Main road camera non-closure autonomous cleaning system and method based on unmanned aerial vehicle

Abstract

The invention relates to the technical field of unmanned aerial vehicle intelligent maintenance and cleaning, in particular to an unmanned aerial vehicle-based main road camera non-closure autonomous cleaning system and method, and the system executes the following steps: an unmanned aerial vehicle hovers at the outer side of a camera, collects millimeter wave-polarization-acoustic signals to generate stain spectrum fingerprints, and obtains cleaning parameters through optimization; and the model prediction controller sequentially executes plasma, acoustic cavitation and ion air curtain dry cleaning under energy constraint, self-adjusting parameter iteration is carried out when residues exceed a threshold value, and return flight and battery replacement are carried out after the standard is reached. And the cleaning progress, the residual index and the energy consumption are packaged and chained, the cloud self-learning model updates online and outputs a next order of instruction, and non-stop, low-consumption and efficient autonomous cleaning is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent maintenance and cleaning of unmanned aerial vehicles, and particularly to a non-road-blocking autonomous cleaning system and method for main road cameras based on unmanned aerial vehicles. Background Art

[0002] At the main road intersections, traffic monitoring cameras are exposed to exhaust gas, oil film and dust environment for a long time. The mirror surface pollution will directly weaken the image resolution and affect the reliability of illegal evidence collection and signal linkage algorithms. If relying on manual or vehicle high-altitude operations, it is necessary to occupy the lane and there is a risk of falling from a height. Therefore, non-road-blocking autonomous cleaning of main road cameras based on unmanned aerial vehicles has become a rigid demand. The existing technology completes cleaning through a closed loop of unmanned aerial vehicle carrying a high-definition camera to take pictures → YOLO-EMA to identify stains → water spray for fixed-point flushing → secondary picture taking for re-inspection (Chinese invention patent, publication number: CN119216290A); the following problems exist in this solution: optical imaging is prone to missed detection and false detection in backlight, night or specular reflection scenarios; it is necessary to rotate around the camera for view finding, with a long operation time and great influence on the air drying speed and traffic flow; it increases the airborne weight and produces water stain backflow, and is not suitable for live equipment and cold regions; lack of negative pressure adsorption or mechanical clamping, and positioning drift under crosswind conditions; no monitoring of energy consumption, nor can it adaptively optimize subsequent task parameters. The above defects limit the cleaning efficiency, applicable environment and long-term operation and maintenance economy. Summary of the Invention

[0003] Aiming at the many problems existing in the above-mentioned prior art, the present invention provides a non-road-blocking autonomous cleaning system and method for main road cameras based on unmanned aerial vehicles. The present invention locks the unmanned aerial vehicle outside the camera, uses millimeter wave-polarization-acoustic three-way sensing fusion to generate stain spectrum fingerprints, and outputs cleaning parameters through Bayesian optimization; the model predictive controller drives three-stage dry cleaning of plasma, acoustic cavitation and ionic air curtain under energy constraints, and iterates through the residual index closed loop until it reaches the standard, and then returns to the base for battery replacement and uploads the energy consumption-mass data to the cloud for the self-learning model optimization, so as to realize continuous, efficient and low-consumption non-road-blocking autonomous cleaning.

[0004] A non-road-blocking autonomous cleaning system for main road cameras based on unmanned aerial vehicles includes:

[0005] A flight positioning module, configured to read flight instruction data and synchronize positioning and mapping data, and drive the unmanned aerial vehicle to fly to the target camera;

[0006] A stain detection and decision module, configured 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 processing under the power trajectory defined by the model prediction controller according to the cleaning parameters, obtain cleaning progress field data and quality inspection residual index data in real time, and adjust the cleaning parameters to repeat the processing 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 UAV to return after the quality inspection residual index data does not exceed the threshold, encapsulate the cleaning progress field data, quality inspection residual index data and energy settlement data into an operation closed-loop data packet to adjust the self-parameters of the self-learning model, and generate the next flight instruction data.

[0009] Preferably, the flight positioning module includes an inertial measurement unit, a lidar and a vision recognizer. The flight positioning module performs joint calculation on 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 vision recognizer through a fusion algorithm to obtain the three-dimensional position and attitude of the UAV.

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

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

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

[0013] Preferably, the cleaning and quality inspection module sequentially includes a plasma generating device, an acoustic cavitation transducer device and an ionic air curtain device. The cleaning and quality inspection module further includes a model prediction controller. The model prediction controller generates a power trajectory under a preset energy constraint and drives the plasma generating device, the acoustic cavitation transducer device and the ionic air curtain device.

[0014] Preferably, the model prediction 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.

[0015] Preferably, the energy module includes 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 battery power before and after replacement and outputs energy settlement data.

[0016] Preferably, after generating the job closed-loop data packet, the energy management unit calculates the hash value of the job closed-loop data packet and attaches a digital signature, writes it into the ledger through a blockchain node, and then sends it to the cloud through a network communication interface. The self-learning model uses an 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 completing the training.

[0017] A non-road-blocking autonomous cleaning method for main road cameras based on drones, which is used to execute the non-road-blocking autonomous cleaning system for main road cameras based on drones. The method includes the following steps:

[0018] Read flight instruction data and simultaneous localization and mapping data, drive the drone to fly to the target camera and establish a fixed posture;

[0019] 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 according to the stain spectrum fingerprint data;

[0020] Execute plasma discharge treatment, acoustic cavitation stripping treatment, and ionic air curtain purging treatment in sequence under the power trajectory defined by the model predictive controller according to the cleaning parameters, and obtain cleaning progress field data and quality inspection residual index data; when the quality inspection residual index data is greater than the threshold, adjust the cleaning parameters and repeat the execution of the treatment 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, control the drone to return to the edge base station, and encapsulate the cleaning progress field data, quality inspection residual index data, and energy settlement data into a job closed-loop data packet for the self-learning model to adjust parameters and generate the next flight instruction data.

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

[0023] Through the multi-modal spectrum fingerprint fusion perception technology means, the present invention realizes highly robust recognition of stain types and thicknesses, and overcomes the problem of missed detection in the visual single modality.

[0024] Through the negative pressure adsorption soft docking ring and model predictive control means, the present invention realizes the dry composite cleaning effect of plasma-acoustic cavitation-ionic air curtain under a single hover, avoids repeated photographing, reduces water use, and takes into account the removal of stubborn oil films.

[0025] Through the energy management-blockchain closed-loop means, the present invention realizes the transparency of job energy consumption and cloud incremental self-learning optimization, and continuously reduces the energy and time costs of a single task.

[0026] Through the automatic battery replacement technology of the edge base station, the present invention realizes second-level battery replacement and all-weather non-stop operation, and solves the endurance bottleneck of long-link inspection. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

[0030] Figure 4 It is a schematic diagram of the structure of the pan-tilt load in the specific embodiment of the present invention;

[0031] Figure 5 It is a schematic diagram of the structure of the airborne device in the specific embodiment of the present invention Figure 1 ;

[0032] Figure 6 It is a schematic diagram of the structure of the airborne device in the specific embodiment of the present invention Figure 2 ;

[0033] Figure 7 It is a schematic diagram of the structure of the airborne device in the specific embodiment of the present invention Figure 3 ;

[0034] Figure 8 It is a schematic diagram of the mechanism of the three-axis pan-tilt in the specific embodiment of the present invention;

[0035] Reference numerals: 1, drone fixing bracket; 2, water pump suction port; 3, injection port; 4, liquid storage cavity; 5, liquid storage barrel; 6, positioning groove; 7, pan-tilt load device; 8, camera fixing seat; 9, camera data line perforation; 10, fixing seat; 11, algorithm core board; 12, water pipe guiding hole; 13, adapter board card slot; 14, clamping point; 15, nozzle; 16, flight airborne device; 17, flight control core board; 18, micro 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 pan-tilt; 28, pan-tilt load fixing seat; 29, pan-tilt interface. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, numerous specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments can also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present disclosure.

[0037] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described 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 should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0039] As Figure 1 - Figure 2 shown, an autonomous cleaning system for main road cameras based on drones includes:

[0040] A flight positioning module, configured to read flight instruction data and simultaneous localization and mapping data, and drive the drone to fly to the target camera;

[0041] In the present invention, the drone must complete non-intrusive approach and attachment above the main road with continuous traffic flow. Therefore, the flight positioning module undertakes two core tasks: one is to provide the drone with continuous and reliable spatial coordinates and attitude angles; the other is to map these coordinates and attitudes in real time to a safe flight path that does not interfere with traffic and simultaneously meets the geometric requirements of negative pressure adsorption. The following will be elaborated from four aspects: sensor cooperation, state inference, path planning, and attitude locking.

[0042] In the urban road environment, a single sensor is often distorted by factors such as occlusion, rain and fog, and metal structure reflection. The combination of an inertial measurement unit, lidar, and machine vision is selected for this module, and the reasons and division of labor are as follows:

[0043] Inertial measurement unit: The internal three-axis accelerometer and gyroscope can give the body acceleration and angular velocity in milliseconds. Even if the external signal is completely interrupted, the body pose can be predicted short-term. It is a "non-dependent" but drift-prone benchmark.

[0044] LiDAR: The rotating laser is most sensitive to the reflection of columnar targets such as roadside lamp posts, billboards, and height limit gantries, and can still output accurate distance and azimuth at night without texture and under low light conditions.

[0045] Machine vision: When the lighting is good, the camera recognizes the reflective markers installed on the edge of the camera, and can resolve the heading angle and relative distance at the sub-pixel level, providing scale calibration for the LiDAR.

[0046] This "triangle" perception system is complementary under various working conditions such as sunny days, rainy days, and dim street lights: lasers and vision alternately become 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 these three types of observations. The module uses the extended Kalman filter (EKF), a recursive algorithm. Its core idea is: using the high-speed integration result of the inertial measurement unit as a short-term prediction, and using the low-frequency measurements of LiDAR and vision ranging to correct the deviation, continuously correcting the 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 at the previous moment; then, the radar point cloud is compared with the road three-dimensional map to give the rough position of the aircraft in the road coordinate system; and then the visual marker ranging provides fine alignment. The filter weighs the uncertainties of the predicted value and the observed value through the gain matrix, making the output attitude angle and position coordinates smooth numerically and able to quickly capture real changes. In this way, even if GPS multipath or occlusion makes the satellite signal unavailable, the flight positioning module can still maintain a relative accuracy of centimeter or even millimeter level.

[0049] After obtaining a high-confidence aircraft state, it is also necessary to plan a feasible flight path for the UAV that does not disrupt traffic. The module completes this task with a global-local two-level 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 road three-dimensional grid map stored locally on the UAV, requiring that the flight altitude always be higher than the lane and laterally leave the safety bandwidth of the street lamp post.

[0051] At the local level, the LiDAR scans the moving vehicles, high trucks or construction platforms above the lane in real time. When the distance between any detected temporary obstacle and the planned flight path is less than the set safety threshold, the local replanner regenerates a detour route within dozens of milliseconds and then smoothly connects it to the original flight path.

[0052] Through this "two-layer" strategy, the UAV will neither have a collision risk due to suddenly emerging vehicles in the air nor can it ensure an ideal incident angle to enter the camera adsorption area globally.

[0053] After the drone arrives in front of the camera, its flight speed is gradually converged to about 0.1 m / s. Subsequently, the module switches the control right to the quaternion attitude controller: the speed outer loop suppresses the horizontal and vertical speed errors to an extremely low level; the attitude inner loop takes the quaternion error as the target and generates the motor angular velocity command to make the nose direction perpendicular to the normal of the camera mirror surface completely.

[0054] When the position deviation and the attitude error simultaneously enter the millimeter-level and angular component-level thresholds, the vacuum pump starts, and the soft adsorption ring forms a pressure difference to gently "attach" the drone to the outside of the camera shield. Thus, the ultimate goal of the flight positioning task is completed: achieving a millimeter-level docking attitude without disturbing the main road traffic flow.

[0055] Preferably, the flight positioning module includes an inertial measurement unit, a lidar, and a visual recognizer. The flight positioning module performs joint resolution on the acceleration and angular velocity output by the inertial measurement unit, the point cloud coordinates output by the lidar, and the attitude angles output by the visual recognizer through a fusion algorithm to obtain the three-dimensional position and attitude of the drone.

[0056] This module plays the role of a "space reference generator" in the entire drone camera cleaning system: it assimilates the original observations scattered in different coordinate systems, different frequencies, and different noise distributions into the three-dimensional position information (x, y, z) and attitude angles (φ, θ, ψ) in the same moment and the same geographical 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. The a b , ω b output by the airborne inertial measurement unit is located in the body coordinate system B. The lidar point cloud is originally located in the radar coordinate system L. The visual recognizer obtains the pose of the camera coordinate system C relative to the target marker through Perspective-n-Point (PnP).

[0058] In the deployment stage, the external parameter matrices T LB , T CB are obtained for B→L and B→C respectively. During operation, the lidar point cloud and the visual measurement are unified into the body coordinate system through the external parameter transformation, and then pushed to the geographical coordinate system E through the rotation matrix R BE . In this way, all measurements can be directly compared numerically without cross-coordinate system interpolation.

[0059] The sampling timestamps of different sensors are first written into the IMU clock domain through the hardware timestamp, and then linearly interpolated at the software layer; the error is controlled within the millisecond level to avoid the Extended Kalman Filter (EKF) mixing "old attitudes" with "new positions" and causing false vibrations.

[0060] Probability integration is the essence of the extended Kalman filter. Inertial navigation provides short-term high-frequency predictions but accumulates biases; lidar / vision provides low-frequency absolute measurements but has its own Gaussian noise. EKF uses linear approximation to pull the non-linear state-observation model back into the Gaussian framework. The core equation is:

[0061]

[0062] where represents the prior (inertial navigation prediction); z t represents the measurement (lidar / vision); K t represents the Kalman gain - a quantitative weight that measures "whom to trust".

[0063] If the visual signal is distorted (night glare), its observation covariance R vision increases, and K t automatically reduces the visual weight, allowing lidar to dominate; vice versa. This adaptive trade-off enables the system to continue operating even when a sensor fails, rather than simply "exiting when the signal is lost".

[0064] The bias estimates converge synchronously. The biases b a , b g are included as state variables in the filtering - meaning that EKF estimates not only the attitude but also the health of the IMU itself. As the visual-lidar ground truth is continuously injected, the bias covariance gradually converges, and the inertial navigation drift is restricted to a small range.

[0065] Control integration transforms the solution results into a safe trajectory and attitude. The design of the trajectory cost function: Global planning is not simply the shortest path but rather the minimization of the comprehensive cost:

[0066]

[0067] where J represents the global planning cost function. The smaller the sum, the better the trajectory; α represents the line length weight; β represents the altitude penalty, keeping the trajectory always above the safe altitude; γ represents the lateral penalty, restricting being too close to the street lamp pole; ||p k - p k-1 || is the Euclidean norm, representing the spatial distance between two adjacent nodes; p k represents the three-dimensional position vector (x k , y k , z k ) of the k-th trajectory node in the geographic coordinate system; p k-1 represents the three-dimensional position vector of the (k - 1)-th trajectory node; f height (p k ) represents the altitude penalty function, which returns a positive value when the altitude z k of the node p k exceeds the set safe altitude, otherwise it returns zero; flateral (p k ) represents the horizontal penalty function, when the node p k When the lateral offset exceeds the safety bandwidth, a positive value is returned; otherwise, zero is returned. This function allows a numerical trade-off between "safety" and "economy" rather than relying on empirical thresholds.

[0068] The attitude quaternion closed loop, the classic Euler angle will be singular around ±90°. The module uses quaternion error instead:

[0069]

[0070] where q d is the expected posture, q c is the current posture; is the quaternion multiplication. The attitude controller uses q err Generate angular velocity commands to avoid gimbal lock.

[0071] In one embodiment, the drone is fixed by using a soft ring adsorption. The soft ring adsorption logic is that when the position error ∈ p and posture error ∈ θ At the same time, the vacuum pump is triggered when the pressure drops below the threshold value, generating an adsorption force F with a pressure difference ΔP. adh =ΔP·A. At this point, the control system freezes the external position loop, retaining only fine-tuning of the attitude to avoid secondary displacement caused by pumping vibration. It should be noted that the method of fixing the drone body for camera cleaning in this invention is limited to soft ring adsorption. Any other fixing method based on three-dimensional coordinates can also be used. In some special cases or needs, the drone can also be used for cleaning in a hovering manner.

[0072] Environmental adaptation and fault tolerance mechanism: if the visual recognizer loses its mark for N consecutive frames, the weight scheduler will directly reduce its weight to zero, but retain the visual thread and do not interrupt the capture opportunity. If the laser radar has scattering noise due to water mist, the system detects that the point cloud density decreases and automatically increases R LiDARR ; When the density recovers, it will gradually return. max Within seconds, the aircraft is instructed to automatically climb to a preset altitude to avoid the vehicle, then descend to the work surface after observation is restored. This layered fault tolerance avoids "single point failure" and is essential for uninterrupted traffic operations on main roads.

[0073] Using this solution, in an urban canyon test section, based on true value comparison with a total station, the positioning root mean square error (RMS) was ≤1mm, and the attitude root mean square error (RMS) was ≤0.2°. Local obstacle avoidance replanning took 20ms, and the attitude loop closed-loop latency was <3ms, sufficient to suppress disturbances in a lightweight multirotor in 5m / s gusts. The docking success rate exceeded 98% in all three scenarios: sunny, nighttime, and drizzling.

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

[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 instruction data.

[0076] The path planning unit in the flight positioning module takes the idea of a continuous potential field as the core: first, maps all risk factors in the road environment into potential energy densities related to spatial coordinates, then searches for the three-dimensional trajectory with the minimum cost in this potential field, and finally smooths the trajectory and writes it into the flight instruction data. Its principle includes the following:

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

[0078]

[0079] where β represents the altitude penalty weight, h c represents the safe altitude threshold, z represents the current altitude; γ represents the lateral penalty weight, y t represents the lateral coordinate of the center line of the street lamp pole, w s represents the lateral safety bandwidth, y represents the current lateral coordinate; κ j represents the repulsive force coefficient of the jth dynamic obstacle, p j represents the center position of the obstacle, σ represents the equivalent radius of the obstacle; 1(·) is an indicator function used to activate the penalty when exceeding the limit.

[0080] Secondly, define the comprehensive cost. Discretize the trajectory into a node sequence p0, p1,..., p N , and minimize the following formula:

[0081]

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

[0083] Finally, perform dynamic consistency shaping. Use a fifth-degree polynomial for each segment of the trajectory:

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

[0085] Perform smoothing and determine the coefficient a with the minimum third - derivative energy as the target i . This can eliminate the sharp corners of the broken line and ensure that the multi - rotor controller will not generate excessive angular acceleration during the tracking process. The shaped time - displacement sequence is written into the flight instruction data for real - time execution by the attitude control loop. The entire process provides an obstacle - avoidance flight path that takes into account safety distance, energy economy, and dynamic smoothing, laying a trajectory foundation for the drone to achieve non - road - closing attachment in the main road environment.

[0086] A stain detection and decision - making module, which 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 the mirror stain and give a one - time executable dry - cleaning parameter combination without lifting the shield and without interrupting traffic. Its mechanism can be decomposed into four levels: "multi - physical quantity acquisition → spectrum fingerprint construction → graph convolution characterization → 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, resulting in a depression in the reflection coefficient at 24 GHz;

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

[0091] Acoustic echo: 40 kHz ultrasonic pulses are scattered at the air - stain film interface, and the rough surface will broaden the main peak frequency band.

[0092] The three - channel complementarity enables the system to obtain separable features in both day - night and rain - fog environments.

[0093] Spectrum fingerprint construction, all original observables are normalized and then spliced 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] where Γ​​24 , Γ 25 represents the amplitude of the reflection coefficient at 24 GHz and 25 GHz; S ij represents the Mueller matrix element; R a represents the ultrasonic roughness index; f ac represents the acoustic main peak frequency; I UV , I VIS represents the fluorescence intensity.

[0096] To make full use of the physical coupling relationship between features, project F onto a ten-node complete graph. The edge weights adopt Gaussian similarity:

[0097]

[0098] where, w ij represents the similarity between node i and node j; f i represents the o-th dimensional eigenvalue; σ represents the width factor. After two-layer graph convolution, the fingerprint embedding vector h is obtained, and its dimension is much lower than the original features, which is convenient for subsequent optimization.

[0099] Clean the parameter search, and denote the parameters to be optimized as:

[0100] p = [N p , τ s , τ f

[0101] where, N p represents the number of plasma pulses; τ s represents the duration of acoustic cavitation frequency sweep; τ f represents the duration of ion wind curtain purging.

[0102] Bayesian optimization minimizes the objective function with the expected improvement criterion:

[0103]

[0104] where, represents the comprehensive cost; R(h, p) represents the residual index predicted by the model; λ represents the energy consumption penalty weight; E(p) represents the energy estimation function. After thirty iterations, it converges to the cleaning parameter p * .

[0105] Through the present invention, the three types of scenarios of oil film, dust, and water mist show partitionable clustering in the ten-dimensional feature space, and the mean square error of the predicted residual index is maintained at the order of 10 -3 ; the parameter search convergence time is less than 0.3 s, which can be parallel with the flight attachment action; after cleaning, resample and update the true residual index, and use it as a supervised signal to incrementally 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 the millimeter wave reflection coefficient, polarization scattering matrix and acoustic echo into a stain spectrum fingerprint. The fingerprint optimization unit outputs a cleaning parameter set according to the stain spectrum fingerprint.

[0107] In unblocked main roads, oil film, dust, and water mist on the camera housing exhibit different electromagnetic, optical, and acoustic properties. The module first uses the millimeter-wave probe, polarization camera, and ultrasonic transducer to select their respective "sensitive areas," interweaving these three physical quantities into a contamination spectrum fingerprint. Then, through Bayesian search, a mapping between the fingerprint and cleaning parameters is established: "minimum residue - minimum energy consumption."

[0108] Multimodal signal complementation includes:

[0109] For millimeter wave reflection, the oil film changes the dielectric constant of the shield surface, and the reflection depth at 24 GHz is monotonically proportional to the thickness.

[0110] Polarization imaging, multilayer interfaces weaken the linear polarization retention, and the non-diagonal elements of the Mueller matrix fluctuate significantly with film thickness.

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

[0112] Each of the three has blind spots: vision failure at night, laser attenuation due to rain and fog, and oil film covering up acoustic texture; integrating them can make the blind spots complement each other.

[0113] The module compresses the raw data into a fingerprint. The module normalizes and sorts the three raw data, extracts the ten most discriminative components, and then uses the complete graph combined with Gaussian edge weights to encode the "who is related to whom" information into the graph structure. Two layers of graph convolution absorb the noisy raw components into a more robust embedding vector h.

[0114] The Gaussian edge weight formula and the meaning 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 ]; "cleanliness" is measured by the residual index prediction R(h,p), and "energy saving" is measured by the energy consumption estimate E(p). The two are linearly combined to form the comprehensive cost:

[0116]

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

[0118] Bayesian optimization Treat it as a "black box" function, use Gaussian process to estimate its shape, and then use expected improvement to select the next set of p. After 30 samplings, p can be found. * ——The most cost-effective number and duration of pulses under the current stain fingerprint.

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

[0120] In this application, even if one channel is completely distorted, graph convolution can still accurately predict parameters based on the remaining channels. A complete search is completed within 0.3 seconds, and parameters are already available before the attachment process is complete. Each real-world quality inspection reduces prediction errors, and over a long period of time, energy consumption can be reduced to less than half that of traditional water washing.

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

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

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

[0124] 2. How many plasma pulses N are needed under the current remaining power and energy consumption budget? p , sweep length acoustic cavitation τ s , blow ion wind 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 near 24 GHz, and therefore react most strongly to oil film thickness; polarization imaging focuses on the "depolarization" effect of the three-layer interface of cover glass, oil film, and air on linear polarization, and can quantify film thickness and distribution uniformity; acoustic echoes at 40 kHz reflect the microscopic roughness of the surface; the larger the dust particles, the wider the main peak.

[0126] If any of the three signals degrades (such as visual overexposure at night or laser absorption caused by rain and fog), the other two signals can still remain separable, allowing the system to construct a stable stain spectrum fingerprint in all weather and full light conditions.

[0127] The low-dimensional embedding vector h (obtained from 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; thus, the cost function uses the expectation of the residual index weighted with the energy consumption E, and the penalty coefficient λ comes from the energy management module.

[0128] Bayesian optimization treats the entire cost surface as a black box and performs posterior estimation on it using a Gaussian process; then it selects the next set of cleaning parameters using the expected improvement (EI) criterion to approach the optimal combination p at the fastest convergence rate * .

[0129]

[0130] p * represents the set of optimal cleaning parameters; p = [N p , τ s , τ f , where N p represents the number of plasma pulses, τ s represents the duration of acoustic cavitation sweep frequency, and τ f represents the duration of ion wind curtain purging; represents the expected value of the residual index predicted using the fingerprint embedding h and the 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 increases monotonically with the number of pulses and the driving duration.

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

[0132] This set of "fingerprint - optimization" chain enables the drone to give a one-time, energy - aware dry cleaning solution for each camera stain without manual judgment; it is measured that in extreme cases such as an oil film of 60μm, a dust roughness of 30μm, and a night illumination of 5lx, the residual index can still be pressed below the threshold, and the energy consumption can be controlled within half of 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 limited by the model prediction controller, obtain real - time cleaning progress field data and quality inspection residual index data, 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 the attachment, the cleaning system must distribute the energy of the dry three-stage process (plasma, acoustic cavitation, ion air curtain) to the most needed positions within a single dwell time window without closing traffic, and prove that the mirror surface is below the residual threshold. This module consists of four internal closed-loop series: power trajectory planning, energy-position coupling execution, residual index quantification, and parameter adaptive update. The core physical-control logic is described layer by layer in text below, and only the calculation formulas that have not appeared before and are indispensable are given.

[0135] Principle of power trajectory planning. The controller takes the current remaining energy E res and the cleaning parameters p = [N p , τ s , τ f as hard constraints, and predicts the future H discrete steps. Generate the discharge power sequence P0, P1,..., P H . The MPC optimization goal is to push the residual index below the threshold with as little energy as possible. Therefore, the following objective function is constructed:

[0136]

[0137] P k represents the discharge power at the k-th step; E k = P k Δt represents the energy consumption at this step; C k represents the contribution of this step's power to the predicted residual index; μ represents the residual penalty weight; Δt represents the discrete time step. Constraint conditions:

[0138] 0 ≤ P k ≤ P max

[0139]

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

[0141] Energy-space coupling execution. The three-stage device works under the drive of P in , and 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 projected energy at the coordinates (x, y) up to the moment t; ρ(x, y, τ) represents the instantaneous energy density; τ represents the integration time variable. The progress field can be imaged into a heat map in real time to detect local dead corners.

[0144] The residual index metric, after each three-stage process cycle is completed, the system re-acquires the stain spectrum fingerprint F new . Using the clean reference fingerprint F ref as a reference, calculate the relative residual index:

[0145]

[0146] η represents the residual index; ||·||2 represents the two-norm. When η ≤ η max (η max is the set threshold), it is determined that the mirror is qualified; otherwise, enter the parameter adaptation stage.

[0147] For parameter adaptive update, if η > η max , the module fine-tunes the cleaning parameters using gradient correction:

[0148]

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

[0150] The learning rate decays exponentially:

[0151] α m = α0e -δm

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

[0153] Through the present invention, in the measured scenario of a common oil film thickness of 60 μm, η can be reduced from 0.12 to 0.04 after two rounds of iteration, which is lower than the threshold of 0.05. The power trajectory optimization in cooperation with the local progress field imaging enables the decontamination efficiency per joule to be increased by about 35% compared with the constant power-constant duration scheme. There is no droplet splashing and no high-brightness visible light throughout the process, which does not affect the traffic flow and the driver's line of sight.

[0154] Preferably, the cleaning and quality inspection module sequentially includes a plasma generating device, an acoustic cavitation transducer device, and an ionic air curtain device. The cleaning and quality inspection module further includes a model predictive controller, and the model predictive controller generates a power trajectory within a preset energy constraint and drives the plasma generating device, the acoustic cavitation transducer device, and the ionic air curtain device.

[0155] After the drone is attached, the module must optimize the decontamination process using three dry methods—plasma, acoustic cavitation, and ion air curtain—in the optimal order and at the optimal power within a fixed energy budget. This process also monitors residual energy and performs secondary enhancements as needed. The core concept is to first use a model predictive controller (MPC) to break down the residual electrochemical energy into a power-time trajectory, which is then consumed sequentially by the three devices. If the real-time residual energy still exceeds a threshold, rescheduling is performed, assuming energy consumption remains high.

[0156] Energy distribution - power-duration coupling, let the remaining energy E res (E res Represents the total energy available in the current battery and supercapacitor), which MPC divides into three energy quotas:

[0157] E1+E2+E3=E res

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

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

[0160] The energy efficiency weight of the stage is calculated based on the spectrum fingerprint given by the stain detection module, and the system queries the empirical coefficient ξ i , represents the efficiency of removing the current fouling film per unit joule at stage i. MPC solves the linear programming with the maximum residual reduction as the goal:

[0161]

[0162] stE1+E2+E3=E res

[0163]

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

[0165] Obtain Then, divide it by the discharge voltage U i The total discharge charge of each section is obtained, which is used as the upper limit of the pulse number, sweep frequency or purge time.

[0166] The local energy density is closed-loop. When the three devices are working, the instantaneous energy density ρ(x,y,t) on the mirror coordinate (x,y) is continuously integrated into the progress field Φ(x,y,t). imgRedraw the heat map in seconds and calculate the local energy difference:

[0167]

[0168] Φ ref represents the empirical sufficient energy threshold; [z] + represents max(z,0). If ΔΦ min >0, it indicates that there are still dead corners not cleaned properly, and local acoustic cavitation frequency sweeping needs to be added. The additional energy is

[0169] For the secondary determination of the residue index, after the cleaning cycle ends, according to:

[0170]

[0171] Calculate the residue index. If η > η max , the MPC takes the remaining energy balance of the previous round ΔE = E res - as the new E res and re-enters the initial step until η ≤ η max or the energy is exhausted.

[0172] Through the present invention, the stage energy efficiency weight ξ i makes the energy flow preferentially to the device with the highest decontamination efficiency; the progress field closed-loop allows any local energy gap to be compensated by secondary frequency sweeping; within the 60s residence window, one or two rounds of iteration can meet the residue threshold without triggering energy overdraw.

[0173] Preferably, the model predictive controller updates the power trajectory according to the cleaning progress field data and the quality inspection residue index data, and sends the updated power trajectory to the plasma generating device and the acoustic cavitation transducer device.

[0174] After the first power trajectory is issued, the three-stage dry process starts to act on the mirror surface. The data bus at the drone end continuously transmits two types of feedback: the cleaning progress field Φ(x,y,t) represents the spatial-temporal distribution of the cumulative energy; the quality inspection residue index η(t) measures the relative error of the mirror surface from the target cleanliness.

[0175] The controller regards the two as "state measurements", re-solves the small-scale MPC at a frequency of ten hertz, and forms a self-tuning power trajectory. This closed-loop mainly involves two coupling mechanisms.

[0176] Gain injection of error potential energy. For any moment t, define the local energy gap on the mirror coordinates (x,y):

[0177] ΔΦ(x,y,t) = [Φ ref - Φ(x,y,t)] +

[0178] Φ ref represents the energy threshold required for the experience, [·] + is a positive truncation function. Map the overall gap to the power adjustment amount:

[0179]

[0180] where k Φ represents the progress-power coupling gain, and A represents the effective area of the mirror surface.

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

[0182]

[0183] k p represents the residual proportional gain, and k i represents the residual integral gain, and τ is the integral time window.[[ID=3o]]

[0184] Online correction of the power trajectory. For the next control step k, the original trajectory power is corrected to:

[0185]

[0186] and is immediately sent to the plasma generating device and the sonochemical cavitation transducer; if exceeds the safety upper limit P max , then take P max . This real-time superposition will not damage the original MPC optimization structure, but can make the power quickly compensate for local dead corners and global residues.

[0187] Through the present invention, the mirror surface area with a large progress field gap automatically obtains additional sonochemical cavitation energy, avoiding "cleaning zebra stripes". The proportional-integral regulation presses the residual index into the threshold within 1-2 rounds and will not cause power oscillation due to noise. All power correction amounts are dynamically deducted within the remaining energy budget of the previous moment, and will never trigger battery under-voltage.

[0188] Through the above "local gap-global residue" dual-channel gain injection, the model predictive controller can autonomously update the power trajectory and accurately drive the plasma and sonochemical cavitation devices under the time pressure of continuous traffic flow, ensuring a quantifiable clean result in one attachment.

[0189] The energy module is used to control the UAV to return after the quality inspection residual index data does not exceed the threshold, and encapsulate the cleaning progress field data, quality inspection residual index data and energy settlement data into an operation closed-loop data packet to adjust the own parameters of the self-learning model, and generate the next flight instruction data.

[0190] After the mirror residual index η has met the threshold η max This module takes over the control right and completes three key tasks: ① Determine whether the remaining energy is sufficient for a safe return; ② Package the energy-mass information of the entire process of one operation into a closed-loop data packet; ③ Perform incremental update on the self-learning model based on this data packet and output the flight instruction data for the next flight. The following explains the core logic layer by layer and gives the necessary calculation formulas for the first appearance.

[0191] Return energy determination and scheduling. After the UAV stays for cleaning and is still hanging outside the camera, first confirm its remaining battery power E rem Can it cover the whole process of return climb, cruise, landing and battery replacement at the edge base station. Define the minimum energy required for return:

[0192] E ret = mgh loss + κd home

[0193] E ret represents the return energy requirement; m represents the total mass of the UAV; g represents the acceleration due to gravity; h loss represents the equivalent height difference for re-climbing after descending; d home represents the horizontal distance to the base station; κ represents the energy consumption coefficient per unit distance of cruising.

[0194] When E rem ≥ E ret , the energy module issues a return instruction; otherwise, immediately enter the "low-power hover and request ground assistance" mode.

[0195] Construction of the operation closed-loop data packet. After the return navigation is started, the module encapsulates four types of data: the cleaning progress field grid Φ(x,y,t end ); the final residual index η end ; the energy settlement triple [E use , E rem , E ret , where E use represents the energy consumed in this round of cleaning; timestamp and pose trajectory hash (to ensure backtracking consistency). The data packet is written into the local linked list after internal salt hashing and then asynchronously uploaded to the cloud blockchain to prevent tampering.

[0196] The self-learning model is incrementally updated. The cloud self-learning model adopts the Proximal Policy Optimization (PPO) framework, and uses the energy-quality data of a single-round operation as the "immediate reward" to continuously fine-tune the policy parameter θ. The reward function defined in the present invention is:

[0197]

[0198] R represents the operation reward value; w1, w2, w3 represent three weight coefficients; η end represents the final residual index; E use , E rem has the same meaning as above; T cycle represents the total duration of this round of operation; T max represents the maximum allowed residence time.

[0199] Gradient update rule:

[0200]

[0201] β represents the learning rate; represents the expectation of the reward under the current policy.

[0202] The updated policy immediately infers the flight instruction data for the next flight (takeoff point, target camera sequence, energy budget, re-tuned 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 replaces the battery after the drone returns. The energy management unit records the battery power before and after replacement and outputs energy settlement data.

[0204] After the drone completes camera cleaning and confirms that the residual index η ≤ η max , it then enters the return flight - battery swapping - energy settlement process. The energy module consists of two parts in the ground edge base station: an automatic battery swapping mechanism and an energy management unit. The two cooperate to ensure the uninterrupted cyclic operation of the fleet and provide real energy consumption labels for the cloud self-learning model.

[0205] The mechanical-electrical coupling idea of the automatic battery swapping mechanism. Mechanical positioning: The base station gripper relies on the V-shaped guide groove to align the drone battery compartment at the sub-millimeter level, avoiding insertion and extraction skew under wind load or ground vibration.

[0206] Electrical isolation and thermal management. When plugging and unplugging, first cut off the DC total negative pole, then the positive pole, and finally release the signal PIN; the order is reversed when installing a new battery. The plug is embedded with a heat pipe and a phase change material, allowing the residual heat after high-rate discharge to be quickly exported.

[0207] The battery swapping cycle takes less than 8 seconds to complete a set of operations of unplugging - plugging - locking. The supercapacitor maintains the power supply for the avionics during the switching interval, and the on - board controller does not need to be restarted.

[0208] Measurement and settlement of the energy management unit. Power measurement: Before and after battery swapping, read the ampere - hour integrator and terminal voltage of the battery pack to obtain Q pre ,U pre and Q post ,U post , where Q pre is the remaining charge amount (Ah) before battery swapping, and U pre is the corresponding terminal voltage; Q post ,U post are the starting values of the new battery. Energy settlement:

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

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

[0211]

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

[0213] Closed - loop data packet. The energy management unit encapsulates [E use ,SOC pre ,SOC post ,Φ(x,y,t end ),η end into a "job closed - loop data packet", where Φ(x,y,t end ) is the final cleaning progress field, and η end is the final residue index. The data packet is timestamped and task - ID - added and then saved on the blockchain.

[0214] The incremental update of the cloud strategy of the self - learning model uses energy consumption - quality data as immediate rewards:

[0215]

[0216] w1, w2, w3: weight coefficients; E rem : remaining energy after returning; T cycle : total duration of this round; T max : allowed maximum residence time; The policy parameter θ is as follows:

[0217]

[0218] (β is the learning rate) for fine-tuning and generating the flight instruction data for the next flight mission.

[0219] With the present invention, 8-second battery swapping ensures seamless connection of "landing - swapping - taking off" of drones during peak hours. The energy consumption traceable dual metering eliminates temperature drift error, and the closed-loop data supports later auditing. The self-evolving strategy continuously optimizes the energy benchmark with the real energy consumption - quality curve, and reduces energy consumption batch by batch under the same cleanliness level.

[0220] The energy module enables metered energy saving and rolling self-optimization for the unmanned cleaning operation of the main road cameras through three mechanisms of "fast battery swapping + precise energy settlement + data-driven strategy update".

[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 a blockchain node, and then sends it to the cloud through a network communication interface. The self-learning model uses the incremental gradient boosting decision tree to perform online training on the received operation closed-loop data packet, and returns the next flight instruction data after completing the training.

[0222] After the end of a cleaning operation, the energy module quickly uploads the "energy - quality" closed-loop data to the blockchain and pushes it to the cloud. The cloud uses the incremental gradient boosting decision tree (GBDT) for online learning, completes the self-evolution of the strategy within seconds, and sends back the flight instruction for the next flight mission. The link has three core links: trusted encapsulation, ledger writing, and online training.

[0223] Trusted encapsulation: After generating the data packet pkg (including 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 private key of the base station. In this way, any backend node can verify the signature with the public key to ensure that the data cannot be forged.

[0224] H = SHA256(pkg)

[0225]

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

[0227] Incremental GBDT online training: After the cloud verifies the signature, it extracts the feature vectors:

[0228] z = [E use , η end , T cycle , …]

[0229] And use it as a new sample to flow into the incremental GBDT. The model performs gradient iteration on the target variable (the theoretical energy consumption of the next task):

[0230] 1. Residual calculation:

[0231]

[0232] 2. Prediction update:

[0233]

[0234] denotes the residual of the i-th sample when the m-th incremental tree is represented; y i denotes the true target energy consumption of the i-th sample; denotes the predicted value of the first m - 1 trees for the i-th sample; denotes the new prediction after adding a new tree; v denotes the learning rate; T (m) (z i ) denotes the output of the m-th new regression tree on the sample z i ; z i denotes the feature vector of the i-th sample. Incremental update does not require replaying historical data and can complete a model refresh within a few seconds.

[0235] The flight instruction is issued, and the updated model combines the distribution of the camera cluster and the remaining energy of the aircraft fleet to generate the take-off base station and target camera list for the next flight and the estimated energy budget The initial cleaning parameter vector p0. The instruction is transmitted back to the standby drone through the MQTT bus to achieve seamless rotation.

[0236] Such as Figure 3 - Figure 8 shown, the applications of the present invention on the drone body include:

[0237] A drone with a flight airborne device (16) and a three-axis gimbal (27). The drone is connected with a three-axis gimbal (27) through a gimbal interface (29). A gimbal load device (7) is provided on the three-axis gimbal (27). The gimbal load device (7) includes a nozzle (15), a micro camera, and 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 instruction data and synchronous positioning - map construction data and drive the drone to fly to the target camera. The stain detection and decision module is used to collect millimeter wave reflection coefficients, polarization scattering matrices, and acoustic echoes and output cleaning parameters. The cleaning and quality inspection module sequentially performs plasma, acoustic cavitation, and ionic air curtain cleaning under the power trajectory limited by the model prediction controller and outputs the cleaning progress field and the quality inspection residual index in real time;

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

[0239] The flight onboard equipment (16) is fixed to the upper end of the drone by screws. The flight onboard 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 instructions 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 replacement mechanism and an energy management unit. After the quality inspection residue index does not exceed the threshold, it controls the drone to return and complete battery replacement, and at the same time records the battery power before and after replacement and encapsulates it into an operation closed-loop data packet to update the self-learning model and generate the next flight instruction data.

[0240] The drone is equipped with a drone fixing bracket (1), and the drone fixing bracket (1) is provided with a positioning groove (6). The liquid storage bucket (5) is installed on both sides of the drone fixing 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 bucket (5) is provided with a liquid storage cavity (4). The liquid storage bucket (5) is provided with an injection port (3) for injecting liquid and a water pump suction port (2) connected to the 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 equipment (7) is provided with a fixing seat (10) and a camera fixing seat (8) for installing a micro camera. One side of the camera fixing seat (8) is provided with a camera data line perforation (9). One side of the gimbal load equipment (7) is provided with a water pipe guiding hole (12). The gimbal load equipment (7) is provided with a transfer board card slot (13) and a clamping point (14) for inserting an eport interface transfer board (22). The gimbal load equipment (7) is provided with a nozzle installation hole for installing the nozzle (15). The outlet of the nozzle (15) is coaxially integrated with a plasma electrode and an ion air curtain nozzle for realizing 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). There is a connection hole (26) outside the flight airborne equipment (16). The opening timing of the electromagnetic three-way valve (24) is scheduled in real time by the model predictive controller in combination with the cleaning progress field data.

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

[0245] The algorithm core board (11) is used to control the flight of the drone and the angle of the pan-tilt. The algorithm core board (11) is electrically connected to the flight control core board (17) through a circuit, and reads the power trajectory generated by the model predictive controller through a high-speed interface to update the working current of the plasma generating device and the acoustic cavitation transducer device 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 a circuit. 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 consumption hover mode.

[0247] A non-road-blocking autonomous cleaning method for main road cameras based on drones, which is used to execute the non-road-blocking autonomous cleaning system for main road cameras based on drones. The method includes the following steps:

[0248] Read the flight instruction data and the simultaneous localization and mapping data, drive the drone to fly to the target camera and establish a fixed attitude;

[0249] 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 according to the stain spectrum fingerprint data;

[0250] Execute plasma discharge treatment, acoustic cavitation stripping treatment and ion air curtain purging treatment in sequence according to the cleaning parameters under the power trajectory defined by the model predictive controller, and obtain cleaning progress field data and quality inspection residue index data; when the quality inspection residue index data is greater than the threshold, adjust the cleaning parameters and repeat the execution of the treatment until the quality inspection residue 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, control the UAV to return to the edge base station, and encapsulate the cleaning progress field data, quality inspection residual index data, and energy settlement data into an operation closed-loop data packet for the self-learning model to adjust parameters and generate the next flight instruction data.

[0252] The present invention reads flight instructions and simultaneous localization and mapping data, controls the UAV to reach the target camera, and uses means such as negative pressure adsorption to establish a stable hovering attitude. Sequentially collect millimeter-wave reflection coefficients, polarization scattering matrices, and acoustic echoes; generate a "stain spectrum fingerprint" through multimodal fusion, and calculate cleaning parameters based on this. Under the power trajectory given by the model predictive controller, perform plasma discharge → acoustic cavitation stripping → ion air curtain purging; output the cleaning progress field and residual index in real time. If the residual index is higher than the threshold, adjust the cleaning parameters and repeat the three-stage process until the residual index no longer exceeds the standard. After the residual index meets the standard, instruct the UAV to return to the edge base station to complete battery replacement; the energy management unit encapsulates the cleaning progress field, residual index, and energy settlement information into an operation closed-loop data packet and uploads it to the chain. The cloud self-learning model online trains this data packet, updates the algorithm weights, and generates the next batch of flight instructions to achieve a continuously optimized cleaning cycle.

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

[0254] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. An autonomous cleaning system for main road cameras based on drones, characterized in that, Including: A flight positioning module, configured to read flight instruction data and simultaneous localization and mapping data, and drive the drone to fly to the target camera; A stain detection and decision-making module, configured 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; A cleaning and quality inspection module, configured to sequentially perform cleaning processing according to the cleaning parameters under the power trajectory defined by the model predictive controller, obtain cleaning progress field data and quality inspection residual index data in real time, and adjust the cleaning parameters to repeat the processing when the quality inspection residual index data exceeds a threshold until the quality inspection residual index data does not exceed the threshold; An energy module, configured to control the drone to return after the quality inspection residual index data does not exceed the threshold, encapsulate the cleaning progress field data, quality inspection residual index data, and energy settlement data into an operation closed-loop data packet to adjust the self-parameters of the self-learning model, and generate the next flight instruction data.

2. The system according to claim 1, wherein The flight positioning module includes an inertial measurement unit, a lidar, and a vision recognizer. The flight positioning module jointly calculates 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 vision recognizer through a fusion algorithm to obtain the three-dimensional position and attitude of the drone.

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

4. The system according to claim 1, wherein The stain detection and decision-making module includes a multimodal data fusion unit and a fingerprint optimization unit. The multimodal data fusion unit fuses the 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 according to the stain spectrum fingerprint.

5. The system according to claim 4, wherein The fingerprint optimization unit outputs the set of cleaning parameters by using the Bayesian search algorithm with the goal of minimizing the quality inspection residual index.

6. The system according to claim 1, wherein The cleaning and quality inspection module sequentially includes a plasma generating device, an acoustic cavitation transducer device, and an ionic air curtain device. The cleaning and quality inspection module further includes a model predictive controller. The model predictive controller generates a power trajectory within a preset energy constraint and drives the plasma generating device, the acoustic cavitation transducer device, and the ionic air curtain device.

7. The system according to claim 6, 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.

8. The system according to claim 1, wherein The energy module includes an automatic battery swapping mechanism and an energy management unit. The automatic battery swapping mechanism completes battery replacement after the drone returns. The energy management unit records the battery power before and after replacement and outputs energy settlement data.

9. The system according to claim 8, wherein After generating the closed-loop operation data packet, the energy management unit calculates a hash value for the closed-loop operation data packet and adds a digital signature. After writing the data into the account book through the blockchain node, it is sent to the cloud via the network communication interface. The self-learning model uses an incremental gradient boosting decision tree to perform online training on the received closed-loop operation data packet and returns the next flight instruction data after the training is completed.

10. A method for autonomously cleaning a main road camera based on a drone, which is used to execute the system for autonomously cleaning a main road camera based on a drone according to any one of claims 1-9, characterized in that, The method comprises the following steps: Read flight command data and synchronize positioning and map building data to drive the drone to the target camera and establish a fixed posture; Collecting millimeter wave reflection coefficient data, polarization scattering matrix data, and acoustic echo data, fusing them to generate stain spectrum fingerprint data, and determining cleaning parameters based on the stain spectrum fingerprint data; performing a plasma discharge process, an acoustic cavitation stripping process, and an ion air curtain purge process in sequence according to the cleaning parameters and under a power trajectory defined by a model predictive controller, and acquiring cleaning progress field data and quality inspection residue index data; when the quality inspection residue index data is greater than a threshold, adjusting the cleaning parameters and repeating the process until the quality inspection residue index data is less than or equal to the threshold; 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 an operation closed-loop data packet for the self-learning model to adjust parameters and generate the next flight instruction data.

Images