A ventilator adaptive cleaning method and system

Abstract

This invention provides an adaptive cleaning method and system for ventilators. The invention relates to the field of industrial cleaning and automation control technology. The evaluation method includes: a drone inspecting a roof containing circular ventilators; an environmental perception module acquiring image information and identifying the circular ventilators; controlling the drone to approach the identified circular ventilators and extracting the geometric parameters of the crescent-shaped vent and the internal dirt state parameters; generating a set of control commands including a jet direction vector, jet area, and jet pressure; executing cleaning operations based on the control commands; real-time analysis of image features generated by the pilot jet to verify the accuracy of the jet direction vector; if an angle deviation is detected, angle compensation is triggered; after confirming the accuracy of the jet direction vector, monitoring image information during the cleaning process to evaluate the cleaning effect; if the cleaning effect is determined to be unsatisfactory, dynamically adjusting the jet pressure.

Description

Technical Field

[0001] This invention relates to the field of industrial cleaning and automation control technology, specifically to an adaptive cleaning method and system for ventilators. Background Technology

[0002] In the architectural structure of large industrial plants and warehousing facilities, rooftop powered and non-powered ventilation systems are core components for ensuring indoor air circulation, regulating temperature and humidity, and discharging production waste gases. Among these, circular turbine ventilators are widely used in industrial buildings due to their excellent aerodynamic layout and the fact that they do not require electricity. From the perspective of structural mechanics and fluid mechanics, these ventilators typically consist of dozens of radially arranged arc-shaped blades with specific curvatures, which overlap to form complex flow channel characteristics. This overlapping design aims to meet all-weather rain protection requirements, ensuring that rainwater slides off the outer edges of the blades under gravity and does not directly enter the room. However, this non-linear physical structure also results in irregular crescent-shaped gaps at the ventilator's exhaust outlets, with significant shielding and geometric depth in three-dimensional space, posing a significant technical challenge to subsequent automated cleaning and maintenance.

[0003] Due to the large amounts of dust, grease, and fine particulate matter generated during industrial production, which are discharged with the airflow, dirt easily accumulates in the narrow gaps formed by overlapping blades due to the vortex at the blade edges and surface adhesion. Particularly at the pointed ends of the crescent-shaped exhaust vents, where the radius of curvature is smaller and the airflow velocity changes abruptly, the dirt deposition density is significantly higher than in the central area. Long-term dust accumulation not only increases the rotational inertia of the ventilator and reduces exhaust efficiency, but may also cause irreversible damage to the metal substrate of the blades due to the accumulation of corrosive substances, shortening the service life of the facility. Therefore, regular deep cleaning of these irregular exhaust vents is a necessary part of industrial operation and maintenance.

[0004] However, existing industrial cleaning technologies exhibit significant limitations when dealing with irregular exhaust vents with complex and concealed characteristics. Currently, the mainstream cleaning method still relies primarily on manual high-altitude operations, using high-pressure water guns for manual spraying. In this mode, because operators are on an unstable high-altitude platform, it is difficult to precisely control the jet direction, resulting in a significant loss of high-pressure water energy on the outer surface of the blades, failing to effectively penetrate into the overlapping gaps. Furthermore, the adjustment of pressure and jet envelope area during manual cleaning depends mainly on subjective experience, lacking quantitative standards, which easily leads to uneven cleaning or localized pressure overload that damages the blade structure.

[0005] With the development of unmanned technology, although drone cleaning systems equipped with spray devices have emerged, existing systems are mostly designed for regular, flat surfaces such as glass curtain walls and photovoltaic backsheets. Their core control logic is primarily based on two-dimensional path planning and fixed-parameter spraying. When dealing with spatially irregular exhaust vents, existing visual recognition algorithms can only achieve a rough overall location of the target object, failing to provide refined perception of the geometric parameters, spatial orientation, and real-time blockage status of each micro-exhaust vent. Due to the lack of an adaptive jet vector adjustment mechanism, existing equipment maintains a fixed spray angle and pressure during movement, unable to dynamically compensate for the geometric characteristics of different locations within gaps. This execution mode, lacking a feedback loop, directly leads to low cleaning efficiency when facing deeply blocked or high-obscuration areas, while resulting in significant waste of water and energy in clean areas. How to achieve accurate identification, dynamic vector alignment, and on-demand energy allocation for nonlinear exhaust vent structures in complex industrial environments has become a critical technical challenge that urgently needs to be addressed in the field of automated cleaning. Summary of the Invention

[0006] This invention provides an adaptive cleaning method for ventilators, the cleaning method comprising:

[0007] The drone inspects the roof containing the circular ventilator. The environmental perception module collects image information, identifies the circular ventilator, and extracts environmental benchmarks that reflect the current ambient light characteristics.

[0008] Control the drone to approach the identified circular ventilator and extract the geometric morphology parameters and internal dirt status parameters of the crescent-shaped vent.

[0009] Based on the geometric parameters, the jet direction vector and jet area for achieving jet penetration are calculated. Based on the internal dirt state parameters, the jet pressure for removing dirt is calculated. A set of control commands containing the jet direction vector, jet area and jet pressure are generated along the central axis of the crescent-shaped vent.

[0010] The cleaning operation is executed based on control commands. The cleaning operation includes starting the pilot jet, analyzing the image features generated by the pilot jet in real time to verify the accuracy of the spray direction vector, triggering angle compensation if an angle deviation is detected, monitoring the image information during the cleaning process to evaluate the cleaning effect after confirming that the spray direction vector is accurate, and dynamically adjusting the spray pressure if the cleaning effect is not as expected.

[0011] The steps for extracting the geometric morphology parameters and internal dirt state parameters of the crescent-shaped vent include:

[0012] The high-resolution image of the crescent-shaped vent is processed by minimizing the energy function E(C) to make the initial contour converge to the true inner and outer contours Cin and Cout of the crescent-shaped vent:

[0013] E(C) = 01g(| IHDR(C(q))|)|C'(q)|dq

[0014] Where C(q) is the contour evolution curve, q is the curve parameter, IHDR is the high-resolution grayscale image, and g is the edge stopping function:

[0015] g(| IHDR|)=11+β| IHDR|2

[0016] Where β is a positive parameter of the sensitivity of the control function;

[0017] Based on the converged inner and outer contours Cin and Cout, the width gradient representing the geometric morphological parameters is calculated. w(s) and the internal shadow gradient Gs(x,y) characterizing the internal dirt state parameters, where s is the arc length parameter along the central axis of the crescent-shaped vent.

[0018] Calculate the local normal vectors nin(s) and nout(s) of the inner and outer contours Cin and Cout at position s on the central axis of the crescent-shaped vent, respectively.

[0019] The local radial vector Nchan(s) and local tangential vector Tchan(s) of the blade passage are calculated based on the local normal vector; the final injection direction vector Vjet(s) is set as follows:

[0020] Vjet(s) = cos( pitch)Tchan(s)+sin( pitch)Nchan(s)

[0021] in, pitch is the pitch angle determined based on the change in the projection of the crescent-shaped vent in the vertical direction.

[0022] The calculation of the jet area Ajets includes:

[0023] Ajet(s) = Amin + (Amax) Amin) w(s)wmax e α| w(s)|

[0024] Where Amax and Amin are the maximum and minimum adjustable spray areas of the nozzle, respectively, w(s) is the opening width at arc length s, wmax is the maximum width of the crescent-shaped vent, and α is a positive attenuation coefficient. w(s) is the width gradient.

[0025] The step of calculating the injection pressure Pjet(s) is achieved using a piecewise function:

[0026] Pjet(s) = PbaseifGs,max(s) <TGPbase+η(Gs,max(s) TG)ifGs,max(s)≥TG

[0027] Where Pbase is the basic cleaning pressure, Gs,max(s) is the maximum value of the internal shadow gradient Gs(x,y) within the local rectangular analysis window at arc length s, TG is the gradient threshold for determining whether the dirt is a stubborn agglomerate, and η is the gain coefficient.

[0028] The step of generating a set of control commands also includes calculating a pressure modulation factor γ(s) based on geometric features and coupling it with the injection pressure Pjet(s) to obtain the final command pressure Pjet'(s):

[0029] γ(s)=1 e δ / | w(s)|

[0030] Pjet'(s)=γ(s) Pjet(s)

[0031] Where δ is a positive adjustment parameter.

[0032] The step of real-time analysis of image features generated by the pilot jet to verify the accuracy of the jet direction vector specifically includes: calculating a jet antisplash flux ΦBS.

[0033] ΦBS=1Δt LC∮C(s)max(0,f(l) nb(l))dl

[0034] Where C(s) is the crescent-shaped vent outline corresponding to the current control point, l is the arc length parameter on the outline, LC is the total length of the outline, Δt is the time interval between image frames, f(l) is the optical flow vector at point l on the outline, and nb(l) is the unit outward normal vector of the outline at point l.

[0035] When the jet backsplash flux ΦBS is not less than a preset threshold TΦ, it is determined that an angle deviation has been detected.

[0036] After determining that an angle deviation has been detected, the step of triggering angle compensation also includes: calculating the splash establishment time tr and the steady-state fluctuation coefficient Cv by analyzing the dynamic characteristics of the jet splash flux ΦBS(t) in the time series;

[0037] If tr is less than a time threshold Tr and Cv is less than a fluctuation threshold Tcv, then it is determined to be geometric splashing, and the optimal jet direction vector is iteratively searched with the goal of minimizing ΦBS;

[0038] If tr is not less than Tr or Cv is not less than Tcv, it is determined to be physical backsplash, the current jet direction vector is locked as the optimal angle, and the current control point is marked as high-resistance blockage.

[0039] The steps for dynamically adjusting the jet pressure specifically include: calculating the dirt removal rate. c(si), the dirt removal rate is defined as the average decay rate of the dirt texture energy field during a cleaning pulse of fixed duration; if the dirt removal rate is detected... c(si) is lower than an expected rate If exp(si) and the jet backsplash flux ΦBS remains low, it is determined to be a deep blockage, and the command pressure at the current control point is gradually increased by a preset increment ΔP until... c(si) recovers to the expected level or reaches the safe limit.

[0040] This invention provides an adaptive cleaning system for ventilators, the system comprising:

[0041] Data Acquisition Module: The drone inspects the roof containing the circular ventilator. The environmental perception module collects image information, identifies the circular ventilator, and extracts environmental benchmarks that reflect the current ambient light characteristics.

[0042] Feature extraction module: Controls the drone to approach the identified circular ventilator and extracts the geometric morphology parameters and internal dirt state parameters of the crescent-shaped vent.

[0043] Command generation module: Based on geometric morphology parameters, calculates the jet direction vector and jet area for achieving jet penetration; based on internal dirt state parameters, calculates the jet pressure for removing dirt; and generates a set of control commands containing the jet direction vector, jet area, and jet pressure along the central axis of the crescent-shaped vent.

[0044] Correction module: Executes cleaning operations based on control commands. The cleaning operations include activating the pilot jet, analyzing the image features generated by the pilot jet in real time to verify the accuracy of the spray direction vector, triggering angle compensation if an angle deviation is detected, monitoring image information during the cleaning process to evaluate the cleaning effect after confirming that the spray direction vector is accurate, and dynamically adjusting the spray pressure if the cleaning effect is not as expected.

[0045] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned adaptive cleaning method for a ventilator.

[0046] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned adaptive cleaning method for a ventilator.

[0047] The adaptive cleaning method and system for ventilators disclosed in this invention have significant advantages over existing technologies. This invention fundamentally solves the core technical problems of low accuracy, poor efficiency, and weak adaptability in traditional cleaning methods when dealing with targets such as industrial rooftop circular turbine ventilators with complex three-dimensional structures and irregular pollution states by constructing a complete workflow from global perception, local modeling, strategy generation to execution.

[0048] After switching to local operation mode, the segmentation algorithm of the present invention can accurately extract the complete shape of the crescent-shaped vent based on the physical reality that the inner and outer edges of the vent have different clarity. Even if the inner edge is blurred or interrupted due to dirt, it can still extract the complete shape by relying on the global geometric properties of the contour.

[0049] In generating the cleaning strategy, by parameterizing the geometry of the crescent-shaped vent, the most geometrically obscured tip area can be accurately identified. Based on this, the system can adaptively adjust the spray area, using a large-area fan-shaped jet in wide areas to improve efficiency, and switching to a high-energy-density needle-shaped jet in narrow tips, achieving ultimate adaptation to the geometry. Regarding energy distribution, this invention analyzes the texture energy field of the dirt to more accurately assess the physical adhesion strength of the dirt, thereby achieving on-demand distribution of spray pressure, increasing the output pressure only in areas where stubborn clumps are detected. Furthermore, this invention actively suppresses high pressure in the most geometrically obscured areas through pressure modulation, prioritizing the penetration of the jet. This intelligent strategy of penetration rather than brute force is the key to fundamentally solving the problem of cleaning narrow gaps.

[0050] This invention features a closed-loop verification and dynamic adjustment mechanism with fault diagnosis capabilities during the execution phase. Before cleaning, an incident vector verification method based on real-time optical flow analysis ensures the absolute accuracy of the spray angle. By analyzing the spatiotemporal dynamic characteristics of the backsplash pattern, it distinguishes between geometric backsplash caused by angle errors and physical backsplash caused by blockages, thus making correct decisions and avoiding incorrect corrections to the correct angle. During the cleaning process, the system quantifies the dirt removal rate through a spray-sensing-decision micro-cycle and can again diagnose online, through dual-condition judgment logic, angle failures caused by drone displacement or ventilator rotation and deep blockages caused by unexpected dirt strength, triggering corresponding angle recalibration or energy compensation procedures. This transforms the cleaning system of this invention from a simple pre-programmed executor into an intelligent operator capable of real-time sensing, decision-making, and adaptation to unknown changes, thereby ensuring the highest quality and efficiency of the final cleaning operation in complex and ever-changing industrial environments. Attached Figure Description

[0051] Figure 1 This is a flowchart of the adaptive cleaning method for ventilators according to the present invention;

[0052] Figure 2 The ventilator to be cleaned according to the present invention. Detailed Implementation

[0053] In a preferred embodiment of this application, an adaptive cleaning method and system for ventilators is provided. This system is primarily deployed in rooftop environments of large industrial plants or warehouses. A typical controlled object in this environment is a circular turbine ventilator, whose physical structure includes a cylindrical support base fixed to the roof and a rotatable turbine blade assembly. The blade assembly consists of several metal arc-shaped blades with a preset curvature, stacked at equal intervals along the circumferential direction. The overlapping structure of adjacent blades in the circumferential direction forms an irregular crescent-shaped vent on the radially outer side of the ventilator. This crescent-shaped vent exhibits a non-linear spiral extension trend in three-dimensional space. The normal vector direction of its opening cross-section dynamically deflects with changes in blade height and circumferential position, and significant geometric shading exists at the tip of the crescent shape. In practical applications, the interior of such crescent-shaped vents and the deep spaces at the blade overlap areas are often covered with industrial dust or grease clumps with uneven physicochemical properties.

[0054] The core of the hardware system in this embodiment is a multi-rotor unmanned aerial vehicle (UAV) flight platform. The main body of the fuselage is preferably constructed of high-strength carbon fiber composite material to support complex onboard payloads. A global navigation satellite system module integrating real-time dynamic differential technology is installed on the top of the flight platform to provide centimeter-level spatial positioning references in open rooftop environments. A flexible water supply pipeline is coupled to the lower part of the flight platform's fuselage via a high-pressure fluid interface. This pipeline extends upwards to a fluid distribution unit inside the fuselage and downwards to a high-pressure water pump station via a drag-and-drop connection.

[0055] An environmental perception module is mounted on the nose and front of the fuselage of the flight platform using a three-axis mechanical stabilization gimbal. This module specifically includes a high-resolution industrial camera with a global shutter and a depth camera based on the infrared time-of-flight principle. To maintain attitude stability in complex lighting or signal obstruction environments, the system also integrates a visual inertial odometry system, which provides real-time six-DOF attitude parameters to the system through a multi-sensor fusion algorithm.

[0056] The drone flight platform is equipped with an adaptive cleaning execution module at its base. The core of this module is a vector pointing mechanism with dual degrees of freedom, which drives the jet terminal to independently deflect in both horizontal azimuth and pitch directions via a high-precision servo motor. The jet terminal is equipped with a multi-functional combined nozzle with a variable orifice. The nozzle contains an orifice adjustment device controlled by an electromagnetic proportional valve, capable of steplessly adjusting the outlet orifice diameter within a range of 1.0 mm to 1.4 mm based on a control signal, thereby altering the jet's convergence and energy distribution. The fluid circuit of this execution module is connected in series with a pressure sensor and a flow sensor. Its rated operating pressure is preferably 8 MPa, and its maximum rated flow rate is 18 L / min, ensuring the jet has sufficient kinetic energy to penetrate the shielding layer in the overlapping area of ​​the blades.

[0057] The system's logic processing and coordination center is the airborne integrated control unit, which employs an embedded industrial computer with a high-performance graphics processing core. This control unit is electrically connected to the environmental perception module, flight control system, and vector pointing mechanism via a dedicated airborne bus. In terms of spatial arrangement, the high-resolution industrial camera is calibrated and fixed to the upper side of the nozzle trajectory, ensuring that the center of the camera's imaging field of view accurately covers the water contact area of ​​the nozzle jet at the crescent-shaped vent.

[0058] In one specific embodiment of the present invention, the initial stage of the adaptive cleaning method requires global perception and environmental baseline determination. The goal of this stage is to identify and lock the spatial pose of the target cleaning object—a circular turbine ventilator—when the UAV is performing a large-scale roof inspection mission, and to establish the necessary environmental reference system for localized fine-tuning operations.

[0059] The drone flight platform performs a full-coverage scan of the industrial building roof following a pre-set "bow"-shaped path. During this process, the onboard visual inertial odometry calculates the drone's pose matrix TWB in the global world coordinate system W in real time. This matrix includes the rotation matrix RWB and translation vector Pu of the drone's body coordinate system B relative to the world coordinate system W. Simultaneously, the onboard integrated control center uses a pre-trained deep learning target detection model to analyze the real-time video stream captured by the high-resolution industrial camera to identify and locate the circular turbine ventilator within the field of view. Once the model detects the target, it outputs the target's bounding box and confidence score in the two-dimensional image coordinate system.

[0060] After successfully detecting the circular turbine ventilator, a coarse-adjustment vector calculation is immediately performed to obtain the radial deflection angle θr between the UAV's current heading and the center of the target ventilator. θr is a globally preset reference for local alignment. The world coordinates Pv of the ventilator's center point in 3D space are calculated by combining the depth map acquired by the depth camera with the bounding box of the detected target. Subsequently, the vector Vuv pointing from the UAV to the ventilator center is calculated.

[0061] Vuv=Pv Pu

[0062] Where, Pv∈ 3 represents the position vector of the center of the circular turbine ventilator in the world coordinate system W, Pu∈ 3 represents the current position vector of the UAV in the world coordinate system W, which is directly provided by TWB.

[0063] Simultaneously, it is necessary to determine the current heading vector Hu of the UAV in the world coordinate system W. This vector is defined as the forward axis of the UAV's body coordinate system B, i.e., the projection of the X-axis onto the horizontal plane of the world coordinate system W. Let the unit vector of the forward axis of the UAV's body coordinate system B be xB, then its representation in the world coordinate system W is:

[0064] Hu'=RWB xB

[0065] Projecting Hu' onto the XY plane of the world coordinate system W (assuming the Z-axis is vertical) yields the final heading vector Hu. Finally, by calculating the angle between these two vectors Vuv and Hu, the radial deflection angle θr is obtained.

[0066] θr=arccosVuv Hu‖Vuv‖‖Hu‖

[0067] The radial deflection angle θr is not used to instantaneously trigger the cleaning action, but rather as an initial heading correction command when the UAV switches from global inspection mode to local operation mode. In the open environment of industrial rooftops, the UAV's heading is easily affected by environmental factors such as crosswinds, causing it to drift. This global preset benchmark ensures that the UAV's nose is roughly facing the ventilator when approaching the target, thus significantly reducing the search space and convergence time of the subsequent local fine-tuning alignment algorithm, demonstrating the guiding role of global perception on local execution.

[0068] While completing target locking and coarse vector calculation, the ambient lighting baseline is extracted. The determination of dirt inside the crescent-shaped vent heavily relies on image contrast and brightness information, which fluctuates drastically with weather and lighting conditions during operation. To eliminate this uncertainty, a ring-shaped background area (ROI) is defined outside the bounding box of the target detection output. This ROI excludes shadows and specular interference from the ventilator itself and accurately reflects the background lighting characteristics of the current environment. The system converts the image within this ROI from the RGB color space to the CIELAB color space and calculates the luminance component L of all pixels within the region. The average value of Lbase and the chromaticity component a and b The average value (abase, bbase) is recorded. This set of parameters (Lbase, abase, bbase) is recorded as the environmental baseline associated with the ventilator. This data serves as an adaptive reference system during the cleaning effectiveness verification phase. For example, when determining whether a cleaned area is clean, the system will use the Lbase value of that area. Instead of using a fixed brightness threshold, the value is compared with Lbase in a normalized manner. This real-time ambient light benchmark establishment method ensures the robustness and consistency of cleaning quality judgment.

[0069] After the UAV completes global perception and environmental baseline anchoring, the adaptive cleaning method enters the second stage. This embodiment will now elaborate on the three-dimensional modeling of the crescent-shaped vent in this stage. The goal of this step is to perform high-resolution imaging of a single crescent-shaped vent after the UAV switches to local operation mode, and to extract its geometric parameters and internal dirt status, providing high-precision input for generating control commands.

[0070] After the UAV completes initial alignment and approaches the target based on the radial deflection angle θr, a high-resolution industrial camera captures images of the crescent-shaped vent. Due to the potential metallic reflection on the blade surface and the dim lighting inside the vent, the images exhibit high dynamic range (HDR) characteristics. Therefore, the system first acquires a sequence of images taken at different exposure times and then generates a single-channel grayscale image IHDR(x,y) with high dynamic range through a multi-exposure fusion algorithm.

[0071] Subsequently, edge detection and contour extraction algorithms are executed to overcome the asymmetry caused by the blurring of the inner edge of the crescent-shaped vent due to dirt coverage and the clearness of the outer edge due to strong light contrast.

[0072] This embodiment preferably employs an edge detection and contour extraction algorithm, specifically including:

[0073] Calculate the structure tensor J of IHDR(x,y), which describes the gradient distribution in the neighborhood of each pixel:

[0074] J=w(Ix)2w(IxIy)w(IxIy)w(Iy)2=J11J12J21J22

[0075] Where Ix and Iy are the partial derivatives of the image in the x and y directions, obtained through convolution with the Sobel operator; w represents a Gaussian weighted window. By performing eigenvalue decomposition on the structure tensor J, two eigenvalues ​​λ1≥λ2≥0 and their corresponding eigenvectors are obtained. In the crescent-shaped edge region, λ1... λ2, and the direction of the eigenvector corresponding to λ1 is consistent with the direction of the edge normal.

[0076] Next, the contour evolution is performed. The initial contour is defined by a rectangular frame that roughly surrounds the vent. Contour evolution curve C(q):→ The energy function E(C) of 2 is defined as:

[0077] E(C) = 01g(| IHDR(C(q))|)|C'(q)|dq

[0078] Where g is an edge stopping function related to the image gradient, and its form is:

[0079] g(| IHDR|)=11+β| IHDR|2

[0080] Here, β is a positive parameter controlling the sensitivity of the function. The energy function E(C) is minimized by solving the gradient descent flow equation, thereby driving the initial contour to converge towards the true edge.

[0081] C t=(gκ g N)N

[0082] Where t is the evolution time, κ is the curvature of curve C, and N is the unit in-line normal vector of the curve. The advantage of this invention lies in utilizing not only local gradient information but also the global geometric properties of the contour, i.e., curvature, thus exhibiting extremely high robustness to cases where the inner edge has partial breaks or blurring. After the evolution finally stabilizes, the precise inner and outer contours Cin and Cout of the crescent-shaped vent are obtained.

[0083] After extracting the contour, geometric parameterization is performed to calculate the opening width gradient.

[0084] The distance from each pixel within the crescent-shaped region to the nearest contour is calculated to find the region's skeleton or central axis S. This central axis S can be parameterized by the arc length parameter s∈[0,LS], where LS is the total length of the central axis. For any point S(s) on the central axis, the Euclidean distance between its normal direction and the intersection of the inner and outer contours Cin and Cout is defined as the opening width w(s) at that point. To quantify the abrupt geometric changes at the two ends of the crescent shape, the system calculates the spatial rate of change of the opening width w(s) along the central axis, i.e., the width gradient. w(s):

[0085] w(s)=dw(s)ds

[0086] This gradient value w(s) represents the degree of spatial shading in the overlapping area of ​​the blades. In the tip region where s is close to 0 or LS, The absolute value of w(s) increases significantly, and the gradient value... w(s) is the basis for the execution of the cleaning control strategy.

[0087] The internal shadow gradient is extracted within the region defined by Cin and Cout to characterize the physical state of the dirt. To eliminate shadow misjudgment caused by uneven illumination, the image IHDR(x,y) of this region is first normalized for brightness using the established environmental baseline brightness Lbase.

[0088] Inorm(x,y)=IHDR(x,y)IROI Lbase

[0089] Wherein, IROI is the average brightness of the vent area. To describe the complex texture of the dirt, the system employs the Laws texture energy method. The normalized image Inorm(x,y) is convolved with a set of predefined 5x5 convolution kernels (preferably L5, E5, S5, W5, R5) to generate multiple texture energy maps. This embodiment focuses particularly on energy maps that reflect surface roughness and edge features, specifically texture energy maps generated by L5E5 and E5L5 kernel pairs. These energy maps are then weighted and fused to obtain a dirt texture energy field Etexture(x,y). The internal shadow gradient Gs(x,y) is defined as the gradient vector of this texture energy field:

[0090] Gs(x,y)= Etexture(x,y)= Etexture x, Etexture y

[0091] The amplitude |Gs(x,y)| quantifies the non-uniformity and complexity of the dirt's adhesion state. Regions with high amplitudes correspond to the edges of dirt clumps or abrupt material changes, representing key areas requiring higher cleaning pressure. Thus, the 3D modeling of the crescent-shaped slit provides clear geometric constraints for the strategy (by...). w(s) characterization) and state constraints (characterized by Gs(x,y).

[0092] In industrial environments, the dirt adhering to ventilator blades is not a uniform thin layer. It is typically a heterogeneous mixture of dust particles of varying sizes, sticky oil droplets, moisture, and chemical reaction products. Over time, these substances undergo processes such as deposition, solidification, and agglomeration, forming an adhesion layer with a complex three-dimensional microstructure. When light shines on the surface of this adhesion layer, the differences in its microstructure lead to highly non-uniform light scattering and absorption. Specifically: for smooth, thin layers of dust, the surface is relatively flat, scattering light more uniformly, appearing in images as areas with gradual brightness changes and poor texture details. Therefore, the corresponding texture energy field Etexture(x,y) has a low and gradual value, and the calculated gradient amplitude |Gs(x,y)| is correspondingly small. For hard particle agglomerates or solidified oil stains, this type of dirt has a significant three-dimensional structure, with a rough and uneven surface containing numerous micro-pits, protrusions, and cracks. These structures produce rich microscopic shadows and highlights, causing drastic changes in local light intensity. In Laws texture energy methods, edge detection kernels (such as E5) and speckle detection kernels (such as S5) are particularly sensitive to such high-frequency signals. Therefore, in these regions, the calculated texture energy field Etexture(x,y) exhibits local peaks, and its spatial rate of change, i.e., the gradient, increases sharply. Especially at the edges of dirt clumps, due to abrupt changes in material and thickness compared to the surrounding area, the strongest texture boundary is formed, causing |Gs(x,y)| to reach a local maximum. Therefore, the amplitude of the internal shading gradient Gs(x,y) is essentially a quantification of the complexity of the three-dimensional microstructure of the dirt layer surface. A higher amplitude indicates a more non-uniform physical morphology and a more complex structure of the dirt, corresponding to higher adhesion strength and stronger physical / chemical bonding (such as hard clumps and solidified oil stains). Overcoming this adhesion requires greater mechanical impact energy during cleaning operations. Correlating the jet pressure Pjet(s) with the maximum value of the gradient amplitude Gs,max(s) directly matches the input of cleaning energy with the energy threshold required to remove dirt, thereby ensuring that energy is allocated on demand and avoiding insufficient energy or excessive impact.

[0093] After completing the 3D modeling of the crescent-shaped vent, the core of the adaptive cleaning method is to generate an energy-angle control matrix. This matrix is ​​not a matrix in the traditional mathematical sense, but rather a set of parameterized control commands distributed along the central axis S of the crescent-shaped vent. For any position of arc length s on the central axis, this matrix defines a set of optimal cleaning execution parameters, including the jet direction vector, jet pressure, and jet area. The purpose of this step is to decouple the extracted, independent geometric constraints from the state constraints and map them into specific, executable physical control quantities.

[0094] The goal of geometric decoupling is to calculate the injection direction vector Vjet(s) and injection area Ajet(s) corresponding to the maximum penetration efficiency at the location of the crescent-shaped vent based on the local geometry of the vent.

[0095] The injection direction vector Vjet(s) is calculated, and its determination must be coupled with the helical channel structure formed by the overlapping blades. For any point S(s) on the central axis, its position in the cylindrical coordinate system of the ventilator is first determined. Through estimation using three-dimensional point cloud data, the local tangential vector Tchan(s) and radial vector Nchan(s) of the blade channel at that point can be obtained. Due to the helical arrangement of the blades, the optimal incident direction is not purely tangential, but rather involves a pitch angle. pitch.

[0096] Specifically, after the drone is aimed at the crescent-shaped vent C(s) and hovers stably,

[0097] For a point S(s) on the central axis, two corresponding contour points Pin(s) and Pout(s) are determined on the inner and outer contours Cin and Cout, respectively, along its normal direction. Using the two-dimensional pixel coordinates (u,v) in the industrial camera image and the depth value d provided after registration with the depth camera, the system can calculate the three-dimensional coordinates of Pin(s) and Pout(s) in the UAV body coordinate system B. To construct the local surface, not only is the point at s sampled, but also the points in the neighborhood of s [s... Within the range [Δs, s+Δs], a series of points on the inner and outer contours are densely sampled to form two local 3D point sets, Pin and Pout. For each 3D point in Pin and Pout, the normal vector of the small surface containing that point is estimated by analyzing the spatial distribution of its neighboring points. Specifically, Principal Component Analysis (PCA) is used. For any point Pi in the point set, its k nearest neighbors are selected to form a local point cloud. The covariance matrix of this local point cloud is calculated as follows:

[0098] Σ=1kj=1k(Pj P)(Pj P)T

[0099] Here, P is the centroid of these k points. Eigenvalue decomposition is performed on the covariance matrix Σ; the eigenvector corresponding to the smallest eigenvalue is the normal vector ni of the local surface. Using this method, the corresponding normal vectors can be calculated for all points in Pin and Pout, thus obtaining two local normal vector fields.

[0100] At position S(s), the local normal vectors nin(s) and nou(s) are calculated at Pin(s) and Pout(s), respectively. These two normal vectors point inwards from the surfaces of two adjacent blades. The ideal channel radial vector should lie in the plane formed by the two blade normal vectors and bisect their included angle. Therefore, Nchan(s) can be approximated by the sum of these two normal vectors:

[0101] Nchan(s)=nin(s)+nout(s)‖nin(s)+nout(s)‖

[0102] The tangential vector must be perpendicular to both the radial vector and the ventilator's axis of rotation. If the ventilator's axis of rotation is Arot, then the tangential vector can be obtained through a cross product operation:

[0103] Tchan(s) = Nchan(s) × Arot

[0104] Therefore, the optimal jet direction vector Vjet(s) is set as follows:

[0105] Vjet(s) = cos( pitch)Tchan(s)+sin( pitch)Nchan(s)

[0106] in, pitch is the tilt angle. Calculating the optimal jet direction vector ensures that the jet water flow follows the physical channel formed by the overlapping blades, rather than directly impacting the outer surface of the blades. This is a prerequisite for achieving deep cleaning.

[0107] Pitch angle The pitch actually reflects the degree of blade tilt relative to the horizontal plane. This tilt can be indirectly inferred by analyzing the change in the projection of the crescent-shaped vent in the vertical direction. After aligning with a crescent-shaped vent, the UAV performs a small vertical lift maneuver, preferably a few centimeters up and down. During this process, the camera continuously acquires images. Because the blade is tilted, the inner and outer contours Cin and Cout in the image will exhibit asymmetrical movement when the UAV moves vertically, i.e., vertical parallax. In particular, the opening width w(s) at a point S(s) on the central axis S will change with the height of the UAV. A simplified geometric model is established to correlate the local tilt angle of the blade with the observed width change rate dw / dz (where z is the height of the UAV). By measuring the value of dw / dz, an approximate tilt angle of the blade relative to the radial plane in this region can be obtained. This tilt angle is used as the pitch angle. Pitch

[0108] The spray area Ajet(s) is determined, and the spray area is directly controlled by the orifice diameter of the multi-functional combined nozzle. The orifice diameter setting is closely related to the local geometric complexity of the crescent-shaped vent, and its core basis is the calculated width gradient. w(s). When | A large w(s) implies a sharply narrowed geometry, requiring a highly concentrated energy beam; conversely, a smaller w(s) allows for a wider-coverage fan-shaped jet. Therefore, the jet area Ajet(s) is defined as a value proportional to the width w(s) and related to the width gradient ||w(s). w(s) is an inversely proportional function.

[0109] Ajet(s) = Amin + (Amax) Amin) w(s)wmax e α| w(s)|

[0110] Where Amax and Amin represent the maximum and minimum adjustable spray area of ​​the nozzle, respectively, and wmax is the maximum width of the crescent-shaped vent. α is a positive attenuation coefficient used to adjust the suppression strength of the width gradient on the spray area. The spray area Ajet(s) ensures that in the wide area in the middle of the crescent shape, the system uses a large-area spray to improve efficiency; while in the narrow areas at both ends, the system automatically switches to a small-area, high-energy-density needle jet to cope with severe geometric shielding.

[0111] The goal of state decoupling is to calculate the jet pressure Pjet(s) required for effective dirt removal based on the physical state of the dirt inside the crescent-shaped vent. The core of this is to calculate the internal shadow gradient Gs(x,y) that characterizes the complexity of the dirt.

[0112] Since the injection pressure is controlled along the one-dimensional central axis s, and Gs(x,y) is a field defined on a two-dimensional region, it is necessary to aggregate the two-dimensional state information onto the one-dimensional control line. For any point S(s) on the central axis, a local rectangular analysis window Ws perpendicular to the tangent at that point is defined. The maximum value of the shadow gradient of all pixels within this window is calculated as the dirt state feature Gs,max(s) at point s:

[0113] Gs,max(s)=max(x,y)∈Ws|Gs(x,y)|

[0114] The magnitude of Gs,max(s) represents the adhesion strength of the most stubborn dirt in that local area. The jet pressure Pjet(s) is defined as a piecewise function positively correlated with Gs,max(s), ensuring energy is distributed as needed.

[0115] Pjet(s) = PbaseifGs,max(s) <TGPbase+η(Gs,max(s) TG)ifGs,max(s)≥TG

[0116] Where Pbase is the base cleaning pressure used to rinse away general floating dust, TG is the gradient threshold for determining whether the dirt is a stubborn agglomerate, and η is a gain coefficient that linearly maps the gradient intensity exceeding the threshold to an additional pressure increment.

[0117] The injection pressure Pjet(s) ensures that the system output pressure is increased only in areas where hard clumps or highly viscous oil contaminants are detected, avoiding unnecessary high-pressure impacts on the blade substrate and saving energy while protecting the ventilator surface.

[0118] Through the above decoupling process, for each control point si discretized along the central axis S, the system generates a control parameter triplet (Vjet(si), Pjet(si), Ajet(si)). The set of these triplets constitutes the complete energy-angle control matrix. To further optimize the cleaning strategy, a pressure modulation factor γ(s) based on geometric features is introduced, ultimately achieving deep coupling between geometry and state.

[0119] γ(s)=1 e δ / | w(s)|

[0120] Where δ is a positive adjustment parameter. The modulation factor γ(s) is in | w(s) approaches 0 at its extreme tip and approaches 1 in the flat middle region. The final command pressure Pjet'(s) is:

[0121] Pjet'(s)=γ(s) Pjet(s)

[0122] Based on this, cleaning strategies were developed for different regions. At the crescent tip (high curvature region), the gradient s in this region is close to 0 or LS, and the width gradient is | w(s)| is maximized. Therefore, Ajet(s) is driven to its minimum value Amin, forming a needle-like jet; simultaneously, the pressure modulation factor γ(s) approaches 0, and even if Gs,max(s) is high, the final command pressure Pjet'(s) will be significantly suppressed. In the region with the most severe geometric shielding and where the jet is prone to backsplashing, the absolute accuracy of the direction vector and the ultimate convergence of the jet are prioritized. The blockage is dismantled through penetration rather than brute force, avoiding energy waste and secondary pollution to the surrounding environment caused by high-pressure backsplashing. In the middle of the crescent (low curvature region), this region| w(s)| is very small, close to 0. Therefore, Ajet(s) is driven to a larger value, forming a fan-shaped jet to improve coverage efficiency; at the same time, the pressure modulation factor γ(s) approaches 1, so that the command pressure Pjet'(s) is completely determined by the state of the dirt Gs,max(s), ensuring that all the calculated necessary pressure can be applied in this area to achieve effective removal of stubborn dirt.

[0123] After generating the complete energy-angle control matrix, the adaptive cleaning method enters the execution and verification phase. This embodiment details the first stage of the cleaning process, namely, incident vector verification, to ensure that the actual jet direction vector Vjet(s) is precisely aligned with the physical channel formed by the overlapping blades before applying high-pressure cleaning, thus resolving the accuracy issue during the cleaning process.

[0124] When the cleaning operation begins along the central axis S, for the first control point s1, the airborne integrated control center first instructs the vector pointing mechanism of the adaptive cleaning execution module to adjust the nozzle to the initial spray direction vector Vjet(s1) defined in the energy-angle control matrix. At the same time, it instructs the orifice adjustment device of the multi-functional combined nozzle to set the corresponding spray area Ajet(s1).

[0125] Subsequently, instead of immediately spraying at the commanded pressure Pjet'(s1), the system initiates a low-pressure pilot jet for verification. This pilot jet, with a pressure Ppilot far below the base cleaning pressure Pbase, aims only to generate a water flow pattern sufficient for clear capture by an industrial camera, without actually stripping away the dirt.

[0126] An industrial camera continuously acquires image sequences of the water splash area at a high frame rate. The onboard integrated control center analyzes these images in real time to quantify the visual features generated by the water flow. This embodiment proposes a jet anti-splash flux evaluation algorithm based on optical flow field analysis to determine the accuracy of the incident vector.

[0127] In the image, the crescent-shaped vent C(s1) is used as the boundary to define the interior as the penetration zone, and the adjacent annular region outside it is the sputtering zone.

[0128] A dense optical flow estimation algorithm is used to calculate the pixel displacement vector field f(x,y) between two consecutive frames. The direction and magnitude of this optical flow vector f(x,y) represent the direction and velocity of the pixel's motion between frames, respectively.

[0129] When there is a deviation in the jet direction vector, the water flow mainly impacts on the outer surface of the blade. According to the principle of conservation of momentum of fluid collision, most of the water droplets will be rebounded, forming an outward sputtering flow. On the contrary, when the direction vector is correct, the water flow enters the interior along the channel, and only a small amount of water mist overflows due to friction or disturbance with the pipe wall. Therefore, the severity of the angle deviation can be evaluated by quantifying the visual motion flux from the penetration area to the sputtering area.

[0130] In this embodiment, the jet back-splash flux ΦBS is defined as a scalar, which is used to quantify the intensity of the outward visual motion passing through a unit length contour line per unit time. Its calculation formula is:

[0131] ΦBS = 1 / Δt LC ∮C(s1) max(0, f(l) nb(l)) dl

[0132] where ∮ C(s1)(…) dl represents the line integral along the closed contour line C(s1) of the crescent-shaped vent, l is the arc length parameter on the contour line, LC is the total length of the contour line, f(l) is the optical flow vector at the point l on the contour line, nb(l) is the unit outer normal vector of the contour line at the point l, and the dot product f(l) nb(l) calculates the projection of the optical flow vector on the normal direction of the contour line. When this projection is positive, it indicates that the pixel motion direction is outward (i.e., sputtering); when it is negative or zero, it indicates inward or parallel to the contour line motion. The function max(0, …) only accumulates all the outward motion components and filters out the inward motion interference. The entire line integral calculates the total optical flow flux passing through the entire contour line outward. Dividing by the total length LC of the contour line and the frame interval Δt normalizes it, so as to obtain a flux value independent of the contour size and the camera frame rate.

[0133] The jet back-splash flux ΦBS is a physical quantity that directly measures whether the water flow sputters rather than penetrates. The larger its value, the more serious the angle deviation and the more intense the energy waste.

[0134] Compare the calculated ΦBS with the threshold TΦ.

[0135] If ΦBS < TΦ: This indicates that almost no back-splash occurs. The system determines that the current jet direction vector Vjet(s1) is accurate, and the incident vector verification passes. Prepare to enter the high-pressure cleaning.

[0136] If ΦBS ≥ TΦ, it indicates that significant splashing has been detected. In this case, it cannot be immediately determined as an angular deviation; instead, a spatiotemporal feature analysis of the splashing pattern should be initiated to analyze the root cause of the splashing. The core of the spatiotemporal feature analysis of the splashing pattern is based on the fact that although both geometric and physical splashing appear as splashing in the image, their flow patterns exhibit significant differences in formation time, spatial distribution, and dynamic stability.

[0137] The system extracts images from the very short time window after the pilot jet is activated, preferably the first 10-20 frames, and performs time-series analysis on the optical flow field.

[0138] During geometric splashing (angle error), the water jet directly and steadily impacts the hard, smooth outer surface of the blade. This splashing forms almost instantaneously upon contact with the jet surface and quickly reaches a steady state. In the image sequence, this is manifested as the jet splashing flux ΦBS(t) rapidly climbing to a high level and then remaining relatively stable with minimal fluctuations. The splashed water mist morphology is spatially regular and concentrated because the geometry of the reflecting surface (blade surface) is determined. The directional distribution of the optical flow field exhibits a regular divergence pattern related to the blade surface normal.

[0139] In the physical splashing (encountering blockage) process, the water flow first enters a narrow and irregular gap, then interacts in complex ways with the blockage, which is of unknown shape, loose material, or viscous texture. The water flow needs time to fill the space within the gap and interact with the blockage before finally being squeezed out or ejected due to pressure buildup; the formation of splashing involves an observable delay. In the image sequence, this manifests as the jet splashing flux ΦBS(t) not immediately reaching its peak, but experiencing a brief rise followed by a highly unsteady state, i.e., irregular fluctuations. This is because, as the water flow interacts with the loose blockage, it continuously opens new microchannels or causes local collapse of the blockage, causing the intensity and direction of the splashing to constantly change. The splashed water mist is typically more diffuse and disordered in space because it does not bounce off a fixed reflective surface but rather emerges unevenly from the entire crescent-shaped gap. The directional distribution of the optical flow field is more chaotic and lacks obvious structure.

[0140] Based on the above analysis, two new quantitative indicators are defined to describe the dynamic characteristics of ΦBS(t) within the time window: backsplash time tr: the time required from the start of the jet to the first time ΦBS(t) reaches 90% of its peak value; and steady-state fluctuation coefficient Cv: the ratio of the standard deviation of ΦBS(t) to its mean value in the time period after tr.

[0141] If tr < Tr (the establishment time is extremely short) and Cv < Tcv (the fluctuation is small and the state is stable), it is determined as geometric backspray, triggering the angle compensation algorithm, aiming to minimize ΦBS, and iteratively searching for the optimal jet direction vector Vjet .

[0142] If tr ≥ Tr (the establishment time is long) or Cv ≥ Tcv (the fluctuation is large and the state is unstable), it is determined as physical backspray, locking the current jet direction vector Vjet(s1), determining it as the geometrically optimal angle, and tagging the control point with a high-resistance blockage label to guide the cleaning strategy in the next stage.

[0143] Among them, Tr and Tcv are experimentally calibrated thresholds.

[0144] After the incident vector verification is successful and it is ensured that the jet direction vector is accurately aligned with the physical channel, the adaptive cleaning method enters the second-level execution stage, that is, energy dynamic adjustment. That is, while ensuring the angle remains accurate, the inner part of the crescent-shaped vent is deeply cleaned according to the command pressure Pjet'(si) defined in the energy-angle control matrix, and the cleaning effect is evaluated in real time through the monitoring logic, and the energy output is dynamically adjusted to cope with the local change of the dirt adhesion strength and unexpected situations during the operation process.

[0145] For each control point si on the central axis S, the system performs the following steps:

[0146] Execute a cleaning pulse with a fixed duration of Δtc at the command pressure Pjet'(si) and the jet area Ajet(si). At the end of each pulse, the jet pressure immediately drops to the pilot jet pressure to eliminate the interference of the high-pressure water mist on the image. The industrial camera captures the clear post-pulse image Ipost(x,y).

[0147] To objectively evaluate the effect of each cleaning pulse, this embodiment defines a dirt stripping rate c, and this rate c measures the attenuation rate of the dirt texture energy field during the cleaning pulse Δtc.

[0148] Before the pulse execution, the initial dirt texture energy field Etexture,pre(x,y) of this area has been recorded. After the post-pulse image Ipost(x,y) is captured, the system performs the same processing on it to obtain the post-pulse dirt texture energy field Etexture,post(x,y). Subsequently, the average energy change within the local rectangular analysis window Wsi is calculated and divided by the pulse duration to obtain the dirt stripping rate:

[0149] c(si) = Etexture,pre(si) Etexture,post(si)Δtc

[0150] Where, Etexture(si) represents the average value of the dirt texture energy field within the window Wsi. The value of c(si) is positive and the larger it is, the higher the cleaning efficiency; if its value is close to zero or negative, it means the cleaning is ineffective or the image is interfered.

[0151] After calculating c(si) in each microcirculation, the system enters the diagnosis to distinguish different reasons for low - efficiency cleaning:

[0152] Case A: Angle failure determination. If during the operation, the drone has a small displacement due to factors such as gusts of wind, or the ventilator itself rotates, the verified optimal angle will fail. This failure will cause the high - pressure water flow to suddenly fail to enter the gap and instead impact the outer surface of the blade. The most direct visual consequence is that the internal dirt is no longer peeled off, and at the same time, there is a strong back - splash outside. When calculating c(si), the jet back - splash flux ΦBS of the post - pulse image is calculated in parallel.

[0153] If the system detects c(si) < T (The dirt stripping rate is lower than a very small positive threshold and the cleaning almost stalls) and ΦBS > TΦ (the jet back - splash flux is significantly higher than the normal threshold), and these two conditions hold simultaneously, then the system makes a determination of angle failure. Immediately abort the second - level execution and force a return to the incident vector verification step to re - search for and verify the angle for the current control point si. This dual - condition determination logic can extremely reliably distinguish angle problems from other problems and avoid misattributing stubborn dirt.

[0154] Case B: Depth blockage determination. The dirt adhesion strength inside the crescent - shaped vent may have local extreme points far exceeding expectations, such as hard crystals formed by special chemical substances. In this case, even if the spraying angle is completely correct, the command pressure Pjet'(si) decoupled from the initial state may not be sufficient to peel it off. Its visual manifestation is no abnormal back - splash, but the dirt stripping speed continues to be significantly lower than expected.

[0155] If the system detects ΦBS < TΦ (the jet back - splash flux remains low, proving that the angle is correct) but c(si) If both conditions are met (the actual stripping rate is significantly lower than the expected rate), the system determines that there is deep blockage and triggers an energy compensation procedure. This procedure gradually increases the command pressure at the current control point by a preset increment ΔP, meaning the new command pressure becomes Pjet'(si) + ΔP. After a new cleaning pulse, the system reassesses... c(si). This process will be iterative until c(si) is reached. c(si) returns to the expected level, or the command pressure reaches the safety limit Pmax set to protect the equipment.

[0156] Through the aforementioned second-level execution and dynamic energy adjustment mechanism, this invention constructs an intelligent closed-loop cleaning system. It not only executes preset cleaning strategies, but more importantly, during execution, it can diagnose and address the two core challenges of angle drift and unexpected dirt intensity in real time through high-frequency perception-decision cycles, thereby ensuring the final quality and efficiency of cleaning operations in complex and ever-changing industrial environments.

[0157] After completing the cleaning operation on the last control point sN, the UAV flight platform adjusts its position, and the system performs global coordinate connection. The onboard integrated control center calls the pose matrix TWB in the global world coordinate system W before the UAV switched to local operation mode. Based on this coordinate, the flight control system plans a safe and smooth flight path, driving the UAV from the current local observation point back to the interrupted "bow"-shaped inspection path. Once the UAV reaches the path point and resumes its original inspection heading, the entire adaptive cleaning process for a single ventilator is completed, and the system will continue to perform the global inspection task until the next target to be cleaned is found.

[0158] This invention provides an adaptive cleaning system for ventilators, the system comprising:

[0159] Data Acquisition Module: The drone inspects the roof containing the circular ventilator. The environmental perception module collects image information, identifies the circular ventilator, and extracts environmental benchmarks that reflect the current ambient light characteristics.

[0160] Feature extraction module: Controls the drone to approach the identified circular ventilator and extracts the geometric morphology parameters and internal dirt state parameters of the crescent-shaped vent.

[0161] Command generation module: Based on geometric morphology parameters, calculates the jet direction vector and jet area for achieving jet penetration; based on internal dirt state parameters, calculates the jet pressure for removing dirt; and generates a set of control commands containing the jet direction vector, jet area, and jet pressure along the central axis of the crescent-shaped vent.

[0162] Correction module: Executes cleaning operations based on control commands. The cleaning operations include activating the pilot jet, analyzing the image features generated by the pilot jet in real time to verify the accuracy of the spray direction vector, triggering angle compensation if an angle deviation is detected, monitoring image information during the cleaning process to evaluate the cleaning effect after confirming that the spray direction vector is accurate, and dynamically adjusting the spray pressure if the cleaning effect is not as expected.

[0163] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned adaptive cleaning method for a ventilator.

[0164] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned adaptive cleaning method for a ventilator.

[0165] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0166] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0167] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for adaptive cleaning of a ventilator, characterized in that, The cleaning method includes: The drone inspects the roof containing the circular ventilator. The environmental perception module collects image information, identifies the circular ventilator, and extracts environmental benchmarks that reflect the current ambient light characteristics. Control the drone to approach the identified circular ventilator and extract the geometric morphology parameters and internal dirt status parameters of the crescent-shaped vent. Based on the geometric parameters, the jet direction vector and jet area for achieving jet penetration are calculated. Based on the internal dirt state parameters, the jet pressure for removing dirt is calculated. A set of control commands containing the jet direction vector, jet area and jet pressure are generated along the central axis of the crescent-shaped vent. The cleaning operation is executed based on control commands. The cleaning operation includes starting the pilot jet, analyzing the image features generated by the pilot jet in real time to verify the accuracy of the spray direction vector, triggering angle compensation if an angle deviation is detected, monitoring the image information during the cleaning process to evaluate the cleaning effect after confirming that the spray direction vector is accurate, and dynamically adjusting the spray pressure if the cleaning effect is not as expected.

2. The adaptive cleaning method for a ventilator according to claim 1, characterized in that, The steps for extracting the geometric morphology parameters and internal dirt state parameters of the crescent-shaped vent include: The high-resolution image of the crescent-shaped vent is processed by minimizing the energy function E(C) to make the initial contour converge to the true inner and outer contours Cin and Cout of the crescent-shaped vent: E(C)=01g(| IHDR(C(q))|)|C'(q)|dq Where C(q) is the contour evolution curve, q is the curve parameter, IHDR is the high-resolution grayscale image, and g is the edge stopping function: g(| IHDR|)=11+β| IHDR|2 Where β is a positive parameter of the sensitivity of the control function; Based on the converged inner and outer contours Cin and Cout, the width gradient representing the geometric morphological parameters is calculated. w(s) and the internal shadow gradient Gs(x,y) characterizing the internal dirt state parameters, where s is the arc length parameter along the central axis of the crescent-shaped vent.

3. The adaptive cleaning method for a ventilator according to claim 1 or 2, characterized in that, The steps to calculate the injection direction vector Vjet(s) specifically include: calculating the local normal vectors nin(s) and nout(s) at position s on the central axis of the crescent-shaped vent, corresponding to the inner and outer contours Cin and Cout respectively; The local radial vector Nchan(s) and local tangential vector Tchan(s) of the blade passage are calculated based on the local normal vector; the final injection direction vector Vjet(s) is set as follows: Vjet(s)=cos( pitch)Tchan(s)+sin( pitch)Nchan(s) in, pitch is the pitch angle determined based on the change in the projection of the crescent-shaped vent in the vertical direction.

4. The adaptive cleaning method for a ventilator according to claim 2, characterized in that, The calculation of the jet area Ajets includes: Ajet(s)=Amin+(Amax Amen) w(s)wmax e a| w(s)| Where Amax and Amin are the maximum and minimum adjustable spray areas of the nozzle, respectively, w(s) is the opening width at arc length s, wmax is the maximum width of the crescent-shaped vent, and α is a positive attenuation coefficient. w(s) is the width gradient.

5. The adaptive cleaning method for a ventilator according to claim 2, characterized in that, The step of calculating the injection pressure Pjet(s) is achieved through a piecewise function: Pjet(s)=PbaseifGs,max(s)<TGPbase+η(Gs,max(s) TG)ifGs,max(s)≥TG Where Pbase is the basic cleaning pressure, Gs,max(s) is the maximum value of the internal shadow gradient Gs(x,y) within the local rectangular analysis window at arc length s, TG is the gradient threshold for determining whether the dirt is a stubborn agglomerate, and η is the gain coefficient.

6. A ventilator adaptive cleaning method according to claim 4 or 5, characterized in that, The step of generating a set of control commands also includes calculating a pressure modulation factor γ(s) based on geometric features and coupling it with the injection pressure Pjet(s) to obtain the final command pressure Pjet'(s): γ(s)=1 e d / | w(s)| Five'(s)=γ(s) Five(s) Where δ is a positive adjustment parameter.

7. The adaptive cleaning method for a ventilator according to claim 1, characterized in that, The step of real-time analysis of image features generated by the pilot jet to verify the accuracy of the jet direction vector specifically includes: calculating a jet antisplash flux ΦBS. ΦBS=1Δt LC∮C(s)max(0,f(l) nb(l))dl Where C(s) is the crescent-shaped vent outline corresponding to the current control point, l is the arc length parameter on the outline, LC is the total length of the outline, Δt is the time interval between image frames, f(l) is the optical flow vector at point l on the outline, and nb(l) is the unit outward normal vector of the outline at point l. When the jet backsplash flux ΦBS is not less than a preset threshold TΦ, it is determined that an angle deviation has been detected.

8. The adaptive cleaning method for a ventilator according to claim 7, characterized in that, After determining that an angle deviation has been detected, the step of triggering angle compensation also includes: calculating the splash establishment time tr and the steady-state fluctuation coefficient Cv by analyzing the dynamic characteristics of the jet splash flux ΦBS(t) in the time series; If tr is less than a time threshold Tr and Cv is less than a fluctuation threshold Tcv, then it is determined to be geometric splashing, and the optimal jet direction vector is iteratively searched with the goal of minimizing ΦBS; If tr is not less than Tr or Cv is not less than Tcv, it is determined to be physical backsplash, the current jet direction vector is locked as the optimal angle, and the current control point is marked as high-resistance blockage.

9. The adaptive cleaning method for a ventilator according to claim 1, characterized in that, The steps for dynamically adjusting the jet pressure specifically include: calculating the dirt removal rate. c(si), the dirt removal rate is defined as the average decay rate of the dirt texture energy field during a cleaning pulse of fixed duration; if the dirt removal rate is detected... c(si) is lower than an expected rate If exp(si) and the jet backsplash flux ΦBS remains low, it is determined to be a deep blockage, and the command pressure at the current control point is gradually increased by a preset increment ΔP until... c(si) recovers to the expected level or reaches the safe limit.

10. A ventilator adaptive cleaning system, characterized in that, The system includes: Data Acquisition Module: The drone inspects the roof containing the circular ventilator. The environmental perception module collects image information, identifies the circular ventilator, and extracts environmental benchmarks that reflect the current ambient light characteristics. Feature extraction module: Controls the drone to approach the identified circular ventilator and extracts the geometric morphology parameters and internal dirt state parameters of the crescent-shaped vent. Command generation module: Based on geometric morphology parameters, calculates the jet direction vector and jet area for achieving jet penetration; based on internal dirt state parameters, calculates the jet pressure for removing dirt; and generates a set of control commands containing the jet direction vector, jet area, and jet pressure along the central axis of the crescent-shaped vent. Correction module: Executes cleaning operations based on control commands. The cleaning operations include activating the pilot jet, analyzing the image features generated by the pilot jet in real time to verify the accuracy of the spray direction vector, triggering angle compensation if an angle deviation is detected, monitoring image information during the cleaning process to evaluate the cleaning effect after confirming that the spray direction vector is accurate, and dynamically adjusting the spray pressure if the cleaning effect is not as expected.

Images