Unmanned aerial vehicle-based main road camera non-closed road autonomous cleaning system and method

Abstract

The present application relates to the technical field of unmanned aerial vehicle intelligent maintenance and cleaning, and particularly relates to a main road camera non-closed road autonomous cleaning system and method based on an unmanned aerial vehicle, which performs the following: the unmanned aerial vehicle hovers outside the camera, collects millimeter wave-polarization-acoustic signals to generate a stain spectrum fingerprint, and obtains cleaning parameters after optimization; a model predictive controller sequentially performs plasma, acoustic cavitation, and ion wind curtain dry cleaning under energy constraints, and iteratively adjusts parameters when the residual exceeds a threshold value, and returns to replace the battery after meeting the standard. The cleaning progress, residual index, and energy consumption are packaged and uploaded, a cloud self-learning model is updated online and outputs the next batch of instructions, and autonomous cleaning is realized without stopping the road, with low consumption and high efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance and cleaning technology for unmanned aerial vehicles (UAVs), and in particular to an autonomous cleaning system and method for main road cameras without road closures based on UAVs. Background Technology

[0002] At major road intersections, traffic monitoring cameras are constantly exposed to exhaust fumes, oil films, and dust. Mirror contamination directly weakens image resolution, affecting the reliability of violation evidence collection and signal linkage algorithms. Relying on manual or aerial vehicle operations requires lane closures and poses a risk of falling from heights. Therefore, autonomous cleaning of main road cameras using drones without road closures has become a rigid requirement. Existing technology completes cleaning in a closed loop: drone carrying a high-definition camera takes photos → YOLO-EMA identifies stains → water sprays for targeted rinsing → secondary photo re-inspection (Chinese Invention Patent, Publication No.: CN119216290A). This solution has the following problems: optical imaging is prone to missed or false detections in backlight, nighttime, or specular reflection scenarios; it requires rotating around the camera for framing, resulting in long operation times and significant impact on drying speed and traffic flow; it increases the onboard weight and causes water backflow, making it unsuitable for electrical equipment and extremely cold regions; it lacks negative pressure adsorption or mechanical clamping, leading to positioning drift under crosswind conditions; and it does not monitor energy consumption or adaptively optimize subsequent task parameters. The aforementioned defects limit cleaning efficiency, applicable environments, and long-term operational and maintenance costs. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides a non-closed road autonomous cleaning system and method based on a drone's main road camera. This invention locks the drone to the outside of the camera and uses millimeter-wave, polarization, and acoustic three-channel sensing fusion to generate a stain spectrum fingerprint, which is then optimized using Bayesian methods to output cleaning parameters. A model predictive controller drives a three-stage dry cleaning process—plasma, acoustic cavitation, and ion curtain—under energy constraints, and iterates through a residual index closed loop until the target is met. The drone then returns to the drone for battery swapping and uploads energy consumption-quality data to the blockchain for cloud-based self-learning model optimization, achieving continuous, efficient, and low-consumption non-closed road autonomous cleaning.

[0004] A drone-based autonomous road cleaning system using cameras for non-closure of main roads includes:

[0005] The flight positioning module is used to read flight command data and synchronize positioning and map building data to drive the drone to fly to the target camera;

[0006] The stain detection and decision module is used to collect millimeter-wave reflection coefficient data, polarization scattering matrix data and acoustic echo data, fuse them to generate stain spectrum fingerprint data, and determine cleaning parameters based on the stain spectrum fingerprint data.

[0007] The cleaning and quality inspection module is used to sequentially perform cleaning processes according to the cleaning parameters under the power trajectory defined by the model prediction controller, acquire cleaning progress field data and quality inspection residual index data in real time, and adjust the cleaning parameters to repeat the process when the quality inspection residual index data exceeds the threshold, until the quality inspection residual index data does not exceed the threshold.

[0008] The energy module is used to control the drone to return to base after the quality inspection residual index data does not exceed the threshold, and to encapsulate the cleaning progress field data, quality inspection residual index data and energy settlement data into a closed-loop data package to adjust the self-learning model's own parameters and generate the next flight command data.

[0009] Preferably, the flight positioning module includes an inertial measurement unit, a lidar, and a visual recognition device. The flight positioning module uses a fusion algorithm to jointly calculate the acceleration and angular velocity output by the inertial measurement unit, the point cloud coordinates output by the lidar, and the attitude angle output by the visual recognition device to obtain the three-dimensional position and attitude of the UAV.

[0010] Preferably, the flight positioning module further includes a path planning unit, which generates an obstacle avoidance trajectory based on the three-dimensional position and attitude, and writes the trajectory into the flight command data.

[0011] Preferably, the stain detection and decision module includes a multimodal data fusion unit and a fingerprint optimization unit. The multimodal data fusion unit fuses millimeter-wave reflection coefficient, polarization scattering matrix and acoustic echo into a stain spectrum fingerprint. The fingerprint optimization unit outputs a set of cleaning parameters based on the stain spectrum fingerprint.

[0012] Preferably, the fingerprint optimization unit outputs the set of cleaning parameters with the goal of minimizing the quality inspection residue index using a Bayesian search algorithm.

[0013] Preferably, the cleaning and quality inspection module includes a plasma generator, an acoustic cavitation transducer, and an ion curtain device in sequence. The cleaning and quality inspection module also includes a model prediction controller, which generates a power trajectory within a preset energy constraint and drives the plasma generator, the acoustic cavitation transducer, and the ion curtain device.

[0014] Preferably, the model prediction controller updates the power trajectory based on the cleaning progress field data and the quality inspection residual index data, and sends the updated power trajectory to the plasma generator and the acoustic cavitation transducer.

[0015] Preferably, the energy module includes an automatic battery swapping mechanism and an energy management unit. The automatic battery swapping mechanism completes battery replacement after the UAV returns to base, and the energy management unit records the power consumption before and after the battery swap and outputs energy settlement data.

[0016] Preferably, after generating the operation closed-loop data packet, the energy management unit calculates the hash value of the operation closed-loop data packet and attaches a digital signature, writes it into the ledger through the blockchain node, and then sends it to the cloud through the network communication interface. The self-learning model uses an incremental gradient boosting decision tree to train the received operation closed-loop data packet online, and returns the next flight command data after the training is completed.

[0017] A method for autonomous cleaning of main road cameras without road closure based on unmanned aerial vehicles (UAVs), used to execute the aforementioned autonomous cleaning system for main road cameras without road closure, includes the following steps:

[0018] Read flight command data and synchronized positioning and mapping data to drive the drone to the target camera and establish a fixed attitude;

[0019] Millimeter-wave reflection coefficient data, polarization scattering matrix data, and acoustic echo data are collected, fused to generate stain spectrum fingerprint data, and cleaning parameters are determined based on the stain spectrum fingerprint data.

[0020] According to the cleaning parameters, plasma discharge treatment, acoustic cavitation stripping treatment and ion curtain purging treatment are sequentially performed under the power trajectory defined by the model prediction controller to obtain cleaning progress field data and quality inspection residual index data; when the quality inspection residual index data is greater than the threshold, the cleaning parameters are adjusted and the treatment is repeated until the quality inspection residual index data is less than or equal to the threshold.

[0021] After the quality inspection residual index data is less than or equal to the threshold, the drone is controlled to return to the edge base station, and the cleaning progress field data, quality inspection residual index data and energy settlement data are encapsulated into a closed-loop data package for the self-learning model to adjust parameters and generate the next flight command data.

[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0023] This invention achieves highly robust identification of stain type and thickness through multimodal spectral fingerprint fusion sensing technology, overcoming the problem of missed detection by visual single-modality.

[0024] This invention achieves a dry composite cleaning effect of plasma-acoustic cavitation-ion air curtain under single hovering by using a negative pressure adsorption soft docking ring and model prediction control, avoiding repeated photography, reducing water usage, and removing stubborn oil films.

[0025] This invention achieves transparency in operational energy consumption and incremental self-learning optimization in the cloud through an energy management-blockchain closed-loop approach, continuously reducing the energy and time costs of a single task.

[0026] This invention achieves second-level battery replacement and 24 / 7 uninterrupted operation through edge base station automatic battery swapping technology, solving the battery life bottleneck of long-link inspection. Attached Figure Description

[0027] Figure 1 This is a structural block diagram of the system of the present invention;

[0028] Figure 2 This is a schematic diagram of the overall system architecture of the present invention;

[0029] Figure 3 This is a schematic diagram of the structure of the drone mounting bracket in a specific embodiment of the present invention;

[0030] Figure 4 A schematic diagram of the gimbal load structure in a specific embodiment of the present invention;

[0031] Figure 5 Schematic diagram of the airborne equipment in a specific embodiment of the present invention Figure 1 ;

[0032] Figure 6 Schematic diagram of the airborne equipment in a specific embodiment of the present invention Figure 2 ;

[0033] Figure 7 Schematic diagram of the airborne equipment in a specific embodiment of the present invention Figure 3 ;

[0034] Figure 8 A schematic diagram of the three-axis gimbal mechanism in a specific embodiment of the present invention;

[0035] Reference numerals: 1. UAV mounting bracket; 2. Water pump inlet; 3. Injection port; 4. Liquid storage chamber; 5. Liquid storage tank; 6. Positioning slot; 7. Gimbal load device; 8. Camera mounting base; 9. Camera data cable through hole; 10. Mounting base; 11. Algorithm core board; 12. Water pipe guide hole; 13. Adapter board slot; 14. Locking point; 15. Nozzle; 16. Flight onboard equipment; 17. Flight control core board; 18. Miniature water pump; 19. Water pump inlet; 20. Water pump outlet; 21. Relay control board; 22. EPORT interface adapter board; 23. Electromagnetic three-way valve outlet; 24. Electromagnetic three-way valve; 25. Electromagnetic three-way valve inlet; 26. Connection hole; 27. Three-axis gimbal; 28. Gimbal load mounting base; 29. ​​Gimbal interface. Detailed Implementation

[0036] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0037] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0038] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0039] like Figures 1-2 As shown, a non-closed autonomous cleaning system for main roads based on drone cameras includes:

[0040] The flight positioning module is used to read flight command data and synchronize positioning and map building data to drive the drone to fly to the target camera;

[0041] In this invention, the drone must approach and attach itself without disturbing pedestrians above a busy main road. Therefore, the flight positioning module undertakes two core tasks: first, to provide the drone with continuous and reliable spatial coordinates and attitude angles; and second, to map these coordinates and attitudes in real time to a safe flight path that does not interfere with traffic while meeting the requirements of negative pressure adsorption geometry. This will be elaborated below from four aspects: sensor collaboration, state inference, path planning, and attitude locking.

[0042] In urban road environments, single sensors are often distorted by factors such as obstruction, rain, fog, and reflections from metal structures. This module uses a combination of inertial measurement unit, LiDAR, and machine vision, for the following reasons and division of labor:

[0043] Inertial Measurement Unit: The internal three-axis accelerometer and gyroscope can provide the body's acceleration and angular velocity in milliseconds. Even if the external signal is completely interrupted, it can predict the body's attitude in a short time. It is a "dependency-free" but easy-to-accumulate drift reference.

[0044] LiDAR: Rotating lasers are most sensitive to reflections from pillar targets such as streetlights, billboards, and height restriction barriers, and can still output accurate distance and direction at night when there are no textures and low light.

[0045] Machine vision: In good lighting conditions, the camera can identify reflective marks installed on the edge of the camera and can distinguish heading angle and relative distance at the sub-pixel level, providing scale correction for LiDAR.

[0046] This "triangular" sensing system complements each other under various conditions, such as sunny days, rainy days, and dim streetlights: laser and vision alternate as the main observation sources, while the inertial measurement unit always provides high-frequency predictions.

[0047] The real-time attitude and position of the aircraft require the fusion of three types of observations. The module employs the Extended Kalman Filter (EKF) recursive algorithm. Its core idea is to use the high-speed integration results of the inertial measurement unit as short-term predictions, and use low-frequency measurements from lidar and visual ranging to correct deviations, continuously correcting inertial navigation drift.

[0048] Within each filtering cycle, the system first extrapolates the acceleration and angular velocity to the new position based on the attitude calculated in the previous moment. Then, it compares the radar point cloud with a 3D road map to provide a rough position of the aircraft in the road coordinate system. Finally, visual marker ranging provides fine alignment. The filter balances the uncertainties of the predicted and observed quantities through a gain matrix, ensuring that the output attitude angles and position coordinates are numerically smooth while quickly capturing real changes. This allows the flight positioning module to maintain relative accuracy at the centimeter or even millimeter level even if GPS multipath propagation or signal blockage renders satellite signals unavailable.

[0049] After obtaining a high-confidence assessment of the drone's status, it is necessary to plan a feasible flight path for the drone that does not disrupt traffic. The module accomplishes this task using a two-level global-local planning framework:

[0050] At the global level, the algorithm searches for the shortest feasible path from the current position to the target camera in the 3D road grid map stored locally on the drone, requiring that the flight altitude always be higher than the lane and that the drone always move laterally away from the street lamp pole within a safe bandwidth.

[0051] At the local level, the lidar scans for vehicles, high-load trucks, or construction platforms above the lane in real time. When any temporary obstacle is detected to be less than a set safety threshold from the predetermined track, the local replanner regenerates a detour route within tens of milliseconds and then smoothly connects it to the original track.

[0052] Through this "two-layer" strategy, the drone will not be at risk of collision with vehicles that suddenly rush out in the air, and it can also ensure the ideal angle of incidence when entering the camera's adsorption area.

[0053] After the drone arrives at the front side of the camera, its flight speed is gradually reduced to about 0.1 meters per second. Then the module switches control to the quaternion attitude controller: the outer speed loop suppresses the horizontal and vertical speed errors to an extremely low level; the inner attitude loop uses the quaternion error as the target to generate motor angular velocity commands, so that the direction of the drone's head is completely perpendicular to the normal of the camera mirror.

[0054] When the position deviation and attitude error simultaneously reach the millimeter-level and angular component-level thresholds, the vacuum pump starts, and the soft adsorption ring forms a pressure difference, gently "attaching" the drone to the outside of the camera housing. This completes the final goal of the flight positioning mission: to achieve millimeter-level docking attitude without interference from traffic on the main road.

[0055] Preferably, the flight positioning module includes an inertial measurement unit, a lidar, and a visual recognition device. The flight positioning module uses a fusion algorithm to jointly calculate the acceleration and angular velocity output by the inertial measurement unit, the point cloud coordinates output by the lidar, and the attitude angle output by the visual recognition device to obtain the three-dimensional position and attitude of the UAV.

[0056] This module acts as a "spatial reference generator" in the entire UAV camera cleaning system: it assimilates the original observations scattered across different coordinate systems, frequencies, and noise distributions into three-dimensional position information (x, y, z) and attitude angles (φ, θ, ψ) at the same time and in the same geographic reference system. The core principle of this process can be divided into three logical chains: coordinate integration, probability integration, and control integration.

[0057] Coordinate integration includes: multi-coordinate system mapping, and the output of the airborne inertial measurement unit (A). b ,ω b Located in system B. The lidar point cloud is natively located in the lidar coordinate system L. The visual recognizer obtains the pose of the camera coordinate system C relative to the target marker through perspective-n points (PnP).

[0058] During the deployment phase, the extrinsic parameter matrices T are obtained for B→L and B→C respectively. LB ,T CB During runtime, the radar point cloud and visual measurements are unified into the machine system through extrinsic parameter transformation, and then processed by the rotation matrix R. BE Translate to geographic coordinate system E. In this way, all measurements can be directly compared numerically without the need for cross-system interpolation.

[0059] The sampling timestamps of different sensors are first written to the IMU clock domain via hardware timestamps, and then linear interpolation is performed at the software layer; the error is controlled at the millisecond level to avoid the EKF mixing the "old attitude" with the "new position" and causing false vibrations.

[0060] Probabilistic integration extends the essence of Kalman filtering. Inertial navigation provides short-term high-frequency predictions but accumulates zero bias; laser / vision provides low-frequency absolute measurements but inherently carries Gaussian noise. EKF uses a linear approximation to pull the nonlinear state-observation model back into a Gaussian framework, with the core equation:

[0061]

[0062] in, Indicates prior knowledge (inertial navigation prediction); z t Indicates measurement (laser / vision); K t Kalman gain represents the quantitative weight that measures "who to trust".

[0063] If the visual signal is distorted (nighttime glare), its observation covariance R vision Increase, K t The system automatically reduces the visual weight, allowing the laser to dominate, and vice versa. This adaptive trade-off allows the system to continue operating even when the sensor fails, rather than simply "shutting down when the signal is lost."

[0064] The zero-biased estimation converges synchronously, and the zero-biased b a ,b g The state variables are included in the filter—meaning that EKF not only estimates attitude but also the health of the IMU itself. As visual-laser ground truth is continuously injected, the zero-bias covariance gradually converges, and the inertial navigation drift is confined to a small range.

[0065] Integrated control transforms the solution results into safe trajectories and attitudes. Trajectory cost function design: global planning is not simply about finding the shortest path, but about minimizing the overall cost.

[0066]

[0067] Where J represents the global planning cost function, and the smaller the sum, the better the trajectory; α represents the route length weight; β represents the altitude penalty, ensuring the trajectory is always above the safe altitude; γ represents the lateral penalty, constraining the trajectory from being too close to the lamppost; ||p k -p k-1 The Euclidean norm represents the spatial distance between two adjacent nodes; p k This represents the three-dimensional position vector (x, y) of the k-th track node in the geographic coordinate system. k ,y k ,z k );p k-1 f represents the three-dimensional position vector of the (k-1)th track node; height (p k ) represents the height penalty function, when node p k height z k Returns a positive value if the set safe altitude is exceeded; otherwise, returns zero.lateral (p k ) represents the lateral penalty function, when node p k It returns a positive value if the lateral offset exceeds the safe bandwidth, otherwise it returns zero; this function allows "safety" and "economy" to be weighed numerically, rather than pieced together from empirical thresholds.

[0068] Attitude quaternion closed-loop calculations exhibit singularities around ±90° with classic Euler angles. The module is then modified to use quaternion errors.

[0069]

[0070] Where q d For the desired posture, q c This is the current stance; This is a quaternion multiplication. The attitude controller uses q. err Generate angular velocity commands to avoid gimbal lock.

[0071] In one embodiment, a soft ring adsorption is used to fix the UAV. The soft ring adsorption logic is as follows: when the position error ∈ p With attitude error ∈ θ Simultaneously, the vacuum pump is triggered when the pressure is below the threshold, generating adsorption force F through pressure difference ΔP. adh =ΔP·A. At this point, the control system freezes the external position ring, retaining only attitude fine-tuning to avoid secondary displacement caused by pump suction vibration. It should be noted that the method of cleaning the camera by fixing the drone body in this invention is limited to soft ring adsorption, but it can also be any other fixing method based on three-dimensional coordinates. In some special cases or under certain requirements, the drone can also be used to perform cleaning by hovering.

[0072] The system incorporates environmental adaptation and fault tolerance mechanisms. If the visual recognizer loses its marker for N consecutive frames, the weight scheduler directly reduces its weight to zero, but retains the visual thread to ensure no further capture opportunities are interrupted. If the LiDAR experiences scattering noise due to water mist, causing a decrease in the density of the detected point cloud, the system automatically increases the R-value. LiDARR When density recovers, it gradually reverts back. Simultaneous failure of both channels (τ) max Within seconds, the command aircraft automatically climbs to a preset altitude to avoid being above the vehicle, and then waits for the observation to recover before descending back to the work surface. This layered fault tolerance avoids "single-point failure" and is a necessary condition for uninterrupted operation on the main road.

[0073] Using the solution of this invention, in an urban canyon test section, based on total station true value comparison, the root mean square error of positioning is ≤1mm, and the root mean square error of attitude is ≤0.2°. The local obstacle avoidance replanning operation time is 20ms; the attitude loop closed-loop delay is <3ms, which is sufficient to suppress the disturbance of lightweight multirotors in 5m / s gusts. The docking success rate is higher than 98% in three scenarios: sunny day, night, and light rain.

[0074] Through the aforementioned multi-layered principles, the flight positioning module achieves a continuous closed loop from multi-source noise observation to high-confidence attitude-position output, providing a spatial reference for "non-road-closed" camera cleaning that meets traffic safety requirements and industrial-grade reliability.

[0075] Preferably, the flight positioning module further includes a path planning unit, which generates an obstacle avoidance trajectory based on the three-dimensional position and attitude, and writes the trajectory into the flight command data.

[0076] The path planning unit in the flight positioning module is based on the concept of a continuous potential energy field: first, all risk factors in the road environment are mapped into potential energy densities related to spatial coordinates; then, the minimum-cost three-dimensional trajectory is searched within this potential energy field; finally, the trajectory is smoothed and written into the flight command data. Its principles include the following:

[0077] First, establish the potential energy field. Calculate the potential energy for each spatial point p = [x, y, z]:

[0078]

[0079] Where β represents the altitude penalty weight, h c The safe flight altitude threshold is represented by z, where z represents the current altitude; γ represents the lateral penalty weight, and y represents the current altitude. t w represents the lateral coordinate of the streetlight pole's centerline. s κ represents the horizontal security bandwidth, y represents the current horizontal coordinate; j p represents the repulsive force coefficient of the j-th dynamic obstacle. j σ represents the center position of the obstacle, σ represents the equivalent radius of the obstacle, and 1(·) is an indicator function used to activate the penalty when the limit is exceeded.

[0080] Secondly, define the comprehensive cost. Discretize the trajectory into a sequence of nodes p0, p1, ..., p N Minimize the following expression:

[0081]

[0082] In the formula, J represents the total cost of the trajectory, α represents the distance weighting coefficient, and ||p k -p k-1 || represents the Euclidean distance between adjacent waypoints; the other symbols are the same as in the previous equation. The search algorithm finds the sequence of nodes in the three-dimensional grid that minimizes J, thus obtaining the initial track.

[0083] Finally, dynamic consistency shaping is performed. A fifth-degree polynomial is used for each segment of the trajectory:

[0084] p(t) = a0 + a1t + a2t 2 +a3t 3 +a4t 4+a5t 5

[0085] Smooth the surface and use the least third derivative energy. Determine the coefficient a for the target i This eliminates sharp angles in the trajectory, ensuring that the multi-rotor controller does not generate excessive angular acceleration during tracking. The time-displacement sequence after shaping is written into the flight command data for real-time execution by the attitude control loop. The entire process provides an obstacle avoidance trajectory that balances safe distance, energy economy, and dynamic smoothness, laying the trajectory foundation for UAVs to achieve non-closed road attachment in main road environments.

[0086] The stain detection and decision module is used to collect millimeter-wave reflection coefficient data, polarization scattering matrix data and acoustic echo data, fuse them to generate stain spectrum fingerprint data, and determine cleaning parameters based on the stain spectrum fingerprint data.

[0087] The task of this module is to determine the type and thickness of stains on a mirror surface without removing the protective cover or disrupting traffic, and to provide a one-time executable combination of dry cleaning parameters. Its mechanism can be broken down into four levels: "multi-physical quantity acquisition → spectral fingerprint construction → graph convolution representation → Bayesian parameter search".

[0088] The principles of multi-physical quantity acquisition include:

[0089] Millimeter wave reflection: 22–26 GHz FMCW scanning, the oil film will reduce the surface dielectric constant, causing the reflection coefficient to drop at 24 GHz;

[0090] Polarization imaging: Linearly polarized light passes through the three-layer interface of oil-glass-air, and the off-diagonal elements of the Mueller matrix are most sensitive to the film thickness;

[0091] Acoustic echo: A 40kHz ultrasonic pulse is scattered at the air-dirt membrane interface, and the rough surface will broaden the main peak frequency band.

[0092] The three-channel complementarity enables the system to achieve separability characteristics in both daytime and rainy / foggy environments.

[0093] Spectral fingerprint construction: All original observations are normalized and concatenated into a ten-dimensional vector.

[0094] F=[Γ 24 ,Γ 25 ,S 11 ,S 12 ,S 21 ,S 22 ,R a ,f ac ,L UV ,I VIS ]

[0095] Among them, Γ24 ,Γ 25 Indicates the amplitude of the reflection coefficient at 24GHz and 25GHz; S ij Represents the elements of the Mueller matrix; R a Indicates the ultrasonic roughness index; f ac Indicates the acoustic peak frequency; I UV I VIS Indicates fluorescence intensity.

[0096] To fully utilize the physical coupling relationships between features, F is projected onto a ten-node complete graph. Edge weights are calculated using Gaussian similarity.

[0097]

[0098] Among them, w ij f represents the similarity between node i and node j; i Let represent the o-th eigenvalue; σ represents the width factor. The fingerprint embedding vector h is obtained after two layers of graph convolution, and its dimension is much lower than the original features, facilitating subsequent optimization.

[0099] Cleaning parameter search, the parameters to be optimized are denoted as:

[0100] p = [N p ,τ s ,τ f ]

[0101] Where, N p Indicates the number of plasma pulses; τ s Indicates the duration of acoustic cavitation sweep; τ f This indicates the duration of the ion curtain purging process.

[0102] Bayesian optimization minimizes the objective function using the desired improvement criterion:

[0103]

[0104] in, Let represent the overall cost; R(h,p) represent the residual exponent predicted by the model; λ represent the energy penalty weight; and E(p) represent the energy estimation function. After thirty iterations, it converges to the cleansing parameter p. * .

[0105] Through this invention, oil film, dust, and water mist scenarios can be partitioned and clustered in a ten-dimensional feature space, and the mean square error of the predicted residual index is maintained at 10. -3 The parameters are of a very low magnitude; the convergence time for parameter search is less than 0.3s, which can be paralleled with the flight attachment action; after cleaning, the actual residual index is updated again by resampling, which is used as a monitoring signal to fine-tune the network weights, so that the energy consumption of subsequent operations decreases batch by batch.

[0106] Preferably, the stain detection and decision module includes a multimodal data fusion unit and a fingerprint optimization unit. The multimodal data fusion unit fuses millimeter-wave reflection coefficient, polarization scattering matrix and acoustic echo into a stain spectrum fingerprint. The fingerprint optimization unit outputs a set of cleaning parameters based on the stain spectrum fingerprint.

[0107] In scenarios where main roads are not closed for cleaning, the oil film, dust, and water mist adhering to the camera housing exhibit differences in electromagnetic, optical, and acoustic properties. The module first allows the millimeter-wave probe, polarization camera, and ultrasonic transducer to each select their "sensitive areas," interweaving the three physical quantities into a dirt spectrum fingerprint; then, a Bayesian search is used to establish a "minimum residue - minimum energy consumption" mapping between the fingerprint and cleaning parameters.

[0108] Multimodal signal complementarity includes:

[0109] Millimeter wave reflection: the oil film changes the dielectric constant of the shield surface, and the reflection depth at 24 GHz is monotonically variable with thickness.

[0110] In polarization imaging, the multilayer interface weakens the linear polarization retention, and the off-diagonal elements of the Mueller matrix fluctuate significantly with film thickness.

[0111] For acoustic echoes, the higher the roughness, the wider the 40kHz main peak and the greater the echo energy.

[0112] Each of the three has its blind spots: visual impairment at night, laser attenuation in rain and fog, and acoustic texture obscured by oil film; combining them can make these blind spots complementary.

[0113] The original data is compressed into a fingerprint. The module normalizes and sorts the three original data streams, extracts the ten most discriminative components, and then encodes "who is related to whom" into the graph structure using a complete graph and Gaussian edge weights. Two layers of graph convolution absorb the noisy original components into a more robust embedding vector h.

[0114] The Gaussian edge weight formula and the interpretation of each symbol have been listed in the previous paragraph and will not be repeated here.

[0115] The fingerprint optimization unit abstracts the cleaning action to be adjusted into a vector p = [N] p ,τ s ,τ f "Doing it cleanly" is measured by the residual index prediction R(h,p), while "doing it energy-efficiently" is measured by the energy consumption estimate E(p). The two are linearly combined to form a comprehensive cost:

[0116]

[0117] in, R(h,p) represents the expected operation; E(p) represents the predicted residual index; and λ represents the energy consumption estimate.

[0118] Bayesian optimization Treating it as a "black box" function, we estimate its shape using a Gaussian process, and then use the expectation value to refine and select the next set of p. After 30 samplings, p can be found. * —The most cost-effective number of pulses and duration for the current smudged fingerprint.

[0119] After cleaning, the quality inspection residue index was measured as the true value R. real The system will use R real With the prediction R(h,p) * The error is used to update the network weights in reverse, making the parameters of the next camera more accurate; energy statistics are also written back as the basis for adjusting λ, reducing energy consumption batch by batch.

[0120] In the application of this invention, even if one channel is completely distorted, the graph convolution can still provide correct parameters based on the remaining channels. The complete search is completed within 0.3 seconds, and the parameters are obtained before the attachment action is finished. Each real-world quality inspection reduces the prediction error, and over long periods, energy consumption can be reduced to less than half that of traditional water washing.

[0121] Preferably, the fingerprint optimization unit outputs the set of cleaning parameters with the goal of minimizing the quality inspection residue index using a Bayesian search algorithm.

[0122] After the drone is attached to the camera housing, this module must quickly answer two questions without spraying water or removing the housing:

[0123] 1. What kind of stain is on the mirror surface (oil film, dust, or fog droplets), and what is its approximate thickness and roughness?

[0124] 2. Given the current remaining power and energy consumption budget, how many plasma pulses N are needed? p , sweeping long sound cavitation τ s Blow the ion curtain for a few seconds τ f Only then can the stains be removed below the "residual threshold".

[0125] The principle of signal complementarity is as follows: millimeter waves are most sensitive to changes in dielectric constant around 24 GHz, thus showing a prominent response to oil film thickness; polarization imaging focuses on the "depolarization" effect of the three-layer interface of the shield glass-oil film-air on linear polarization, which can quantify film thickness and distribution uniformity; acoustic echoes at 40 kHz reflect surface micro-roughness, and the larger the dust particles, the wider the main peak.

[0126] If any one of the three signals degrades (e.g., due to visual overexposure at night or laser absorption caused by rain or fog), the other two signals can still remain separable, thus enabling the system to construct a stable stain spectrum fingerprint under all weather conditions and lighting.

[0127] The low-dimensional embedding vector h (obtained by the aforementioned graph convolution) is fed into the residual index predictor. The predictor outputs the probability distribution of the "cleaned residual index," rather than a single-point value; therefore, the cost function... Expected value using residual index The penalty coefficient λ is derived from the energy management module and is weighted by energy consumption E.

[0128] Bayesian optimization treats the entire cost surface as a black box and performs a posterior estimation using a Gaussian process; then, it uses the Expected Improvement (EI) criterion to select the next set of clean parameters, approximating the optimal combination p with the fastest convergence speed. * .

[0129]

[0130] p * Represents the optimal set of cleaning parameters; p = [N] p ,τ s ,τ f ], where N p τ represents the number of plasma pulses. s τ represents the duration of acoustic cavitation sweep. f Indicates the duration of ion curtain purging; λ represents the expected residual index predicted by fingerprint embedding h and parameter p; λ represents the energy consumption penalty weight, which is dynamically given by the energy management module; E(p) represents the cleaning energy consumption estimation function, which monotonically increases with the number of pulses and the driving duration.

[0131] The fingerprint vector F and the edge weight formula w have been given previously. ij and comprehensive costs This will not be repeated here.

[0132] This "fingerprint-optimization" chain allows drones to provide a one-time, energy-aware dry cleaning solution for each camera stain without human judgment. In actual tests, even under extreme conditions such as an oil film of 60μm, a dust roughness of 30μm, and a nighttime illumination of 5lx, it can still keep the residual index below the threshold and control the energy consumption to less than half that of traditional water washing.

[0133] The cleaning and quality inspection module is used to sequentially perform cleaning processes according to the cleaning parameters under the power trajectory defined by the model prediction controller, acquire cleaning progress field data and quality inspection residual index data in real time, and adjust the cleaning parameters to repeat the process when the quality inspection residual index data exceeds the threshold, until the quality inspection residual index data does not exceed the threshold.

[0134] After the drone completes attachment, the cleaning system must, within a single dwell time window without closing traffic, allocate the energy of the dry three-stage process (plasma, acoustic cavitation, and ion curtain) to the most needed locations and prove that the mirror surface is below the residue threshold. This module consists of four internal closed-loop systems connected in series: power trajectory planning, energy-position coupled execution, residue index quantization, and parameter adaptive updating. The core physics-control logic is explained layer by layer in textual form below, only providing calculation formulas that have not yet appeared but are indispensable.

[0135] The principle of power trajectory planning is that the controller uses the current remaining energy E res With cleaning parameter p = [N p ,τ s ,τ f Treating this as a hard constraint, predict the discharge power sequence P0, P1, ..., P over the next H discrete steps. H The goal of MPC optimization is to push the residual exponent below a threshold with as little energy as possible; therefore, the objective function is constructed as follows:

[0136]

[0137] P k E represents the discharge power at step k; k =P k Δt represents the energy consumed in this step; C k This represents the contribution of the power at this step to the predicted residual exponent; μ represents the residual penalty weight; Δt represents the discrete time step. Constraints:

[0138] 0≤P k ≤P max

[0139]

[0140] P max This represents the upper limit of the device's safe power. The solution obtained... This is the power trajectory of this cleaning process.

[0141] Energy-space coupled execution, the three-stage device operates in sequence according to P. Driven by [the system], it generates a local energy density ρ(x,y,t) at the mirror coordinates (x,y) at each moment. The module integrates the energy density over time to obtain the cleaning progress field:

[0142]

[0143] Φ(x,y,t) represents the cumulative energy projected at coordinates (x,y) up to time t; ρ(x,y,τ) represents the instantaneous energy density; and τ represents the integration time variable. The progress field can be imaged into a heatmap in real time to identify local dead zones.

[0144] Residual index measurement: After each completion of a three-stage process cycle, the system re-collects the stain spectrum fingerprint F. new Clean reference fingerprint F ref Based on this, calculate the relative residual index:

[0145]

[0146] η represents the residual exponent; ||·||2 represents the L2 norm. When η≤η max (η max If a threshold is set, the mirror surface is deemed to meet the standard; otherwise, it enters the parameter adaptation stage.

[0147] Parameters are updated adaptively if η > η max The module uses gradient correction to fine-tune the cleaning parameters:

[0148]

[0149] α represents the learning rate; This represents the gradient of the residual exponent with respect to the parameter vector, and its value is approximated by finite differences.

[0150] The learning rate decays exponentially:

[0151] α m =α0e -δm

[0152] α m Let represent the learning rate for the m-th iteration; α0 represent the initial learning rate; and δ represent the decay coefficient. The updated p re-enters MPC until the residual exponent meets the threshold or energy is exhausted.

[0153] This invention, in a typical scenario with an oil film thickness of 60μm, showed that η could be reduced from 0.12 to 0.04 in just two iterations, below the threshold of 0.05. The combination of power trajectory optimization and local progress field imaging improved the decontamination efficiency per unit joule by approximately 35% compared to the constant power-constant duration scheme. The entire process is free of droplet splashing and high-brightness visible light, thus not affecting traffic flow or driver visibility.

[0154] Preferably, the cleaning and quality inspection module includes a plasma generator, an acoustic cavitation transducer, and an ion curtain device in sequence. The cleaning and quality inspection module also includes a model prediction controller, which generates a power trajectory within a preset energy constraint and drives the plasma generator, the acoustic cavitation transducer, and the ion curtain device.

[0155] After the drone attaches, the module needs to decontaminate the device using three dry methods—plasma, acoustic cavitation, and ion curtain—in the optimal order and at the optimal power within a fixed energy budget. Simultaneously, it monitors the residue and performs secondary enhancements as needed. The core idea is to first have the Model Predictive Controller (MPC) break down the remaining electrochemical energy into a power-time trajectory, which is then consumed sequentially by the three devices. If the real-time residue still exceeds the threshold, the energy can be replanned before the energy is depleted.

[0156] Energy distribution—power-duration coupling, assuming residual energy E res (E res (This represents the total available energy of the current battery and supercapacitor), which MPC breaks down into three energy quotas:

[0157] E1 + E2 + E3 = E res

[0158] E1 represents the energy allocated to the plasma generator; E2 represents the energy allocated to the acoustic cavitation transducer; and E3 represents the energy allocated to the ion curtain device.

[0159] Power trajectory The device is sliced ​​so that the power in step k only drives the device currently in the working window, ensuring that the three devices do not compete for the power peak concurrently.

[0160] The stage energy efficiency weight is determined by querying the empirical coefficient ξ based on the spectral fingerprint provided by the stain detection module. i , representing "the efficiency of removing the current fouling membrane per unit joule in stage i". MPC solves a linear programming problem with the objective of maximizing residue reduction:

[0161]

[0162] stE1+E2+E3=E res

[0163]

[0164] ξ i Indicates the unit joule decontamination efficiency of device i; This indicates the upper limit of energy that a single device can withstand while maintaining structural safety.

[0165] Seek Then, divide it by the discharge voltage U. i The total charge of each discharge segment is obtained and used as the upper limit for the number of pulses, sweep frequency, or purge duration.

[0166] In a local energy density closed-loop configuration, when the three devices are operating, the instantaneous energy density ρ(x,y,t) on the mirror coordinates (x,y) is continuously integrated into a progress field Φ(x,y,t). Each module Δt... imgRedraw the heatmap and calculate the local energy difference in seconds:

[0167]

[0168] Φ ref This represents the empirically sufficient energy threshold; [z] + Represents max(z,0). If ΔΦ min A value >0 indicates that there are still blind spots that have not been thoroughly cleaned, requiring additional local acoustic cavitation sweeps with an additional energy of [energy value missing].

[0169] Secondary determination of residual index: After the cleaning cycle is completed, proceed as follows:

[0170]

[0171] Calculate the residual index. If η > η max MPC uses the unused energy balance from the previous round, ΔE = E res - As the new E res Re-enter the initial step until η≤η max Or the energy is exhausted.

[0172] Through this invention, the stage energy efficiency weight ξ i Energy is prioritized for the most efficient decontamination device; the progress field closed loop allows any local energy gap to be compensated by secondary frequency sweep; one or two iterations within a 60s dwell window are sufficient to meet the residual threshold without triggering energy overdraft.

[0173] Preferably, the model prediction controller updates the power trajectory based on the cleaning progress field data and the quality inspection residual index data, and sends the updated power trajectory to the plasma generator and the acoustic cavitation transducer.

[0174] After the initial power trajectory is issued, the three-stage dry process begins to apply to the mirror surface. The data bus at the UAV continuously transmits two types of feedback: the cleaning progress field Φ(x,y,t) characterizes the spatial-temporal distribution of accumulated energy; and the quality inspection residue index η(t) measures the relative error of the mirror surface from the target cleanliness level.

[0175] The controller treats both as "state measurements" and resolves the small-range MPC at a frequency of 10 Hz to form a self-tuning power trajectory. This closed loop mainly involves two coupling mechanisms.

[0176] The gain injection of error potential energy is defined as follows: for any time t, a local energy gap is defined on the mirror coordinates (x, y):

[0177] ΔΦ(x,y,t)=[Φ ref -Φ(x,y,t)] +

[0178] Φ ref This represents the energy threshold required for experience, [·] + The function is a positive cutoff function. Mapping the overall gap to the power adjustment:

[0179]

[0180] Where, k Φ The gain represents the power coupling gain, and A represents the effective area of ​​the mirror.

[0181] Global correction of the residual index; if the latest residual index η(t) is still higher than the threshold η max The controller then increases the power by one level according to the proportional-integral law:

[0182]

[0183] k p Represents the residual proportional gain, k i This represents the residual integral gain, and τ is the integration time window.

[0184] Power trajectory online correction: For the next control step k, the original trajectory power... Revised to:

[0185]

[0186] And immediately issue it to the plasma generator and the acoustic cavitation transducer; if Exceeding the safety limit P max Then take P max This real-time overlay does not disrupt the original MPC optimization structure, but allows for rapid compensation of power for local dead zones and global residuals.

[0187] This invention automatically provides additional acoustic cavitation energy to mirror areas with large gaps in the progress field, avoiding the "zebra stripe cleaning" effect. Proportional-integral adjustment ensures the residual index is pushed into the threshold within 1-2 rounds without power oscillations due to noise. All power corrections are dynamically deducted from the remaining energy budget of the previous moment, never triggering battery undervoltage.

[0188] Through the aforementioned dual-channel gain injection of "local gap-global residue", the model predictive controller can autonomously update the power trajectory and precisely drive the plasma and acoustic cavitation devices under the time pressure of continuous traffic flow, ensuring that quantifiable cleanliness results are achieved with a single application.

[0189] The energy module is used to control the drone to return to base after the quality inspection residual index data does not exceed the threshold, and to encapsulate the cleaning progress field data, quality inspection residual index data and energy settlement data into a closed-loop data package to adjust the self-learning model's own parameters and generate the next flight command data.

[0190] The specular residue index η has met the threshold η max Afterwards, this module takes over control and completes three key tasks: ① Determine if the remaining energy is sufficient for a safe return; ② Encapsulate the energy-mass information of the entire operation into a closed-loop data packet; ③ Incrementally update the self-learning model based on this data packet and output the flight command data for the next sortie. The core logic is explained in layers below, and the necessary calculation formulas appearing for the first time are given.

[0191] Return-to-home energy assessment and scheduling: After the drone has stopped for cleaning and is still hanging outside the camera, the first step is to confirm its remaining battery power E. rem Can it cover the entire process of return climb, cruise, landing, and battery swapping at edge base stations? Define the minimum energy required for return:

[0192] E ret =mgh loss +κd home

[0193] E ret The return-to-home energy requirement is represented by m; the total mass of the drone is represented by g; and the gravitational acceleration is represented by h. loss This represents the equivalent height difference between a descent and a subsequent ascent; d home κ represents the horizontal distance to the base station; κ represents the energy consumption coefficient per unit distance of cruise.

[0194] When E rem ≥E ret The energy module issues a return command; otherwise, it immediately enters the "low-power hovering and requesting ground assistance" mode.

[0195] After the operation closed-loop data packet is constructed and the return navigation is initiated, the module encapsulates four types of data in a unified manner: the cleaning progress field grid Φ(x,y,t). end ); Final residual index η end Energy settlement triplet [E] use E rem E ret ], where E use This indicates the energy consumed in this round of cleaning; timestamp and pose trajectory hash (ensuring backtracking consistency). Data packets are written to the local linked list after being internally salted and hashed, and then asynchronously uploaded to the cloud blockchain to prevent tampering.

[0196] The self-learning model uses incremental updates and a proximal policy optimization (PPO) framework, employing energy-quality data from a single round of work as the "instant reward" to continuously fine-tune the policy parameters θ. The reward function defined in this invention is:

[0197]

[0198] R represents the task reward value; w1, w2, w3 represent the three weighting coefficients; η end E represents the final residual index; use E rem Same meaning as above; T cycle Indicates the total duration of this round of operations; T max This indicates the maximum permitted length of stay.

[0199] Gradient update rule:

[0200]

[0201] β represents the learning rate; This represents the expected reward under the current strategy.

[0202] The updated strategy immediately infers the flight command data for the next sortie (takeoff point, target camera sequence, energy budget, and recalibrated initial cleaning parameters) and transmits it back to the local drone queue.

[0203] Preferably, the energy module includes an automatic battery swapping mechanism and an energy management unit. The automatic battery swapping mechanism completes battery replacement after the UAV returns to base, and the energy management unit records the power consumption before and after the battery swap and outputs energy settlement data.

[0204] The drone completed the camera cleaning and confirmed that the residual index η ≤ η max Then, the process transitions to return-to-base – battery swap – energy settlement. The energy module within the ground-based edge base station consists of two parts: an automatic battery swapping mechanism and an energy management unit. These two components work together to ensure uninterrupted fleet operation and provide accurate energy consumption labels for the cloud-based self-learning model.

[0205] The mechanical-electric coupling concept of the automatic battery swapping mechanism includes mechanical positioning: the base station gripper uses a V-shaped guide groove to align the UAV battery compartment with sub-millimeter precision, avoiding misalignment during insertion and removal under wind load or ground vibration.

[0206] Electrical isolation and thermal management: when plugging or unplugging, first disconnect the DC main negative terminal, then the positive terminal, and finally release the signal PIN; the order is reversed when inserting a new battery. The plug has an embedded heat pipe and phase change material, allowing residual heat to be quickly dissipated after high-rate discharge.

[0207] The battery swapping cycle takes less than 8 seconds to complete one complete set of actions: plugging, inserting, and locking. During the switching interval, the avionics are powered by a supercapacitor, and the airborne controller does not need to be restarted.

[0208] Energy management unit metering and billing, power measurement: The battery pack's ampere-hour integrator and terminal voltage are read before and after battery swapping to obtain Q. pre U pre With Q post U post Q pre U represents the remaining charge (Ah) before battery swapping. pre Q represents the corresponding terminal voltage. post U post This is the initial value for the new battery. Energy calculation:

[0209] E use =U nom (Q pre -Q post (Wh)

[0210] E use This indicates the actual energy consumption of this round of cleaning; U nom This indicates the battery's nominal voltage.

[0211]

[0212] SOC pre Indicates the relative remaining capacity of the old battery; Q nom Indicates the rated capacity.

[0213] Closed-loop data packets, the energy management unit puts [E use SOC pre SOC post ,Φ(x,y,t end ),η end Encapsulated as a "job closed-loop data packet", where Φ(x,y,t) end ) represents the final cleaning progress field, η end This is the final residual index. Data packets are uploaded to the blockchain after being timestamped and include the task ID.

[0214] The incremental update cloud strategy for self-learning models uses energy consumption-quality data as an immediate reward:

[0215]

[0216] w1, w2, w3: Weighting coefficients; E rem Remaining energy after return; T cycle Total duration of this round; T max : Maximum allowed dwell time; strategy parameter θ is:

[0217]

[0218] (β is the learning rate) is fine-tuned, and flight command data for the next sortie is generated.

[0219] This invention enables battery swapping in under 8 seconds, ensuring seamless "landing-battery swapping-takeoff" for drones during peak hours. Energy consumption is traceable with dual metering to eliminate temperature drift errors, and closed-loop data supports subsequent auditing. The self-evolving strategy continuously optimizes the energy benchmark using the real energy consumption-quality curve, reducing energy consumption batch by batch under the same cleanliness level.

[0220] The energy module utilizes three mechanisms—rapid battery swapping, precise energy calculation, and data-driven strategy updates—to enable unmanned cleaning operations of main road cameras to achieve measurable energy savings and continuous self-optimization.

[0221] Preferably, after generating the operation closed-loop data packet, the energy management unit calculates the hash value of the operation closed-loop data packet and attaches a digital signature, writes it into the ledger through the blockchain node, and then sends it to the cloud through the network communication interface. The self-learning model uses an incremental gradient boosting decision tree to train the received operation closed-loop data packet online, and returns the next flight command data after the training is completed.

[0222] After a cleaning operation is completed, the energy module quickly uploads the "energy-quality" closed-loop data to the blockchain and pushes it to the cloud. The cloud uses incremental gradient boosting decision tree (GBDT) for online learning, completing policy self-evolution within seconds and sending back the flight instructions for the next sortie. The chain consists of three core links: trusted encapsulation, ledger writing, and online training.

[0223] Trusted encapsulation involves generating a data packet (pkg) containing energy settlement, residual index, progress field hash, timestamp, and task ID. The energy management unit first performs a one-way hash and then signs it with the base station's private key. This allows any backend node to verify the signature using its public key, ensuring the data is tamper-proof.

[0224] H = SHA256 (pkg)

[0225]

[0226] H represents the packet hash value; pkg represents the original closed-loop data packet; S represents the digital signature; k prv This represents the base station's private key. The signed packet is written to the ledger via the local blockchain node and then uploaded to the cloud via the 5G interface.

[0227] Incremental GBDT online training, feature vector extraction after cloud signature verification:

[0228] z = [E use ,η end ,T cycle ,…]

[0229] This data is then fed into the incremental GBDT as a new sample. The model performs gradient iterations on the target variable (theoretical energy consumption for the next task):

[0230] 1. Residual calculation:

[0231]

[0232] 2. Forecast Update:

[0233]

[0234] y represents the residual of the i-th sample when the m-th increment tree is used; i This represents the actual target energy consumption of the i-th sample; This represents the predicted value of the first m-1 trees for the i-th sample; This represents the new prediction after adding a new tree; v represents the learning rate; T (m) (z i ) represents the m-th new regression tree in sample z. i Output on; z i Let represent the feature vector of the i-th sample. Incremental updates do not require replaying historical data and can complete a model refresh within seconds.

[0235] Once the flight command is issued, the updated model, combining the camera cluster distribution and the remaining energy of the aircraft group, generates a list of takeoff base stations and target cameras for the next sortie, as well as an estimated energy budget. The initial cleaning parameter vector is p0. Commands are transmitted back to the standby UAVs via the MQTT bus, enabling seamless rotation.

[0236] like Figures 3-8 As shown, the application of this invention in the body of a drone includes:

[0237] The UAV is equipped with a flight-borne equipment (16) and a three-axis gimbal (27). The UAV is connected to the three-axis gimbal (27) via a gimbal interface (29). The three-axis gimbal (27) is equipped with a gimbal load device (7). The gimbal load device (7) includes a nozzle (15), a miniature camera, an algorithm core board (11), and integrates a "flight positioning module", a "stain detection and decision module" and a "cleaning and quality inspection module". The flight positioning module is used to read flight command data and synchronous positioning-map construction data and drive the UAV to fly to the target camera. The stain detection and decision module is used to collect millimeter wave reflection coefficient, polarization scattering matrix and acoustic echo and output cleaning parameters. The cleaning and quality inspection module performs plasma, acoustic cavitation and ion air curtain cleaning in sequence under the power trajectory limited by the model prediction controller and outputs the cleaning progress field and quality inspection residue index in real time.

[0238] Two sets of liquid storage tanks (5) are symmetrically installed on both sides of the drone for storing cleaning fluid and clean water respectively. The liquid storage tanks (5) are connected to a micro water pump (18) through an electromagnetic three-way valve (24). The nozzle (15) is connected to the micro water pump (18) through a hose.

[0239] The flight-borne equipment (16) is fixed to the top of the UAV by screws. The flight-borne equipment (16) includes a micro water pump (18), a relay control board (21), a flight control core board (17), and an electromagnetic three-way valve (24). The flight control core board (17) and the algorithm core board (11) synchronously push flight attitude, cleaning parameters, and power trajectory commands through a high-speed bus and are connected to the data link of the "energy module". The energy module is located at the edge base station and includes an automatic battery swapping mechanism and an energy management unit. After the quality inspection residual index does not exceed the threshold, it controls the UAV to return and completes the battery replacement. At the same time, it records the power before and after the replacement and encapsulates it into a closed-loop data package to update the self-learning model and generate the next flight command data.

[0240] The drone is equipped with a drone mounting bracket (1), and the drone mounting bracket (1) is equipped with a positioning groove (6). The liquid storage tank (5) is installed on both sides of the drone mounting bracket (1) through the positioning groove (6). The installation attitude information of the positioning groove (6) is corrected in real time by the flight positioning module to ensure load symmetry and flight balance.

[0241] The liquid storage tank (5) is provided with a liquid storage cavity (4). The liquid storage tank (5) is provided with an injection port (3) for injecting liquid and a water pump inlet (2) connected to a micro water pump (18). The remaining liquid level data of the liquid storage cavity (4) is reported to the energy module through the algorithm core board (11) for calculating the energy budget of the next flight.

[0242] The gimbal load device (7) is provided with a mounting base (10) and a camera mounting base (8) for installing a miniature camera. The camera mounting base (8) has a camera data cable through hole (9) on one side and a water pipe guide hole (12) on one side. The gimbal load device (7) is provided with an adapter plate slot (13) and a locking point (14) for inserting the eport interface adapter plate (22). The gimbal load device (7) is provided with a nozzle mounting hole for installing the nozzle (15). The nozzle (15) outlet is coaxially integrated with a plasma electrode and an ion air curtain nozzle to realize the three-stage dry process of the cleaning and quality inspection module.

[0243] The electromagnetic three-way valve (24) includes an electromagnetic three-way valve inlet (25) and an electromagnetic three-way valve outlet (23). The micro water pump (18) includes a water pump inlet (19) and a water pump outlet (20). The flight airborne equipment (16) is provided with a connection hole (26). The opening sequence of the electromagnetic three-way valve (24) is scheduled in real time by the model prediction controller in combination with the cleaning progress field data.

[0244] The inlet (25) of the electromagnetic three-way valve is connected to the liquid storage tank (5) through a water pipe. The outlet (23) of the electromagnetic three-way valve is connected to the inlet (19) of the water pump. The outlet (20) of the water pump passes through the connection hole (26) and is connected to the nozzle (15). The outlet pressure of the water pump is controlled in a closed loop by the cleaning parameters output by the stain detection and decision module.

[0245] The algorithm core board (11) is used to control the flight of the UAV and the angle of the gimbal. The algorithm core board (11) is electrically connected to the flight control core board (17) through the line, and reads the power trajectory generated by the model prediction controller through the high-speed interface, and updates the working current of the plasma generator and the acoustic cavitation transducer in real time.

[0246] The relay control board (21) is used to control the micro water pump (18) and the electromagnetic three-way valve (24). The flight control core board (17), the algorithm core board (11) and the relay control board (21) are electrically connected through the line. The relay control board (21) also receives the remaining energy threshold signal returned by the energy module to determine whether to enter the low-energy hovering mode.

[0247] A method for autonomous cleaning of main road cameras without road closure based on unmanned aerial vehicles (UAVs), used to execute the aforementioned autonomous cleaning system for main road cameras without road closure, includes the following steps:

[0248] Read flight command data and synchronized positioning and mapping data to drive the drone to the target camera and establish a fixed attitude;

[0249] Millimeter-wave reflection coefficient data, polarization scattering matrix data, and acoustic echo data are collected, fused to generate stain spectrum fingerprint data, and cleaning parameters are determined based on the stain spectrum fingerprint data.

[0250] According to the cleaning parameters, plasma discharge treatment, acoustic cavitation stripping treatment and ion curtain purging treatment are sequentially performed under the power trajectory defined by the model prediction controller to obtain cleaning progress field data and quality inspection residual index data; when the quality inspection residual index data is greater than the threshold, the cleaning parameters are adjusted and the treatment is repeated until the quality inspection residual index data is less than or equal to the threshold.

[0251] After the quality inspection residual index data is less than or equal to the threshold, the drone is controlled to return to the edge base station, and the cleaning progress field data, quality inspection residual index data and energy settlement data are encapsulated into a closed-loop data package for the self-learning model to adjust parameters and generate the next flight command data.

[0252] This invention reads flight commands and synchronizes positioning and mapping data to control a drone to reach the target camera and establish a stable hovering attitude using negative pressure adsorption and other methods. Millimeter-wave reflection coefficients, polarization scattering matrices, and acoustic echoes are collected sequentially; multimodal fusion generates a "stain spectrum fingerprint," based on which cleaning parameters are calculated. Under the power trajectory given by the model predictive controller, plasma discharge → acoustic cavitation stripping → ion air curtain purging are executed; the cleaning progress field and residual index are output in real time. If the residual index is higher than the threshold, the cleaning parameters are adjusted and the three-stage process is repeated until the residual index no longer exceeds the limit. After the residual index reaches the standard, the drone is instructed to return to the edge base station to complete battery swapping; the energy management unit encapsulates the cleaning progress field, residual index, and energy settlement information into a closed-loop data package and uploads it to the blockchain. A cloud-based self-learning model trains this data package online, updates the algorithm weights, and generates the next batch of flight commands, achieving a continuously optimized cleaning cycle.

[0253] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0254] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An unmanned aerial vehicle (UAV)-based arterial camera non-enclosed road autonomous cleaning system, characterized in that, Comprise: Flight positioning module for reading flight instruction data and simultaneous localization and mapping data, driving unmanned aerial vehicle to fly to target camera; Stain detection and decision module for collecting millimeter wave reflection coefficient data, polarization scattering matrix data and acoustic echo data, generating stain spectrum fingerprint data by fusion, determining cleaning parameters based on the stain spectrum fingerprint data, its processing flow is in turn multi-physical quantity collection, spectrum fingerprint construction, graph convolution representation and Bayesian parameter search, wherein: the multi-physical quantity collection is to obtain millimeter wave reflection coefficient data, polarization scattering matrix data and acoustic echo data; the spectrum fingerprint construction is to normalize and aggregate the collected data to generate stain spectrum fingerprint data; the graph convolution representation is to code the graph structure according to the physical correlation between fingerprint elements to obtain fingerprint representation; the Bayesian parameter search determines the cleaning parameters by taking the fingerprint representation as input; The stain detection and decision module comprises a multi-modal data fusion unit and a fingerprint optimization unit, the multi-modal data fusion unit fuses millimeter wave reflection coefficient, polarization scattering matrix and acoustic echo into stain spectrum fingerprint, and the fingerprint optimization unit outputs a set of cleaning parameters according to the stain spectrum fingerprint; the fingerprint optimization unit outputs the set of cleaning parameters by minimizing the quality inspection residual index as the target through the Bayesian search algorithm; The cleaning and quality inspection module is used for executing cleaning processing in turn under the power trajectory defined by the model predictive controller according to the cleaning parameters, acquiring cleaning progress field data and quality inspection residual index data in real time, and adjusting the cleaning parameters to repeat the processing when the quality inspection residual index data exceeds the threshold value, until the quality inspection residual index data does not exceed the threshold value; The energy module is used for controlling the unmanned aerial vehicle to return after the quality inspection residual index data does not exceed the threshold value, encapsulating the cleaning progress field data, quality inspection residual index data and energy settlement data as a job closed loop data package, adjusting the parameters of the self-learning model, and generating next flight instruction data.

2. The system of claim 1, wherein, The flight positioning module comprises an inertial measurement unit, a laser radar and a visual recognizer, the flight positioning module jointly solves the acceleration and angular velocity output by the inertial measurement unit, the point cloud coordinates output by the laser radar and the attitude angle output by the visual recognizer through a fusion algorithm to obtain the three-dimensional position and attitude of the unmanned aerial vehicle.

3. The system of claim 2, wherein, The flight positioning module further comprises a path planning unit, which generates an obstacle avoidance track according to the three-dimensional position and attitude, and writes the track into the flight instruction data.

4. The system of claim 1, wherein, The cleaning and quality inspection module comprises in turn a plasma generating device, an acoustic cavitation transducer device and an ion wind curtain device, and further comprises a model predictive controller, which generates a power trajectory within a preset energy constraint and drives the plasma generating device, the acoustic cavitation transducer device and the ion wind curtain device.

5. The system of claim 4, wherein, The model predictive controller updates the power trajectory according to the cleaning progress field data and the quality inspection residual index data, and sends the updated power trajectory to the plasma generating device and the acoustic cavitation transducer device.

6. The system of claim 1, wherein, The energy module comprises an automatic battery replacement mechanism and an energy management unit, the automatic battery replacement mechanism completes battery replacement after the UAV returns, and the energy management unit records the power before and after replacement and outputs energy settlement data.

7. The system of claim 6, wherein, The energy management unit calculates the hash value of the job closed loop data packet after generating the job closed loop data packet and attaches a digital signature, writes it into the ledger through the blockchain node, and then sends it to the cloud through the network communication interface, the self-learning model adopts incremental gradient boosting decision tree to perform online training on the received job closed loop data packet, and returns the next flight instruction data after training.

8. A method for unmanned aerial vehicle based main road camera non-closed road autonomous cleaning, used for executing the unmanned aerial vehicle based main road camera non-closed road autonomous cleaning system in any one of claims 1-7, characterized in that, The method comprises the following steps: reading the flight instruction data and the simultaneous localization and mapping data, driving the UAV to fly to the target camera and establish a fixed attitude; collecting millimeter wave reflection coefficient data, polarization scattering matrix data and acoustic echo data, fusing to generate stain spectrum fingerprint data, and determining cleaning parameters according to the stain spectrum fingerprint data; According to the cleaning parameters, sequentially execute plasma discharge processing, acoustic cavitation stripping processing and ion wind curtain blowing processing under the power trajectory defined by the model predictive controller, obtain cleaning progress field data and quality inspection residual index data; when the quality inspection residual index data is greater than the threshold value, adjust the cleaning parameters and repeat the processing until the quality inspection residual index data is less than or equal to the threshold value; After the quality inspection residual index data is less than or equal to the threshold value, control the UAV to return to the edge base station, and package the cleaning progress field data, quality inspection residual index data and energy settlement data into a job closed loop data packet, so that the self-learning model adjusts the parameters and generates the next flight instruction data.

Images