems, they can achieve autonomous exploration, data collection, and target identification of the inspection area. Currently, the industry is placing higher demands on the environmental adaptability, data processing speed, and recognition accuracy of inspection drones. The application of edge computing and visual feature alignment technologies has become key to improving drone inspection performance, driving the transformation of intelligent inspection from manual assistance to autonomous intelligence, and helping various industries reduce inspection costs, mitigate safety risks, and improve operational efficiency.

[0003] Existing technologies for intelligent inspection drones suffer from two significant drawbacks: First, insufficient visual feature processing and alignment accuracy. Images captured from different angles and distances in complex environments are prone to feature misalignment and scale inconsistencies, leading to large feature matching deviations during target recognition and affecting the reliability of subsequent data processing. Furthermore, traditional feature processing methods are ill-suited to the dynamic attitude changes during drone flight, making it impossible to quickly achieve spatial calibration of multi-source visual data. Second, a lack of coordination between data processing and transmission. Most devices do not fully integrate edge lightweight inference and efficient communication technologies. This results in either delayed real-time response due to reliance on cloud-based data processing or limitations in local computing power to extract complex features. Additionally, data transmission is susceptible to distance and interference, hindering efficient and stable transmission of processing results and restricting the continuity and accuracy of the inspection process. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides an intelligent inspection drone device based on AI visual recognition.

[0005] The technical solution adopted in this invention is an intelligent inspection drone device based on AI visual recognition, comprising: a visual image acquisition module, an edge visual lightweight inference module, an airspace visual feature alignment module, a multi-source data fusion processing module, a flight attitude collaborative control module, and an inspection result transmission module. The visual image acquisition module captures continuous frame images of the inspection area using a high-definition CMOS sensor and a multispectral imaging component, and transmits the image data stream to the edge vision lightweight inference module in real time. The edge vision lightweight inference module uses a neural network structure optimized with a deep spatially separable convolution and attention mechanism to extract features and perform preliminary target recognition on the image data. Its output feature data is fed into the spatial vision feature alignment module. The spatial vision feature alignment module performs spatial calibration and alignment processing on visual features at different angles and distances through adaptive pixel coordinate mapping and cross-scale feature matching algorithms. The aligned feature data, along with the angular velocity and altitude data collected by the flight attitude sensor, is input into the multi-source data fusion processing module. The multi-source data fusion processing module uses a weighted feature-level fusion strategy to associate and integrate heterogeneous data, generate a unified inspection feature vector, and transmit it to the flight attitude collaborative control module. The flight attitude collaborative control module adjusts the pitch angle, yaw angle, and flight speed parameters of the UAV based on this feature vector. The inspection result transmission module establishes an encrypted data link with the edge computing gateway through the 5G millimeter-wave communication module, and transmits the processed inspection feature vector and target recognition results to the ground control terminal in real time.

[0006] Furthermore, the spatial visual feature alignment module includes: a feature scale normalization unit, a spatial coordinate mapping unit, a cross-frame feature matching unit, and an alignment error correction unit. The feature scale normalization unit normalizes the input multi-scale visual features through dynamic standard deviation adjustment and feature value interval compression algorithms. The spatial coordinate mapping unit constructs a mapping matrix from three-dimensional space to a two-dimensional image plane based on the real-time position coordinates of the UAV and image shooting parameters, and performs spatial transformation of feature point coordinates. The cross-frame feature matching unit uses the K-nearest neighbor algorithm and feature descriptor similarity calculation to match and associate corresponding feature points in consecutive frame images. The alignment error correction unit corrects the deviation of the matched feature point coordinates through a least squares iterative algorithm and outputs spatially aligned visual feature data.

[0007] Furthermore, the multi-source data fusion processing module includes: a heterogeneous data preprocessing unit, a feature association analysis unit, a weighted fusion calculation unit, and a fusion result optimization unit. The heterogeneous data preprocessing unit preprocesses visual feature data and flight attitude data through data format conversion and outlier removal algorithms. The feature association analysis unit uses mutual information calculation and correlation matrix construction methods to mine the intrinsic correlation between different types of data. The weighted fusion calculation unit performs weighted summation on the associated feature data based on a data reliability weight allocation strategy. The fusion result optimization unit smooths the fused data through a sliding window filtering algorithm and outputs a stable inspection feature vector.

[0008] Furthermore, the flight attitude collaborative control module includes: a control parameter parsing unit, an attitude adjustment command generation unit, a speed closed-loop control unit, and an attitude stability adjustment unit. The control parameter parsing unit extracts target distance and angle deviation control parameters from the inspection feature vector. The attitude adjustment command generation unit constructs a command generation model based on PID control logic and fuzzy control rules, and outputs pitch and yaw angle adjustment commands. The speed closed-loop control unit adjusts the UAV's flight speed by calculating the difference between real-time speed feedback and commands. The attitude stability adjustment unit dynamically corrects the adjustment commands by combining the attitude data collected by the gyroscope, thereby performing dynamic control of the flight attitude.

[0009] Furthermore, the feature extraction process expression of the edge vision lightweight inference module is as follows: ,in, For the extracted visual feature vector, For GELU activation function, For the first Layer depth can separate the weights of the convolution. For depthwise separable convolution operations, For the first Frame input image data, For the first Frame input image data, For the first Layer convolution kernel size, For the first Convolution stride, These are the weighting coefficients for the attention mechanism. For multi-head attention computation function, For the first Feature dimensions of a frame image For Hadamard product operations, For LayerNorm normalization operation, For the first Weighting coefficients of the frame image The number of input image frames, This represents the number of convolutional layers.

[0010] Furthermore, the feature alignment accuracy expression of the spatial visual feature alignment module is as follows: ,in, This is a feature alignment accuracy metric. For the aligned first The three-dimensional coordinates of each feature point For the true 3D coordinates of the feature points, To align the first The three-dimensional coordinates of each feature point The total number of feature points. These are the weighting coefficients for the three-dimensional coordinates. The angular deviation influence coefficient is... This represents the deviation value of the drone's shooting angle. It is an exponential function.

[0011] Furthermore, the fused data output expression of the multi-source data fusion processing module is as follows: ,in, This is the fused inspection feature vector. For the first The fusion weights of visual feature data, Feat( ) is the visual feature extraction function. For the first Visual feature data, The number of visual feature types, For the first The fusion weights of flight attitude data For attitude data transformation function, For the first Flight attitude data, The number of pose data types, For cross-modal interaction coefficients, For cross-modal feature interaction functions, This is the variance calculation function.

[0012] Furthermore, the expression for the attitude adjustment amount of the flight attitude cooperative control module is as follows: in, This is the pitch angle adjustment amount. This is the yaw angle adjustment amount. These are the PID control parameters for the pitch angle. These are the PID control parameters for the yaw angle. This is the pitch angle deviation value. This is the yaw angle deviation value. To control time, For integration variables, For the real-time flight speed of the drone, The maximum flight speed threshold, For real-time flight altitude, For reference flight altitude, For feature correlation correction coefficient, This is a feature correlation correction function. For the angular velocity of the drone, Angle of attack for drone flight.

[0013] Furthermore, the data transmission rate expression of the inspection result transmission module is: ,in, For data transmission rate, For communication bandwidth, For transmission power, For the transmit antenna gain, For receiving antenna gain, For communication link efficiency, The distance between the drone and the ground terminal. For communication carrier wavelength, For noise power spectral density, The total power of the interference signal. For the number of interference sources, For the first The interference coefficient of each interference source. For the first The interference power of each interference source Select a function for the transmission mode. To transmit data volume, Service quality level.

[0014] An intelligent inspection drone device based on AI visual recognition is disclosed. The device operates through the following steps: S1, capturing continuous frame images of the inspection area at a preset sampling frequency using a high-definition CMOS sensor and multispectral imaging component of the visual image acquisition module, while simultaneously acquiring exposure parameters, focal length parameters, and drone position coordinates during image capture; S2, transmitting the acquired image data and auxiliary parameters to a lightweight edge vision inference module, where a neural network structure optimized with depth-space separable convolution and attention mechanisms is used to extract hierarchical features from the image data, filtering out candidate feature regions containing potential targets; S3, inputting the candidate feature region data into a spatial visual feature alignment module, where an adaptive pixel coordinate mapping algorithm combined with drone flight attitude data is used to calibrate the spatial coordinates of feature regions in different frame images, and cross-scale feature matching algorithms are then applied. The system performs precise alignment of features from multiple perspectives; S4, the aligned visual feature data and the angular velocity, altitude, and acceleration data collected by the flight attitude sensor are input into the multi-source data fusion processing module, and a weighted feature-level fusion strategy is used to associate and integrate the heterogeneous data to generate a unified inspection feature vector; S5, the unified inspection feature vector is transmitted to the flight attitude collaborative control module, and the pitch angle, yaw angle, and flight speed of the UAV are adjusted through PID control logic and fuzzy control rules based on the target distance and angle deviation information contained in the feature vector, so that the UAV always maintains the optimal inspection attitude; S6, an encrypted data link is established between the 5G millimeter-wave communication module of the inspection result transmission module and the edge computing gateway, and the processed inspection feature vector and target recognition results are transmitted to the ground control terminal in real time, completing the data acquisition, processing, control, and transmission of the entire inspection process.

[0015] Beneficial Effects: This invention proposes an intelligent inspection drone device based on AI visual recognition. The device's onboard visual image acquisition module accurately captures image data of the inspection area, and, in conjunction with a lightweight edge vision inference module, efficiently completes feature extraction and target recognition, avoiding reliance on cloud computing power and significantly improving the real-time performance of data processing. Simultaneously, the spatial visual feature alignment module, through adaptive spatial calibration and cross-scale feature matching technology, effectively solves the problems of feature misalignment and scale inconsistency in images taken from different angles and distances in complex environments, significantly improving feature matching accuracy and ensuring the reliability of subsequent data processing. This successfully overcomes the shortcomings of traditional technologies in visual feature processing and alignment accuracy. The multi-source data fusion processing module integrates visual feature data and flight attitude data to generate a unified inspection feature vector, providing accurate decision-making basis for the flight attitude collaborative control module and ensuring that the UAV maintains the optimal inspection attitude. The inspection result transmission module achieves stable transmission of processing results through an efficient encrypted communication link, completely solving the problems of insufficient coordination between data processing and transmission, delayed response, and unstable transmission in the existing technology. This significantly improves the continuity and accuracy of the inspection process, while reducing the cost and safety risks of manual inspection. It meets the needs of large-scale, high-precision, and real-time inspection in various industries and promotes the efficient development of intelligent inspection towards autonomous intelligence. Attached Figure Description

[0016] Figure 1 This is a diagram showing the modular composition of the device of the present invention; Figure 2 This is a flowchart of the operation steps of the device of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, an intelligent inspection drone device based on AI visual recognition includes: a visual image acquisition module, an edge vision lightweight inference module, an airspace visual feature alignment module, a multi-source data fusion processing module, a flight attitude collaborative control module, and an inspection result transmission module. The visual image acquisition module captures continuous frame images of the inspection area using a high-definition CMOS sensor and a multispectral imaging component, and transmits the image data stream to the edge vision lightweight inference module in real time. The edge vision lightweight inference module uses a neural network structure optimized with a deep spatially separable convolution and attention mechanism to extract features and perform preliminary target recognition on the image data. Its output feature data is fed into the spatial vision feature alignment module. The spatial vision feature alignment module performs spatial calibration and alignment processing on visual features at different angles and distances through adaptive pixel coordinate mapping and cross-scale feature matching algorithms. The aligned feature data, along with the angular velocity and altitude data collected by the flight attitude sensor, is input into the multi-source data fusion processing module. The multi-source data fusion processing module uses a weighted feature-level fusion strategy to associate and integrate heterogeneous data, generate a unified inspection feature vector, and transmit it to the flight attitude collaborative control module. The flight attitude collaborative control module adjusts the pitch angle, yaw angle, and flight speed parameters of the UAV based on this feature vector. The inspection result transmission module establishes an encrypted data link with the edge computing gateway through the 5G millimeter-wave communication module, and transmits the processed inspection feature vector and target recognition results to the ground control terminal in real time.

[0019] The visual image acquisition module employs a combination of a high-definition CMOS sensor and a multispectral imaging component. The sensor has an effective pixel count of 24 million, a pixel size of 1.4 micrometers, and supports a continuous shooting rate of up to 60 frames per second. The multispectral imaging component covers the visible and near-infrared bands from 450 nanometers to 900 nanometers and is equipped with a zoom lens with a focal length of 16mm to 70mm and an aperture range of F1.8 to F6.3, dynamically adjusting the focal length according to the inspection distance. The module features a three-axis mechanical image stabilization mechanism with a stabilization angle range of ±0.5 degrees, effectively compensating for minor shakes during drone flight and ensuring image clarity. When the module is working, it obtains real-time location information through the GPS positioning module on the drone. Combined with the preset inspection path, it starts shooting at a sampling frequency of once every 0.5 seconds. The sensor sensitivity is automatically adjusted between ISO100 and ISO3200, and the exposure time is dynamically adapted to the ambient light intensity within the range of 1 / 1000 second to 1 / 30 second. The multispectral component simultaneously acquires image data of different bands. All image data is stored in 10-bit RAW format and transmitted to the subsequent processing module through a high-speed serial bus. Its core implementation lies in achieving full-dimensional, high-definition image data capture of the inspection area through high-parameter hardware configuration and adaptive shooting parameter adjustment, providing a high-quality data foundation for subsequent feature extraction and target recognition.

[0020] The lightweight edge vision inference module constructs a neural network structure based on a depth-space separable convolution and attention mechanism. The network has 18 layers: 12 convolutional layers, 3 fully connected layers, and 3 attention layers. The convolutional layers use alternating 3×3 and 5×5 kernel sizes with strides of 1 or 2. The fully connected layers have 2048, 1024, and 512 neurons respectively. The attention layers employ a dual attention mechanism combining channel and spatial attention. The module's edge computing chip uses a 16nm process, achieving a computing power of 8 TOPS while consuming less than 5 watts. It supports INT8 quantization and can compress the neural network model to below 200 megabytes, meeting the limited computing power and power consumption requirements of drones. During implementation, the module receives RAW format image data transmitted from the visual image acquisition module, first performs format conversion and channel separation, and then performs hierarchical feature extraction through a pre-trained lightweight neural network. The first 6 convolutional layers extract low-order texture features, the middle 4 convolutional layers are combined with attention layers to extract mid-order shape features, and the last 2 convolutional layers are combined with fully connected layers to extract high-order semantic features. At the same time, the attention mechanism is used to strengthen the weights of key feature regions. The entire inference process takes less than 50 milliseconds, achieving rapid output of preliminary target recognition results and feature data. Its core value lies in achieving efficient and accurate feature extraction and preliminary recognition with limited hardware resources, providing concise and critical feature data for subsequent feature alignment.

[0021] The spatial visual feature alignment module employs adaptive pixel coordinate mapping and cross-scale feature matching technology. It incorporates a pixel coordinate transformation algorithm and feature matching engine, supporting image data with resolutions ranging from 640×480 to 4096×3072. The feature point detection density is set to 5 to 8 per square millimeter, and the coordinate mapping accuracy is controlled within ±1 pixel. After receiving feature data from the edge vision lightweight inference module, the module first acquires real-time angular velocity and altitude data transmitted from the UAV's flight attitude sensor. The angular velocity measurement range is ±200 degrees per second, and the altitude data accuracy is ±0.1 meters. Combined with the focal length and exposure parameters during image capture, it constructs a mapping model from three-dimensional space to a two-dimensional image plane. An adaptive algorithm adjusts the mapping parameters to adapt to variations in shooting angle (range 0 to 90 degrees) and shooting distance (range 5 to 50 meters). Subsequently, a cross-scale feature matching engine is activated to calculate the similarity of feature points in consecutive frames of images. A matching threshold of 0.85 is set, and spatial coordinates are calibrated for successfully matched feature points. Simultaneously, an iterative optimization algorithm corrects alignment errors, with the number of error correction iterations not exceeding 10, ensuring that the spatial deviation of aligned feature points is less than 2 pixels. The key implementation of this module lies in eliminating feature misalignment and scale differences under different shooting conditions through dynamically adapted coordinate mapping and high-precision feature matching. This provides a spatially consistent feature foundation for multi-source data fusion, ensuring the accuracy of subsequent data processing.

[0022] The multi-source data fusion processing module employs a weighted feature-level fusion strategy, supporting three types of data sources: visual feature data, angular velocity data, and height data. The visual feature data has 512 dimensions, the angular velocity data has a sampling frequency of 100 Hz, and the height data has a sampling frequency of 50 Hz. The module incorporates a data correlation analysis engine and a weighted fusion calculation unit. First, it standardizes the format of the input heterogeneous data, converting physical quantities such as angular velocity and height into numerical ranges of the same scale as the visual feature data. Then, it analyzes the correlation between different data sources using a mutual information calculation method. The correlation calculation window size is set to 20 frames, and weight coefficients are assigned based on the calculation results. The weight coefficients for visual feature data range from 0.6 to 0.8, for angular velocity data from 0.1 to 0.2, and for height data from 0.1 to 0.2. These weight coefficients are dynamically updated every 50 frames. During the fusion computation process, various types of data are first synchronized and aligned in time, with the synchronization error controlled within 10 milliseconds. Then, the data is integrated using a weighted summation method to generate a unified 512-dimensional inspection feature vector. Finally, a sliding window filtering algorithm is used to smooth the fusion result, with a window length of 10 frames to filter out random noise in the data. The core of this module lies in achieving the organic integration of heterogeneous data through precise correlation analysis and dynamic weight allocation, eliminating data redundancy and conflicts, and generating an inspection feature vector that is both complete and reliable, providing comprehensive and accurate decision-making basis for flight attitude control.

[0023] The flight attitude cooperative control module takes a 512-dimensional unified inspection feature vector as input, and incorporates PID control logic and a fuzzy control rule library. It supports control of three parameters: pitch angle, yaw angle, and flight speed. The pitch angle control range is -30 degrees to 30 degrees with a control accuracy of ±0.5 degrees; the yaw angle control range is -180 degrees to 180 degrees with a control accuracy of ±1 degree; and the flight speed control range is 1 m / s to 10 m / s with a control accuracy of ±0.1 m / s. After receiving the feature vector output from the multi-source data fusion processing module, the module extracts key control parameters such as target distance and angle deviation through a parameter parsing unit. The target distance detection range is 5 meters to 50 meters, and the angle deviation detection accuracy is ±0.3 degrees. Subsequently, based on the PID control parameters (proportional coefficient range 0.5 to 2.0, integral coefficient range 0.01 to 0.1, and derivative coefficient range 0.1 to 0.5) and fuzzy control rules, it generates attitude adjustment commands and speed adjustment commands. After the command is generated, it is transmitted to the UAV's power system via pulse width modulation (PWM) signal, with a transmission delay controlled within 20 milliseconds. Simultaneously, it receives real-time attitude data from the gyroscope (sampling frequency 100 Hz) and dynamically corrects the adjustment command. The correction magnitude adaptively adjusts according to the magnitude of the attitude deviation; the larger the deviation, the larger the correction magnitude, with a maximum correction magnitude not exceeding 5 degrees. The key to this module's implementation lies in achieving real-time adjustment of the UAV's flight attitude and speed through precise parameter analysis and closed-loop control. This ensures the UAV is always in the optimal state to meet inspection requirements, guaranteeing the stability and targeted nature of the inspection process.

[0024] The inspection result transmission module adopts a combined architecture of a 5G millimeter-wave communication module and an edge computing gateway. The communication module operates in the 24 GHz to 30 GHz frequency band, supports a maximum communication bandwidth of 10 Gbps, and controls the transmission latency to within 50 milliseconds. The edge computing gateway adopts an industrial-grade design and supports multi-protocol compatibility and encrypted transmission. After receiving attitude adjustment information from the flight attitude collaborative control module and inspection feature vectors and target recognition results output by the multi-source data fusion processing module, the module first compresses the data. The compression algorithm uses adaptive entropy coding, and the compression ratio is set to 3:1 to 5:1 to ensure that the data volume is controlled within a reasonable range. Subsequently, the compressed data is encrypted using the AES-256 encryption algorithm, and the encryption key is dynamically updated every 60 seconds. After encryption, the 5G millimeter-wave communication module establishes a communication link with the ground control terminal. The link establishment time is no more than 3 seconds, and data transmission is performed in time-division duplex mode with an uplink transmission rate of no less than 1 gigabits per second. Simultaneously, a link quality detection unit monitors the transmission error rate in real time, with an error rate threshold set to 10^-6. When the error rate exceeds the threshold, the communication channel is automatically switched, with a channel switching time of no more than 100 milliseconds. The core of this module's implementation lies in achieving real-time and secure transmission of inspection data through a high-speed, encrypted, and interference-resistant communication architecture. This ensures that the ground control terminal can obtain accurate inspection results in a timely manner, providing data support for subsequent decision-making.

[0025] Preferably, the spatial visual feature alignment module includes: a feature scale normalization unit, a spatial coordinate mapping unit, a cross-frame feature matching unit, and an alignment error correction unit. The feature scale normalization unit normalizes the input multi-scale visual features through dynamic standard deviation adjustment and feature value interval compression algorithms. The spatial coordinate mapping unit constructs a mapping matrix from three-dimensional space to a two-dimensional image plane based on the real-time position coordinates of the UAV and image shooting parameters, and performs spatial transformation of feature point coordinates. The cross-frame feature matching unit uses the K-nearest neighbor algorithm and feature descriptor similarity calculation to match and associate corresponding feature points in consecutive frame images. The alignment error correction unit corrects the deviation of the matched feature point coordinates through a least squares iterative algorithm and outputs spatially aligned visual feature data.

[0026] Specifically, the feature scale normalization unit, spatial coordinate mapping unit, cross-frame feature matching unit, and alignment error correction unit of the spatial visual feature alignment module work in sequence and collaboratively. The feature scale normalization unit, for the input multi-scale visual features, uses a dynamic standard deviation adjustment algorithm to control the standard deviation of the feature data between 0.8 and 1.2, and uses feature value interval compression technology to limit the feature value range to between 0 and 1, eliminating numerical differences between features of different scales. The spatial coordinate mapping unit receives the real-time position coordinates (positioning accuracy ±0.5 meters) transmitted from the UAV's GPS module, along with the focal length (16 mm to 70 mm) and exposure time (1 / 1000 sec to 1 / 30 sec) parameters during image capture, and constructs a mapping matrix from three-dimensional space to a two-dimensional image plane. The matrix elements are mapped using the least squares method. Iterative optimization, with 8 iterations, achieves accurate conversion of feature point 3D coordinates to 2D image coordinates. The cross-frame feature matching unit uses the K-nearest neighbor algorithm (K value set to 3) to perform preliminary matching of feature points in consecutive frames. Valid matching pairs are filtered through feature descriptor similarity calculation (similarity threshold set to 0.85), with matching efficiency controlled within 30 milliseconds per frame. The alignment error correction unit corrects the feature point coordinate deviation after matching using a least squares iterative algorithm, with each iteration step size set to 0.01. The iteration terminates when the deviation change is less than 0.001, ensuring that the spatial deviation of the corrected feature points is less than 2 pixels. The collaborative operation of the four units achieves accurate alignment of multi-scale and multi-view visual features, providing high-quality feature data for subsequent data fusion.

[0027] Preferably, the multi-source data fusion processing module includes: a heterogeneous data preprocessing unit, a feature association analysis unit, a weighted fusion calculation unit, and a fusion result optimization unit. The heterogeneous data preprocessing unit preprocesses visual feature data and flight attitude data through data format conversion and outlier removal algorithms. The feature association analysis unit uses mutual information calculation and correlation matrix construction methods to mine the intrinsic correlation between different types of data. The weighted fusion calculation unit performs weighted summation on the associated feature data based on a data reliability weight allocation strategy. The fusion result optimization unit smooths the fused data through a sliding window filtering algorithm and outputs a stable inspection feature vector.

[0028] Specifically, the heterogeneous data preprocessing unit uses a format conversion protocol to uniformly convert visual feature data (512-dimensional), angular velocity data (sampling frequency 100 Hz), and height data (sampling frequency 50 Hz) into JSON format. Outliers are removed using the 3σ criterion, with an outlier detection window set to 15 frames to ensure the integrity of the input data. The feature association analysis unit uses mutual information calculation to mine the correlation between different types of data. The mutual information calculation window size is set to 20 frames, and the correlation value ranges from 0 to 1. Data with a correlation greater than 0.7 is considered strongly correlated, and a data association index table is established. The weighted fusion calculation unit assigns weight coefficients based on the data correlation. The weight coefficients of visual feature data are dynamically adjusted between 0.6 and 0.8 based on the correlation, while the weight coefficients of angular velocity and height data are each adjusted between 0.1 and 0.2. The weight coefficients are updated every 50 frames. The correlated data are integrated using a weighted summation formula, with the summation precision controlled to four decimal places. The fusion result optimization unit uses a sliding window filtering algorithm with a window length of 10 frames. The filtering coefficients are dynamically adjusted between 0.1 and 0.3 based on the data fluctuation, filtering out random noise and ensuring that the fluctuation amplitude of the fused inspection feature vector is less than 5%. The synergistic effect of the four units achieves efficient integration of heterogeneous data, generating feature vectors that are both complete and reliable.

[0029] Preferably, the flight attitude collaborative control module includes: a control parameter parsing unit, an attitude adjustment command generation unit, a speed closed-loop control unit, and an attitude stability adjustment unit. The control parameter parsing unit extracts target distance and angle deviation control parameters from the inspection feature vector. The attitude adjustment command generation unit constructs a command generation model based on PID control logic and fuzzy control rules, and outputs pitch and yaw angle adjustment commands. The speed closed-loop control unit adjusts the UAV's flight speed by calculating the difference between real-time speed feedback and commands. The attitude stability adjustment unit dynamically corrects the adjustment commands by combining the attitude data collected by the gyroscope, thereby performing dynamic control of the flight attitude.

[0030] Specifically, the attitude coordination control module includes a control parameter parsing unit, an attitude adjustment command generation unit, a speed closed-loop control unit, and an attitude stability adjustment unit. The control parameter parsing unit extracts control parameters such as target distance (detection range 5 meters to 50 meters) and angle deviation (detection accuracy ±0.3 degrees) from the 512-dimensional inspection feature vector. The parameter parsing time is controlled within 10 milliseconds to ensure real-time control. The attitude adjustment command generation unit has built-in PID control logic and a fuzzy control rule library. The proportional coefficient of the PID control is set between 0.5 and 2.0, the integral coefficient is set between 0.01 and 0.1, and the derivative coefficient is set between 0.1 and 0.5. Based on the magnitude of the angle deviation, the corresponding fuzzy control rule is called to generate the attitude adjustment command. The system provides pitch and yaw angle adjustment commands with a resolution of 0.1 degrees. The speed closed-loop control unit calculates and adjusts the flight speed based on the difference between the real-time speed feedback (sampling frequency 100 Hz) and the command. The speed adjustment step size is set to 0.1 meters per second. When the speed deviation is greater than 0.5 meters per second, a fast adjustment mode is activated, with an adjustment response time of less than 20 milliseconds. The attitude stability adjustment unit receives real-time attitude data (sampling frequency 100 Hz) transmitted from the gyroscope and dynamically corrects the adjustment commands. The correction magnitude is positively correlated with the attitude deviation, with a maximum correction magnitude of no more than 5 degrees. The correction frequency is consistent with the command generation frequency. The coordinated operation of these four units achieves precise and stable control of the UAV's flight attitude and speed, ensuring the continuity and targeted nature of the inspection process.

[0031] Preferably, the feature extraction process of the edge vision lightweight inference module is expressed as follows: ,in, For the extracted visual feature vector, For GELU activation function, For the first Layer depth can separate the weights of the convolution. For depthwise separable convolution operations, For the first Frame input image data, For the first Frame input image data, For the first Layer convolution kernel size, For the first Convolution stride, These are the weighting coefficients for the attention mechanism. For multi-head attention computation function, For the first Feature dimensions of a frame image For Hadamard product operations, For LayerNorm normalization operation, For the first Weighting coefficients of the frame image The number of input image frames, This represents the number of convolutional layers.

[0032] Specifically, the feature extraction formula of the lightweight inference module for edge vision is based on the parameter dimensionality reduction principle of depthwise separable convolution and the feature enhancement logic of the attention mechanism. By splitting the standard convolution into depthwise convolution and pointwise convolution, the computational load is reduced while retaining the ability to extract key features. An attention mechanism is introduced to assign weights to the features of the target region, compensating for the loss of feature information during the lightweighting process. When constructing the formula, the activation function is combined to perform non-linear mapping on the output features, and the data distribution is stabilized through normalization to ensure the stability of feature extraction. The parameter values ​​have been verified through multiple sets of experiments. The number of convolutional layers is set to 12, the number of input image frames is controlled between 8 and 16, the weight coefficients of depthwise separable convolution are assigned between 0.5 and 1.5 according to the importance of features, the weight coefficients of the attention mechanism are set between 0.3 and 0.7, the activation function is an adaptive gradient optimization type, and the weight coefficients of each frame are equally distributed during the normalization process. During implementation, after receiving image data, the module sequentially performs multi-layer convolution operations, attention weighting, activation mapping, and normalization processing according to the formula. The feature extraction time for each frame of image is controlled within 50 milliseconds. This formula enables efficient feature extraction under a lightweight model, which not only meets the computing power limitations of UAVs but also ensures the integrity and discriminability of feature data, providing high-quality feature support for target recognition.

[0033] Preferably, the feature alignment accuracy expression of the spatial visual feature alignment module is: ,in, This is a feature alignment accuracy metric. For the aligned first The three-dimensional coordinates of each feature point For the true 3D coordinates of the feature points, To align the first The three-dimensional coordinates of each feature point The total number of feature points. These are the weighting coefficients for the three-dimensional coordinates. The angular deviation influence coefficient is... This represents the deviation value of the drone's shooting angle. It is an exponential function.

[0034] Specifically, the feature alignment accuracy formula of the airspace visual feature alignment module calculates the impact of 3D spatial coordinate mapping error and angle deviation. Based on the principle of spatial geometric transformation, a model for calculating the coordinate deviation before and after feature point alignment is constructed. An exponential decay factor for angle deviation is introduced to quantify the impact of flight attitude changes on alignment accuracy. The ratio of the sum of squared deviations to the original deviation modulus reflects the alignment effect. The formula design takes into account both coordinate deviation and angle interference to ensure comprehensive accuracy evaluation. Parameter values ​​are selected based on the actual needs of UAV inspection scenarios. The total number of feature points is set to 100 to 200 per frame, the 3D coordinate weight coefficient is set to 0.6 to 0.8, and the angle deviation influence coefficient is set to 0.1 to 0.3. The angle deviation value is collected in real time by the attitude sensor, with the range controlled between 0 and 10 degrees. During implementation, the module first obtains the 3D coordinates and shooting angle deviation of the feature points before and after alignment, calculates the alignment accuracy index according to the formula, and triggers secondary alignment optimization when the index is below 0.8. The coordinate mapping parameters are iteratively adjusted to reduce the deviation, ensuring that the alignment accuracy meets the requirements of subsequent data fusion. The construction of this formula enables quantitative evaluation and dynamic optimization of the alignment effect, improving the reliability of feature alignment.

[0035] Preferably, the fused data output expression of the multi-source data fusion processing module is: ,in, This is the fused inspection feature vector. For the first The fusion weights of visual feature data, Feat( ) is the visual feature extraction function. For the first Visual feature data, The number of visual feature types, For the first The fusion weights of flight attitude data For attitude data transformation function, For the first Flight attitude data, The number of pose data types, For cross-modal interaction coefficients, For cross-modal feature interaction functions, This is the variance calculation function.

[0036] Specifically, the multi-source data fusion processing module's output formula is based on the principle of feature complementarity of heterogeneous data. It eliminates fusion conflicts caused by data type differences by performing feature extraction and format conversion on visual feature data and flight attitude data separately. A cross-modal interaction term is introduced to capture potential correlations between different data types. The data fusion weights are dynamically adjusted using variance ratios to balance the contributions of various data types. A weighted summation model is used in formula construction to integrate multi-source data, ensuring the completeness and representativeness of the fusion results. Parameter values ​​have undergone extensive experimental calibration. The number of visual feature types is set to 8 to 12, the number of flight attitude data types to 3 to 5, the fusion weights for visual feature data are allocated between 0.6 and 0.8, the fusion weights for flight attitude data are allocated between 0.1 and 0.2, and the cross-modal interaction coefficients are set between 0.2 and 0.4. During implementation, the module first preprocesses and extracts features from the input heterogeneous data, assigns weights according to the formula, and completes weighted summation, cross-modal interactive calculation, and variance ratio adjustment. The time synchronization error of the fusion process is controlled within 10 milliseconds. Through this formula, the organic integration of multi-source heterogeneous data is realized, generating a unified inspection feature vector, which provides a comprehensive and accurate decision basis for flight attitude control.

[0037] Preferably, the expression for the attitude adjustment amount of the flight attitude cooperative control module is: in, This is the pitch angle adjustment amount. This is the yaw angle adjustment amount. These are the PID control parameters for the pitch angle. These are the PID control parameters for the yaw angle. This is the pitch angle deviation value. This is the yaw angle deviation value. To control time, For integration variables, For the real-time flight speed of the drone, The maximum flight speed threshold, For real-time flight altitude, For reference flight altitude, For feature correlation correction coefficient, This is a feature correlation correction function. For the angular velocity of the drone, Angle of attack for drone flight.

[0038] Specifically, the attitude adjustment formula of the flight attitude cooperative control module is based on the proportional-integral-derivative (PID) control principle. Combined with the UAV flight dynamics model, normalization factors for flight speed and altitude are introduced to balance control sensitivity under different flight conditions. A feature correlation correction term is added to associate visual feature data with attitude parameters, achieving adaptive control based on target features. The formula design considers steady-state error elimination, dynamic response speed, and overshoot suppression to ensure the accuracy and stability of attitude adjustment. Parameter values ​​are optimized according to the UAV model and inspection scenario. The proportional coefficient is set between 0.5 and 2.0, the integral coefficient between 0.01 and 0.1, the derivative coefficient between 0.1 and 0.5, the maximum flight speed threshold is 10 meters per second, the reference flight altitude is set between 5 and 50 meters depending on the inspection task, and the feature correlation correction coefficient is allocated between 0.1 and 0.3. During implementation, the module collects attitude deviation, flight speed, altitude and visual feature data in real time, calculates the pitch and yaw angle adjustments according to the formula, and controls the generation and execution delay of adjustment commands to within 20 milliseconds. Through this formula, closed-loop adaptive control of flight attitude is achieved, ensuring that the UAV maintains the optimal attitude in complex inspection environments and improving the continuity and targeting of inspections.

[0039] Preferably, the data transmission rate expression of the inspection result transmission module is: ,in, For data transmission rate, For communication bandwidth, For transmission power, For the transmit antenna gain, For receiving antenna gain, For communication link efficiency, The distance between the drone and the ground terminal. For communication carrier wavelength, For noise power spectral density, The total power of the interference signal. For the number of interference sources, For the first The interference coefficient of each interference source. For the first The interference power of each interference source Select a function for the transmission mode. To transmit data volume, Service quality level.

[0040] Specifically, the data transmission rate formula for the inspection result transmission module is based on an extension of Shannon's formula for communication links. It incorporates practical influencing factors such as antenna gain, transmission distance, and interference suppression. The impact of distance on signal strength is quantified using a path loss model. A power superposition term for multiple interference sources is added, and the selection of transmission mode is considered in relation to the amount of transmitted data and the quality of service level. The formula integrates channel capacity calculation, interference suppression, and transmission mode adaptation to ensure the accuracy and practicality of the rate calculation. Parameter values ​​conform to 5G millimeter-wave communication standards, with a communication bandwidth of 200 to 400 MHz, transmit power controlled between 10 and 20 dBmW, transmit and receive antenna gains set to 15 to 25 dB, communication link efficiency between 0.7 and 0.9, and a fixed standard value for noise power spectral density. A maximum of eight interference sources are considered, with the interference coefficient of each source ranging from 0.1 to 0.3. During implementation, the module detects the communication distance, interference signal strength, and data transmission volume in real time, calculates the maximum transmission rate of the current channel according to the formula, and dynamically adjusts the transmission mode and encoding method to ensure that the data transmission rate is not less than 1 gigabits per second and the bit error rate is controlled below 10 to the power of -6. The dynamic optimization of the transmission rate is achieved through this formula, ensuring the real-time and stable transmission of inspection results.

[0041] like Figure 2As shown, an intelligent inspection drone device based on AI visual recognition operates through the following steps: S1, capturing continuous frame images of the inspection area at a preset sampling frequency using a high-definition CMOS sensor and multispectral imaging component of the visual image acquisition module, while simultaneously acquiring exposure parameters, focal length parameters, and drone position coordinates during image capture; S2, transmitting the acquired image data and auxiliary parameters to the edge vision lightweight inference module, performing hierarchical feature extraction on the image data using a neural network structure optimized by deep spatial separable convolution and attention mechanisms, and filtering out candidate feature regions containing potential targets; S3, inputting the candidate feature region data into the spatial visual feature alignment module, performing spatial coordinate calibration of feature regions in different frame images using an adaptive pixel coordinate mapping algorithm combined with drone flight attitude data, and utilizing cross-scale feature matching. The algorithm performs precise alignment of features from multiple perspectives; S4, the aligned visual feature data and the angular velocity, altitude, and acceleration data collected by the flight attitude sensor are input into the multi-source data fusion processing module, and a weighted feature-level fusion strategy is used to associate and integrate the heterogeneous data to generate a unified inspection feature vector; S5, the unified inspection feature vector is transmitted to the flight attitude collaborative control module, and the pitch angle, yaw angle, and flight speed of the UAV are adjusted through PID control logic and fuzzy control rules based on the target distance and angle deviation information contained in the feature vector, so that the UAV always maintains the optimal inspection attitude; S6, an encrypted data link is established between the 5G millimeter-wave communication module of the inspection result transmission module and the edge computing gateway, and the processed inspection feature vector and target recognition results are transmitted to the ground control terminal in real time, completing the data acquisition, processing, control, and transmission of the entire inspection process.

[0042] The formula in this invention integrates different scalar and vector parameters for unified calculation, and constructs a collaborative calculation logic through standardization, dimensional adaptation, and physical meaning association. First, addressing the dimensional differences between scalars (such as flight speed, angle deviation, and weighting coefficients) and vectors (such as visual feature vectors, 3D coordinates, and attitude parameters), the formula is constructed by normalizing various parameters to the same numerical range. For example, the dimensional values ​​of visual feature vectors are compressed to a fixed range, forming a computable basis with the scalar weighting coefficients and angle deviation influence coefficients. Second, a cross-modal association mechanism clarifies the physical relationship between different types of parameters. For example, the vector form of visual feature vectors is linked to the scalar flight speed and altitude through a feature correlation correction term, making the spatial feature information of vectors complementary to the motion state parameters of scalars. Third, dimensional matching is achieved through operations such as weighted summation and Hadamard product. For example, multi-dimensional visual feature vectors are weighted element-wise with scalar weighting coefficients, and the 3D coordinate deviation of vectors and the scalar angle deviation influence coefficient are integrated into a unified evaluation index through multiplication. Taking the spatial visual feature alignment accuracy formula as an example, the three-dimensional coordinate deviation of the vector and the angular deviation of the scalar are normalized and weighted, and then fused by the calculation of the sum of squared deviations and the ratio of the magnitude. This retains the spatial position information of the vector and incorporates the attitude interference of the scalar, ultimately forming a calculation result that is both complete and reasonable. This ensures that different types of parameters work together efficiently in the formula and accurately support the implementation of the functions of each module.

[0043] An intelligent inspection drone device based on AI visual recognition achieves a breakthrough improvement in visual feature processing accuracy through the collaboration of different modules. After the visual image acquisition module captures multi-dimensional continuous frame images, the edge vision lightweight inference module efficiently extracts target features with an optimized neural network structure. The spatial visual feature alignment module adopts adaptive pixel coordinate mapping and cross-scale feature matching technology to accurately calibrate image features at different angles and distances, dynamically adapting to changes in the drone's flight attitude. This completely solves the problems of feature misalignment and scale inconsistency, significantly reducing feature matching deviation in target recognition and providing a highly reliable foundation for subsequent data processing. This significantly outperforms the adaptation capability and calibration efficiency of traditional feature processing methods.

[0044] This invention achieves closed-loop optimization in data processing and transmission coordination, successfully overcoming the challenges of response lag and transmission instability in existing technologies. The multi-source data fusion processing module integrates visual feature data and flight attitude data, combining edge lightweight inference technology to achieve efficient local processing of complex features, eliminating reliance on cloud computing power and significantly improving real-time response speed. The inspection result transmission module constructs a stable data transmission channel through encrypted communication links, resisting the influence of distance and interference factors to ensure efficient transmission of processing results. Simultaneously, the flight attitude collaborative control module precisely adjusts the UAV's attitude based on the fused feature vectors, ensuring the continuity of the inspection process. The collaborative mechanism formed by these modules not only resolves the contradiction between limited local computing power and complex feature extraction but also achieves efficient connection between data processing and transmission, comprehensively improving the accuracy and stability of inspections.

[0045] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent inspection drone device based on AI visual recognition, characterized in that, include: Visual image acquisition module, edge vision lightweight inference module, airspace visual feature alignment module, multi-source data fusion processing module, flight attitude collaborative control module, and inspection result transmission module; The visual image acquisition module captures continuous frame images of the inspection area using a high-definition CMOS sensor and a multispectral imaging component, and transmits the image data stream to the edge vision lightweight inference module in real time. The edge vision lightweight inference module uses a neural network structure optimized with a deep spatially separable convolution and attention mechanism to extract features and perform preliminary target recognition on the image data. Its output feature data is fed into the spatial vision feature alignment module. The spatial vision feature alignment module performs spatial calibration and alignment processing on visual features at different angles and distances through adaptive pixel coordinate mapping and cross-scale feature matching algorithms. The aligned feature data, along with the angular velocity and altitude data collected by the flight attitude sensor, is input into the multi-source data fusion processing module. The multi-source data fusion processing module uses a weighted feature-level fusion strategy to associate and integrate heterogeneous data, generate a unified inspection feature vector, and transmit it to the flight attitude collaborative control module. The flight attitude collaborative control module adjusts the pitch angle, yaw angle, and flight speed parameters of the UAV based on this feature vector. The inspection result transmission module establishes an encrypted data link with the edge computing gateway through the 5G millimeter-wave communication module, and transmits the processed inspection feature vector and target recognition results to the ground control terminal in real time.

2. The intelligent inspection unmanned aerial vehicle device based on AI visual recognition according to claim 1, characterized in that, The spatial visual feature alignment module includes: a feature scale normalization unit, a spatial coordinate mapping unit, a cross-frame feature matching unit, and an alignment error correction unit. The feature scale normalization unit normalizes the input multi-scale visual features through dynamic standard deviation adjustment and feature value interval compression algorithms. The spatial coordinate mapping unit constructs a mapping matrix from three-dimensional space to a two-dimensional image plane based on the real-time position coordinates of the UAV and image shooting parameters, and performs spatial transformation of feature point coordinates. The cross-frame feature matching unit uses the K-nearest neighbor algorithm and feature descriptor similarity calculation to match and associate corresponding feature points in consecutive frame images. The alignment error correction unit corrects the deviation of the matched feature point coordinates through a least squares iterative algorithm and outputs spatially aligned visual feature data.

3. The intelligent inspection unmanned aerial vehicle device based on AI visual recognition according to claim 1, characterized in that, The multi-source data fusion processing module includes: a heterogeneous data preprocessing unit, a feature association analysis unit, a weighted fusion calculation unit, and a fusion result optimization unit. The heterogeneous data preprocessing unit preprocesses visual feature data and flight attitude data through data format conversion and outlier removal algorithms. The feature association analysis unit uses mutual information calculation and correlation matrix construction methods to mine the intrinsic correlation between different types of data. The weighted fusion calculation unit performs weighted summation on the associated feature data based on a data reliability weight allocation strategy. The fusion result optimization unit smooths the fused data through a sliding window filtering algorithm and outputs a stable inspection feature vector.

4. The intelligent inspection unmanned aerial vehicle device based on AI visual recognition according to claim 1, characterized in that, The flight attitude collaborative control module includes: a control parameter analysis unit, an attitude adjustment command generation unit, a speed closed-loop control unit, and an attitude stability adjustment unit. The control parameter analysis unit extracts target distance and angle deviation control parameters from the inspection feature vector. The attitude adjustment command generation unit constructs a command generation model based on PID control logic and fuzzy control rules, and outputs pitch and yaw angle adjustment commands. The speed closed-loop control unit adjusts the UAV's flight speed by calculating the difference between real-time speed feedback and commands. The attitude stability adjustment unit dynamically corrects the adjustment commands by combining the attitude data collected by the gyroscope, thereby performing dynamic control of the flight attitude. 5.The AI vision-identification-based intelligent inspection unmanned aerial vehicle device according to claim 1, characterized in that, The feature extraction process of the edge vision lightweight inference module is expressed as follows: ,in, For the extracted visual feature vector, For GELU activation function, For the first Layer depth can separate the weights of the convolution. For depthwise separable convolution operations, For the first Frame input image data, For the first Frame input image data, For the first Layer convolution kernel size, For the first Convolution stride, These are the weighting coefficients for the attention mechanism. For multi-head attention computation function, For the first Feature dimensions of a frame image For Hadamard product operations, For LayerNorm normalization operation, For the first Weighting coefficients of the frame image, The number of input image frames, This represents the number of convolutional layers. 6.The AI vision-identification-based intelligent inspection unmanned aerial vehicle device according to claim 1, characterized in that, The feature alignment accuracy expression for the spatial visual feature alignment module is: ,in, This is a feature alignment accuracy metric. For the aligned first The three-dimensional coordinates of each feature point For the true 3D coordinates of the feature points, To align the first The three-dimensional coordinates of each feature point The total number of feature points. These are the weighting coefficients for the three-dimensional coordinates. The angular deviation influence coefficient is... This represents the deviation value of the drone's shooting angle. It is an exponential function.

7. The intelligent inspection unmanned aerial vehicle device based on AI visual recognition according to claim 1, characterized in that, The fused data output expression of the multi-source data fusion processing module is: ,in, This is the fused inspection feature vector. For the first The fusion weights of visual feature data, Feat( ) is the visual feature extraction function. For the first Visual feature data, The number of visual feature types, For the first The fusion weights of flight attitude data For attitude data transformation function, For the first Flight attitude data, For the number of pose data types, For cross-modal interaction coefficients, For cross-modal feature interaction functions, This is the variance calculation function. 8.The AI vision-identification-based intelligent inspection unmanned aerial vehicle device according to claim 1, wherein The attitude adjustment expression of the flight attitude cooperative control module is as follows: in, This is the pitch angle adjustment amount. This is the yaw angle adjustment amount. These are the PID control parameters for the pitch angle. These are the PID control parameters for the yaw angle. This is the pitch angle deviation value. This is the yaw angle deviation value. To control time, For integration variables, For the real-time flight speed of the drone, The maximum flight speed threshold, For real-time flight altitude, For reference flight altitude, For feature correlation correction coefficient, This is a feature correlation correction function. For the angular velocity of the drone, Angle of attack for drone flight. 9.The AI vision-identification-based intelligent inspection unmanned aerial vehicle device according to claim 1, wherein The data transmission rate expression for the inspection result transmission module is: ,in, For data transmission rate, For communication bandwidth, For transmission power, For the transmit antenna gain, For receiving antenna gain, For communication link efficiency, The distance between the drone and the ground terminal. For communication carrier wavelength, For noise power spectral density, The total power of the interference signal. For the number of interference sources, For the first The interference coefficient of each interference source. For the first The interference power of each interference source Select a function for the transmission mode. To transmit data volume, Service quality level.

10. An intelligent inspection drone device based on AI visual recognition according to any one of claims 1-9, characterized in that, The device operates through the following steps: S1, capturing continuous frame images of the inspection area at a preset sampling frequency using a high-definition CMOS sensor and multispectral imaging components in the visual image acquisition module, while simultaneously acquiring exposure parameters, focal length parameters, and UAV position coordinates during image capture; S2, transmitting the acquired image data and auxiliary parameters to the edge vision lightweight inference module, where a neural network structure optimized with depth-space separable convolution and attention mechanisms is used to extract hierarchical features from the image data, filtering out candidate feature regions containing potential targets; S3, inputting the candidate feature region data into the spatial visual feature alignment module, using an adaptive pixel coordinate mapping algorithm combined with UAV flight attitude data to calibrate the spatial coordinates of feature regions in different frame images, and using a cross-scale feature matching algorithm to accurately match features from multiple perspectives. Alignment; S4, input the aligned visual feature data and the angular velocity, altitude, and acceleration data collected by the flight attitude sensor into the multi-source data fusion processing module, and use a weighted feature-level fusion strategy to associate and integrate the heterogeneous data to generate a unified inspection feature vector; S5, transmit the unified inspection feature vector to the flight attitude collaborative control module, and adjust the pitch angle, yaw angle, and flight speed of the UAV according to the target distance and angle deviation information contained in the feature vector through PID control logic and fuzzy control rules to ensure that the UAV always maintains the optimal inspection attitude; S6, establish an encrypted data link with the edge computing gateway through the 5G millimeter-wave communication module of the inspection result transmission module, and transmit the processed inspection feature vector and target recognition results to the ground control terminal in real time, completing the data acquisition, processing, control, and transmission of the entire inspection process.

Images