Intelligent cleaning unmanned aerial vehicle system based on high-pressure water jet and method thereof

Abstract

The invention provides an intelligent cleaning unmanned aerial vehicle system and method based on high-pressure water jet, and relates to the technical field of unmanned aerial vehicles. The optimized path acquisition module utilizes a sensor to optimize obstacle avoidance to obtain a new path; the real-time positioning acquisition module realizes positioning by means of PTK; the cleaning area determination module determines a cleaning range according to the positioning acquisition image; and the water flow cleaning execution module starts equipment and adjusts a nozzle to complete high-pressure water jet cleaning. The technical problems that a traditional high-altitude cleaning mode is high in safety risk, low in cleaning efficiency, inaccurate in environment perception and insufficient in dynamic adaptability are solved, and safe unmanned cleaning, path optimization precision and cleaning efficiency maximization of high-rise building glass curtain wall cleaning are achieved; and therefore, the technical effects of operation reliability and intelligent level in a complex environment can be obviously improved.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) technology, specifically to an intelligent cleaning UAV system and method based on high-pressure water jet. Background Technology

[0002] In the field of high-rise building glass curtain wall cleaning, efficient and precise cleaning operations are crucial for building maintenance, equipment safety and environmental quality, and the automation and intelligent control of the cleaning process are key performance indicators.

[0003] In existing technologies, traditional high-altitude cleaning drones, which rely on manual operation or single sensors, exhibit significant limitations in complex environments. Firstly, manual cleaning carries the risk of falling from heights and is inefficient due to factors such as wind and curtain wall structures. Secondly, traditional drones, using only single sensors like binocular vision for obstacle avoidance, are prone to misjudging obstacles (such as open windows or curtain wall frames) in semi-open environments with glass reflections or water mist, leading to collision risks. Furthermore, their cleaning processes often follow fixed paths, making it impossible to dynamically adjust nozzle parameters based on real-time water pressure fluctuations and cleaning results. This results in uneven cleaning or water waste, and the lack of real-time monitoring and autonomous replenishment mechanisms for water tank levels and battery power makes it difficult to meet the demands of large-scale, long-term automated operations. Summary of the Invention

[0004] This application provides an intelligent cleaning drone system and method based on high-pressure water jet, which solves the technical problems of high safety risks, low cleaning efficiency, inaccurate environmental perception and insufficient dynamic adaptability in traditional high-altitude cleaning methods. It achieves safe and unmanned cleaning of high-rise building glass curtain walls, precise path optimization and maximized cleaning efficiency, thereby significantly improving the reliability and intelligence level of operations in complex environments.

[0005] In view of the above problems, on the one hand, this application provides an intelligent cleaning drone system based on high-pressure water jet, the system comprising: a cleaning area acquisition module, used to receive and parse task instructions to acquire the cleaning area and an initial cleaning path; an optimized path acquisition module, used for the drone to fly to the cleaning area according to the initial cleaning path, and to optimize the initial cleaning path by using millimeter-wave radar and binocular vision sensors to obtain an optimized cleaning path; a real-time positioning acquisition module, used to perform real-time positioning of the drone through a PTK mobile terminal and a PTK base station to obtain real-time positioning data; a cleaning area determination module, used for controlling the drone to hover through a flight control unit when the real-time positioning data reaches the cleaning position of the optimized cleaning path, to acquire images of the glass curtain wall through a camera component, and to transmit the images to an image transmission terminal for image analysis to determine the area of ​​the glass curtain wall to be cleaned; and a water flow cleaning execution module, used to activate the drone's water tank and high-pressure nozzle to perform high-pressure water jet cleaning of the glass curtain wall area, and to adjust the high-pressure nozzle for cleaning through a two-degree-of-freedom swing mechanism.

[0006] On the other hand, this application also provides a method for intelligent cleaning drones based on high-pressure water jets. The method includes: receiving and parsing a task command to obtain a cleaning area and an initial cleaning path; the drone flying to the cleaning area according to the initial cleaning path, using millimeter-wave radar and binocular vision sensors to optimize obstacle avoidance perception on the initial cleaning path to obtain an optimized cleaning path; real-time positioning of the drone through a PTK mobile terminal and a PTK base station to obtain real-time positioning data; when the real-time positioning data reaches the cleaning position of the optimized cleaning path, controlling the drone to hover through a flight control unit, acquiring images of the glass curtain wall through a camera component, transmitting them to an image transmission terminal for image parsing, and determining the area of ​​the glass curtain wall to be cleaned; activating the drone's water tank and high-pressure nozzle to perform high-pressure water jet cleaning on the area of ​​the glass curtain wall, and adjusting the high-pressure nozzle through a two-degree-of-freedom swing mechanism for cleaning.

[0007] One or more technical solutions provided in this application have at least the following technical effects: The cleaning area acquisition module receives and parses the task instructions to obtain the cleaning area and initial cleaning path; the path optimization acquisition module guides the drone to fly to the cleaning area according to the initial cleaning path, and uses millimeter-wave radar and binocular vision sensors to optimize the initial cleaning path for obstacle avoidance perception to obtain an optimized cleaning path; the real-time positioning acquisition module obtains real-time positioning data for the drone through PTK mobile terminal and base station terminal; when the real-time positioning data reaches the cleaning position of the optimized cleaning path, the cleaning area determination module controls the drone to hover through the flight control unit, uses the camera component to collect images of the glass curtain wall and transmits them to the image transmission sky terminal for parsing to determine the cleaning area; the water flow cleaning execution module starts the drone's water tank and high-pressure nozzle to perform high-pressure water jet cleaning on the cleaning area, and adjusts the high-pressure nozzle through a two-degree-of-freedom swing mechanism, achieving safe and unmanned cleaning of high-rise building glass curtain walls, precise path optimization and maximizing cleaning efficiency, thereby significantly improving the technical effect of operational reliability and intelligence level in complex environments.

[0008] In summary, this application integrates multi-sensor data during drone cleaning operations, processes it using algorithms such as millimeter-wave radar and binocular vision for obstacle distance mean processing, and cosine similarity trend analysis of cleaning feature sequences, to obtain data on obstacle types and cleaning effects. It then calculates and optimizes cleaning paths and nozzle dynamic adjustment parameters, and adjusts them in conjunction with central water tank center-of-gravity control and energy monitoring results. This allows for precise obstacle avoidance, adaptive adjustment of cleaning parameters, and fully autonomous operation in complex curtain wall environments, significantly improving the safety, efficiency, and intelligence of high-rise building glass curtain wall cleaning systems.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the structure of an intelligent cleaning drone system based on high-pressure water jet provided in an embodiment of this application.

[0012] Figure 2 This is a flowchart illustrating the intelligent cleaning drone method based on high-pressure water jet provided in an embodiment of this application.

[0013] Explanation of reference numerals in the attached diagram: Cleaning area acquisition module 1, Optimized path acquisition module 2, Real-time positioning acquisition module 3, Cleaning area determination module 4, Water flow cleaning execution module 5. Detailed Implementation

[0014] This application provides an intelligent cleaning drone system and method based on high-pressure water jets, which solves the technical problems of high safety risks, low cleaning efficiency, inaccurate environmental perception, and insufficient dynamic adaptability in traditional high-altitude cleaning methods. It achieves safe and unmanned cleaning of high-rise building glass curtain walls, precise path optimization, and maximized cleaning efficiency, thereby significantly improving the reliability and intelligence level of operations in complex environments.

[0015] Example 1, as Figure 1 As shown in the figure, this application provides an intelligent cleaning drone system based on high-pressure water jet, the system comprising: Cleaning area acquisition module 1 is used to receive and parse task instructions to obtain the cleaning area and initial cleaning path.

[0016] In this embodiment, the cleaning area is a specific area of ​​the glass curtain wall of a high-rise building to be cleaned. The initial cleaning path is a preliminary flight path of the cleaning drone generated based on the cleaning area information.

[0017] Specifically, firstly, those skilled in the art issue task instructions via a ground station or mobile terminal. The instructions include basic information such as the geographical location of the target building, the distribution range of the glass curtain wall, and the cleaning priority. The system processes the instructions in a structured manner through a built-in task parsing module. For example, it uses a Geographic Information System (GIS) to convert the address information into latitude and longitude coordinates, and combines this with the building's BIM model or historical work data to identify the specific floors, facade areas, and preliminary characteristics of obstacle distribution (such as the known location of the fixed curtain wall frame) that need to be cleaned.

[0018] Then, based on the parsed clean area information, the system calls the A* path planning algorithm to generate an initial clean path. The algorithm searches for a path using a heuristic function f(n) = g(n) + h(n), where g(n) is the actual cost from the starting point to the current node, and h(n) is the estimated cost from the current node to the destination (such as Manhattan distance or Euclidean distance). Starting from the starting point, the algorithm adds it to an open list and iteratively selects the node with the smallest f(n) in the open list as the current node, traversing its neighboring nodes: if the neighboring node is the destination, it backtracks the path; if the neighboring node is in the closed list, it skips it; if the neighboring node is in the open list, it compares the g(n) values ​​and retains the better path; if the neighboring node has not been visited, it calculates its f(n) value and adds it to the open list. This process is repeated until the destination is found or the open list is empty, ultimately generating the shortest path from the starting point to the destination. This algorithm generates an initial clean path that balances coverage efficiency and safety distance based on the geographical coordinates of the clean area and the preliminary characteristics of obstacles.

[0019] Through the above steps, the mission instructions are transformed into an executable sequence of spatial coordinates and flight trajectories, ensuring that the UAV enters the clean area in an orderly and efficient manner, laying the logical starting point and spatial benchmark for subsequent multi-sensor fusion obstacle avoidance and dynamic path optimization.

[0020] The optimized path acquisition module 2 is used for the UAV to fly to the cleaning area according to the initial cleaning path, and to perform obstacle avoidance perception optimization on the initial cleaning path using millimeter-wave radar and binocular vision sensors to obtain an optimized cleaning path.

[0021] In this embodiment, the millimeter-wave radar is an active detection sensor mounted on a drone. Its core characteristic is its ability to penetrate interference media such as glass curtain walls, reflections, and water mist to acquire accurate distance data of obstacles in real time. The binocular vision sensor is a passive image acquisition device based on the parallax principle. By simulating the human eye's line-of-sight difference, it acquires images of the environment surrounding the drone's flight path and generates obstacle perception images.

[0022] Specifically, when the UAV flies to the cleaning area along the initial cleaning path, it uses a binocular vision sensor to acquire obstacle perception images and a millimeter-wave radar to acquire obstacle radar perception distances. The obstacle perception images are analyzed to obtain obstacle types and binocular perception distances. The average of the binocular perception distances and radar perception distances is processed to obtain the target obstacle perception distance. Based on the obstacle types and target obstacle perception distances, the initial cleaning path is optimized for obstacle avoidance to obtain an optimized cleaning path. The specific steps are explained in detail in the sections on obstacle image acquisition and optimized path acquisition.

[0023] The real-time positioning acquisition module 3 is used to perform real-time positioning of the UAV through the PTK mobile terminal and the PTK base station terminal to obtain real-time positioning data.

[0024] In this embodiment of the application, PTK is short for Precision Tracking and Kinematic, which achieves centimeter-level real-time positioning of UAVs through the collaborative work of the base station and the mobile terminal.

[0025] Specifically, firstly, PTK base stations are deployed near the clean area. These base stations receive satellite signals (such as GPS and BeiDou) through high-precision antennas and calculate the current satellite signal transmission errors (including ionospheric delay, tropospheric delay, etc.) based on known fixed geographical coordinates (such as those pre-calibrated through surveying). The base stations then transmit this error data in real time to the PTK mobile terminal carried by the drone via wireless communication links (such as 4G / 5G or data radios).

[0026] During flight, the PTK mobile terminal synchronously receives satellite signals and error correction data sent by the base station. It then corrects the original positioning data using a differential positioning algorithm (such as RTK real-time dynamic technology). After acquiring the transmission error from the base station and transmitting it to the mobile terminal, the mobile terminal subtracts the satellite pseudorange (the distance between the satellite and the receiver, calculated by multiplying the time it takes for the signal to reach the receiver by the speed of light) from the error data from the base station to eliminate common error terms. This yields the centimeter-level precise coordinates (longitude, latitude, and altitude) of the drone's current position. This positioning data is fed back to the flight control unit in real time for dynamic adjustment of flight attitude and path.

[0027] For example, when a drone approaches a glass curtain wall on a certain floor, PTK real-time positioning data can accurately determine whether it has reached the preset cleaning position (error < 5cm). If the position deviates, the flight control unit will be triggered to fine-tune the motor speed to ensure that the drone hovers stably in front of the target area, providing a spatial reference for subsequent image acquisition and cleaning operations.

[0028] By using differential positioning technology between PTK base stations and mobile terminals, centimeter-level real-time positioning of drones in complex building environments has been achieved. This solves the cleaning deviation and safety risks caused by insufficient accuracy of traditional positioning technologies, and provides key technical support for closed-loop control of precise hovering, image recognition, and fixed-point cleaning.

[0029] The cleaning area determination module 4 is used to control the drone to hover through the flight control unit when the real-time positioning data reaches the cleaning position of the optimized cleaning path, and to collect images of the glass curtain wall through the camera component, transmit them to the image transmission sky terminal for image analysis, and determine the area of ​​the glass curtain wall to be cleaned.

[0030] In this embodiment, the flight control unit is the core control center of the UAV system, integrated inside the UAV fuselage. The image transmission terminal is a key component for UAV image data interaction, responsible for acquiring images and transmitting them to the ground station.

[0031] Specifically, when the real-time positioning data (centimeter-level coordinates) output by the PTK mobile terminal matches the preset cleaning position in the optimized cleaning path (error < 5cm), the flight control unit triggers a hovering command: by adjusting the speed of the quadcopter motors to compensate for wind interference, and combining the data from the inertial measurement unit (a device used to measure the three-axis attitude angles or angular rates and acceleration of an object), the drone's attitude is maintained stable (roll / pitch angle fluctuation < 2°), ensuring that the camera assembly is facing the target curtain wall area. At this time, the high-definition camera assembly (such as a 12-megapixel RGB camera) acquires images of the glass curtain wall at a rate of 25 frames per second and transmits them to the image transmission terminal for image analysis.

[0032] The image transmission terminal is equipped with an image analysis algorithm module. First, it separates the glass curtain wall and the building frame through adaptive threshold segmentation (such as the Otsu algorithm, which analyzes the two-dimensional distribution of the gray values ​​of image pixels and the average gray values ​​of their neighborhoods, calculates the probability density of the binary tuples, and determines the optimal threshold vector based on the maximum inter-class variance, thereby separating the glass curtain wall from the building frame). Then, it uses Canny edge detection to extract the curtain wall boundary contour (first, the image is denoised by Gaussian filtering, then the gradient magnitude and direction are calculated, the edges are refined by non-maximum suppression, and finally the real edges are determined by double threshold screening, thus accurately extracting the curtain wall boundary contour).

[0033] Next, semantic segmentation of the image is performed based on a deep learning model (such as U-Net). When building the U-Net model, the encoder (shrinking path) is first built, which consists of multiple convolutional blocks. Each convolutional block contains two 3×3 convolution operations (preserving the feature map size and padding with 1), followed by a ReLU activation function to introduce non-linearity, and then downsampling through 2×2 max pooling to gradually extract high-level semantic features. Then, the decoder (expanding path) is built. Each decoder first performs 2×2 deconvolution upsampling to expand the feature map size, and then splices it with the feature map after being cropped from the corresponding level of the encoder through skip connections to fuse high and low-level features. Then, it is further processed by two 3×3 convolutions and ReLU activation. Finally, a 1×1 convolution is used in the output layer to adjust the number of channels, and the Softmax activation function is used to classify each pixel to output the semantic segmentation result, realizing accurate identification and pixel coordinate labeling of stains such as dust and water stains.

[0034] Finally, by combining PTK positioning data with image pixel coordinates, a precise vector map of the cleaning area (with centimeter-level resolution) is generated, thereby determining the coordinate area from the upper left to the lower right of the glass curtain wall to be cleaned.

[0035] By using PTK centimeter-level positioning to ensure hovering accuracy, flight control dynamic stability to suppress external interference, and intelligent image analysis algorithms to accurately identify stain boundaries, a closed-loop control system of positioning-hovering-identification is formed, realizing the dynamic and precise delineation of the glass curtain wall cleaning area, providing a key prerequisite for subsequent high-pressure water jet point-to-point and efficient cleaning.

[0036] The water flow cleaning execution module 5 is used to activate the drone's water tank and high-pressure nozzle to perform high-pressure water jet cleaning on the glass curtain wall area, and to adjust the high-pressure nozzle for cleaning through a two-degree-of-freedom swing mechanism.

[0037] In this embodiment, the two-degree-of-freedom swing mechanism is driven by two sets of micro servo motors, which support the high-pressure nozzle to swing independently in two orthogonal directions, forming a two-dimensional spatial motion degree of freedom.

[0038] Specifically, the real-time water pressure of the water tank is first obtained through the real-time water pressure acquisition unit. Then, with the goal of improving the spray stability, the real-time spray parameter acquisition unit dynamically adjusts the spray parameters based on the water pressure to obtain real-time values. Finally, the parameters are transmitted to the microcontroller through the real-time spray parameter transmission unit, which controls the micro servo motor to drive the two-degree-of-freedom swing mechanism to adjust the high-pressure nozzle, thereby completing the cleaning. The specific steps are explained in detail in the section from the real-time water pressure acquisition unit to the real-time spray parameter transmission unit.

[0039] In one possible implementation, the optimized path acquisition module 2 further includes: An obstacle image acquisition unit is used to obtain obstacle perception images by using a binocular vision sensor to perceive the surrounding image while the UAV is flying along the initial cleaning path; an obstacle distance acquisition unit is used to obtain obstacle radar perception distance by using millimeter-wave radar; an obstacle type acquisition unit is used to perceive the obstacle type and location in the obstacle perception image to obtain the obstacle type and the binocular obstacle perception distance; a target distance acquisition unit is used to perform averaging processing on the binocular obstacle perception distance and the obstacle radar perception distance to obtain the target obstacle perception distance; and an optimized path acquisition unit performs obstacle avoidance perception optimization on the initial cleaning path based on the obstacle type and the target obstacle perception distance to obtain an optimized cleaning path.

[0040] Specifically, firstly, when the drone flies along the initial cleaning path, the binocular vision sensor acquires images of the surrounding environment at a rate of 25 frames per second. The binocular parallax algorithm extracts feature points (such as SIFT or ORB features, i.e., scale-invariant feature transformation and acceleration robust features) from the left and right images and matches corresponding points. The pixel-level parallax (i.e., the difference between the horizontal and vertical coordinates of the corresponding points in the left and right images) is calculated. Then, based on the camera baseline distance and focal length parameters, the depth value of each pixel is calculated using the principle of triangulation, generating an obstacle perception image containing the outline of obstacles (such as identifying the rectangular boundary of a curtain wall opening window).

[0041] Meanwhile, the millimeter-wave radar continuously transmits high-frequency millimeter-wave signals. When the signal encounters an obstacle, it reflects to form an echo. The radar measures the time difference between the transmitted signal and the echo (with nanosecond precision) and combines this with the electromagnetic wave propagation speed (approximately 3 × 10⁻⁶ m / s²). 8 The radar detection distance of obstacles is calculated using the formula (distance / time difference of light speed) to achieve accurate distance measurement of obstacles.

[0042] For image data output from binocular vision, semantic segmentation is performed using a convolutional neural network (such as YOLOv5, whose core function is environmental perception, i.e., obstacle classification and preliminary ranging) to identify obstacle types (such as fixed frames, openable windows, and temporary obstacles), and the binocular perception distance of obstacles is calculated based on the parallax principle. In the model construction phase, for the obstacle type recognition task, a YOLOv5 convolutional neural network architecture is adopted. Its backbone network extracts multi-scale features through CSPDarknet (Cross Stage Partial Darknet), the neck uses an FPN+PAN (Feature Pyramid Network + Path Aggregation Network) structure to fuse features from different levels, and the head uses an anchor box mechanism to classify and locate obstacles (fixed frames, openable windows, and temporary obstacles). The model input is a 1024×1024 resolution RGB image output from the binocular vision sensor. After data augmentation (random flipping, brightness adjustment) and normalization preprocessing, it is input into the network. Semantic features are extracted through multi-layer convolution and pooling operations, and the final output includes prediction results containing obstacle category probabilities and bounding box coordinates. In the distance calculation stage, based on the calibration parameters of the binocular camera (baseline distance, focal length), disparity is calculated for the center pixels of the obstacle bounding box output by YOLOv5. By matching the difference between the horizontal and vertical coordinates (disparity d) of corresponding points in the left and right images, the binocular perception distance of the obstacle is calculated using the triangulation formulas: distance from baseline distance, focal length, and disparity. Finally, the binocular perception distance (1.1 meters) and the radar perception distance (1.2 meters) of the same obstacle are averaged to obtain the target obstacle perception distance (1.15 meters). Data fusion reduces the error of a single sensor, effectively suppresses single sensor noise, and improves ranging accuracy.

[0043] For obstacle type and target distance, an obstacle avoidance strategy neural network model is invoked. It adopts a multi-layer fully connected network (MLP) architecture. The construction and training process is as follows: First, the input layer is defined as a structured feature vector, which contains one-hot encoding of obstacle type (e.g., 3 types of obstacles correspond to 3-dimensional input) and target obstacle perception distance (1-dimensional value). After passing through two hidden layers (64 neurons in each layer, with ReLU activation function), feature fusion and nonlinear transformation are performed. The output layer is mapped to continuous control quantity (e.g., 2-dimensional vectors such as motor speed difference, yaw angle adjustment value, etc.). The training data comes from a multimodal simulation and experimental fusion dataset: a virtual scene containing different curtain wall obstacles (fixed frame, operable window, temporary obstacle) is constructed using the Gazebo simulation platform to simulate the sensor input (type label, distance value) when the UAV approaches the obstacle at a speed of 0.5-5m / s, and manually designed safe obstacle avoidance actions (such as lateral offset of 0.6-1.0m for dynamic obstacles and safe distance of 0.3-0.7m for static obstacles) are used as labels; at the same time, more than 1,000 sets of effective obstacle avoidance data (trajectory segments verified by millimeter-wave radar and binocular vision sensors to have no collision risk) are collected during actual flight to form a sample training set.

[0044] The training process employs a hybrid strategy of supervised learning and reinforcement learning. Initially, the mean squared error (MSE) loss function is used to calculate the average of the squared differences between the model's output control quantity and the manually designed labels. Backpropagation updates the network parameters, keeping the deviation within 5%, thus achieving basic obstacle avoidance strategy fitting for known obstacle types and distances. Later, the Proximal Policy Optimization (PPO) reinforcement learning algorithm is introduced, using the reciprocal of the number of collisions and path efficiency as reward functions. The truncation probability ratio is calculated using sampled data to limit the policy update magnitude. The model undergoes 200 iterations of iterative training in a simulation environment, enabling it to autonomously optimize obstacle avoidance actions based on real-time reward signals, improving its generalization adaptability to unknown obstacle layouts. To adapt to the embedded computing power limitations of UAVs, the model undergoes pruning (preserving key connections) and quantization (converting floating-point to 8-bit fixed-point), keeping the inference latency within 10ms. Finally, real-time deployment is achieved on the flight control platform, ensuring that optimized flight control commands are output within one control cycle (50ms) after receiving sensor input.

[0045] By following the steps above, the risk of drone collisions is significantly reduced, while avoiding redundancy in cleaning paths due to misjudgment, thereby improving operational efficiency and safety.

[0046] In one possible implementation, the water flow cleaning execution module 5 further includes: The image sequence acquisition unit is used to continuously acquire images of the cleaned glass curtain wall area through the camera component to obtain a clean image sequence; the feature sequence acquisition unit is used to extract features from the clean image sequence according to preset cleaning indicators to obtain a clean feature sequence; the clean trend recognition unit is used to recognize the clean trend of the clean feature sequence and make feedback adjustments to the high-pressure nozzles according to the target trend cleaning features.

[0047] In this embodiment, the preset cleaning index is a quantitative evaluation standard predefined based on the cleaning quality requirements of glass curtain walls. The target trend cleaning feature is a comprehensive feature vector obtained by performing trend enhancement processing on the cleaning feature sequence.

[0048] Specifically, when the high-pressure nozzles perform water jet cleaning on the glass curtain wall area via a two-degree-of-freedom swing mechanism, the onboard high-definition camera components (such as a 12-megapixel RGB camera with a frame rate of 25 frames per second) continuously acquire cleaning image sequences at a frequency synchronized with the nozzle swing, ensuring that the state of each sprayed area before and after is recorded (e.g., the grayscale value of a certain curtain wall glass is 30±5 before cleaning, and dynamically changes to 80±10 during cleaning). The image data is transmitted to the ground station in real time via a 5.8GHz image transmission link, and then enters the feature extraction stage of the preset cleaning indicators.

[0049] The preset cleaning index is constructed based on a stain detection algorithm: First, the cleaning area and background are separated by adaptive threshold segmentation (such as the Otsu algorithm), and stain boundaries are extracted using Canny edge detection. Second, a pre-trained stain recognition model (such as an improved U-Net network) is used to perform semantic segmentation on each frame of the image. Based on the semantic segmentation results, parameters defined by the preset cleaning index are calculated, including time-varying parameters such as stain area ratio (e.g., decreasing from 15% to 5% before cleaning), grayscale mean change (reflecting stain concentration), and edge blurring (characterizing the uniformity of water jet coverage), thus forming a single-frame cleaning feature vector. The single-frame feature vectors are arranged in chronological order (each 200ms) to generate a cleaning feature sequence, reflecting the dynamic changes in cleaning progress (e.g., the area ratio gradually decreases from 15% to 5% over 5 consecutive frames). Finally, the real-time parameters in the feature sequence are compared with the preset cleaning index to determine whether the current area has reached the preset termination conditions such as no visible stains and grayscale values ​​close to the cleaning standard (the above algorithm process is the same as described in the cleaning area determination module 4, and will not be repeated due to space limitations in the manual).

[0050] Next, the first and second cleaning features are extracted from the cleaning feature sequence. The similarity between the two is calculated using the cosine similarity function and normalized to construct the first trend interaction matrix. The second cleaning feature is then enhanced using this matrix to obtain the first trend cleaning feature. Based on this, the remaining cleaning features in the sequence are enhanced sequentially to obtain the target trend cleaning feature. If the target trend cleaning feature does not meet the preset cleaning requirements, the high-pressure nozzle is adjusted accordingly. The specific steps are explained in detail in the cleaning feature extraction subunit to the target cleaning feature judgment subunit.

[0051] Through real-time image sequence feature processing and closed-loop control, the automation level and operation quality of glass curtain wall cleaning have been significantly improved. While reducing water waste, it avoids the cost of manual re-inspection and provides key technical support for the intelligent upgrade of high-pressure water jet cleaning systems.

[0052] In one possible implementation, the cleanliness trend recognition unit further includes: The system includes a cleaning feature extraction subunit for extracting a first cleaning feature and a second cleaning feature from the cleaning feature sequence; a trend matrix construction subunit for calculating the similarity between the first and second cleaning features using a cosine similarity function and normalizing the calculation results to construct a first trend interaction matrix; a first cleaning feature acquisition subunit for enhancing the trend of the second cleaning feature using the first trend interaction matrix to obtain a first trend cleaning feature; a target cleaning feature acquisition subunit for sequentially enhancing the trend of the remaining cleaning features in the cleaning feature sequence based on the first trend cleaning feature to obtain a target trend cleaning feature; and a target cleaning feature judgment subunit for judging whether the target trend cleaning feature meets the preset cleaning requirements. If not, the high-pressure nozzle is adjusted based on the target trend cleaning feature.

[0053] In this embodiment, the first trend interaction matrix is ​​a two-dimensional matrix used to quantify the trend correlation between adjacent features in the cleaning feature sequence. Trend enhancement is a process of optimizing the cleaning feature sequence based on the first trend interaction matrix.

[0054] Specifically, after the cleaning feature sequence (including time-varying parameters such as stain area percentage and grayscale mean change) is generated, the feature vectors of two adjacent frames are first extracted in chronological order, defined as the first cleaning feature (feature vector at time t, such as [12% area percentage, 75 grayscale mean]) and the second cleaning feature (feature vector at time t+Δt, such as [10% area percentage, 80 grayscale mean]). These two feature vectors are treated as points in a multi-dimensional space, and their directional consistency is calculated using a cosine similarity function (formula: ...). , Indicates the first cleaning characteristic, The result is normalized (mapped to the [0,1] interval) to construct the first trend interaction matrix (such as a 2×2 matrix, where the element values ​​reflect the trend similarity of adjacent features in dimensions such as stain area and gray level).

[0055] Next, the second cleaning feature is enhanced using the first trend interaction matrix: the matrix elements are used as weights to weight the parameters of each dimension of the second cleaning feature (e.g., if the similarity of area proportion is 0.9 and the similarity of grayscale mean is 0.8, the enhanced area proportion trend contribution increases by 10%), suppressing random noise (e.g., single-frame grayscale abrupt changes caused by light reflection), highlighting the true cleaning trend (e.g., the stability of continuous area decrease), and generating the first trend cleaning feature, for example [10% × 1.1 area proportion, 80 × 1.05 grayscale mean]. The weight coefficients are dynamically adjusted by the matrix values. Then, based on the first trend cleaning feature, subsequent features are enhanced sequentially according to the same logic, forming a target trend cleaning feature that includes temporal correlation enhancement (e.g., parameters such as the weighted average decrease rate within a 5-frame sliding window and trend stability).

[0056] Finally, this feature is compared with preset cleaning requirements (such as the area ratio decrease rate after three consecutive frames of trend enhancement > 2% / frame and grayscale mean fluctuation < 5%). If it is not met (such as the area decrease rate after trend enhancement being only 1% / frame, indicating insufficient cleaning efficiency), a feedback adjustment mechanism is triggered: the microcontroller sends instructions to control the servo motor of the high-pressure nozzle to increase the water pressure (such as from 10MPa to 15MPa) or reduce the nozzle diameter (such as from 5mm to 3mm to enhance the impact force), while adjusting the swing frequency of the two-degree-of-freedom swing mechanism (such as from 1Hz to 1.5Hz to increase the coverage density).

[0057] By constructing a trend interaction matrix using cosine similarity, the temporal correlation model of cleaning feature sequences was realized. Based on the target trend features, dynamic feedback control was formed, enabling the system to accurately adjust nozzle parameters according to the dynamic evolution of the cleaning process, thereby improving the intelligence level and resource utilization efficiency of glass curtain wall cleaning.

[0058] In one possible implementation, the water flow cleaning execution module 5 further includes: The system includes a real-time water pressure acquisition unit for acquiring the real-time water pressure of the water tank; a real-time spray parameter acquisition unit for dynamically adjusting the spray parameters of the high-pressure nozzle based on the real-time water pressure to improve the spray stability of the high-pressure nozzle, thereby obtaining real-time spray parameters; and a real-time spray parameter transmission unit for transmitting the real-time spray parameters to a microcontroller, which then controls a micro servo motor to drive a two-degree-of-freedom swing mechanism to adjust the high-pressure nozzle.

[0059] In this embodiment, the microcontroller is an integrated microcomputer chip.

[0060] Specifically, when the drone starts the water tank and high-pressure nozzle for cleaning, it uses a built-in pressure sensor (such as a MEMS pressure sensor with an accuracy of ±1%FS) to collect water pressure data in the water tank in real time and outputs it to the flight control unit at a frequency of 100Hz. This data simultaneously reflects the remaining water volume in the water tank and changes in pipeline resistance (such as local pressure increase caused by water pipe bends), providing a basis for adjusting the spray parameters.

[0061] Next, with the goal of improving the spray stability of the high-pressure nozzle, it is first determined whether the real-time water pressure exceeds the preset stable water pressure threshold determined based on the nozzle type and spray stability requirements. If it exceeds, a spray parameter adjustment command is obtained. Then, the real-time water pump flow rate and real-time nozzle diameter are adjusted according to the command and the real-time water pressure to obtain the real-time spray parameters. The specific steps are explained in detail in the real-time water pressure judgment and spray parameter acquisition subunit.

[0062] Then, the parameters are transmitted to a microcontroller (such as the STM32H7 series) using the CAN bus or UART protocol. After parsing, the microcontroller outputs a PWM signal to drive the micro servo motor of the two-degree-of-freedom swing mechanism. The servo motor is linked to the nozzle bracket through a gear set, adjusting the nozzle pitch and yaw angles within a range of ±45°, so that the water jet direction is always perpendicular to the glass curtain wall surface, while coordinating with water pressure parameters to achieve uniform coverage of a 3m×2m area.

[0063] Through real-time sensing by pressure sensors, dynamic calculation of multivariable parameter adjustment models, and precise control of servo motors, the high-pressure nozzles maintain stable spray pressure, coverage, and impact angle in complex operating environments, significantly improving the uniformity and efficiency of glass curtain wall cleaning, while reducing water waste and repetitive cleaning costs caused by parameter drift.

[0064] In one possible implementation, the real-time injection parameter acquisition unit further includes: The real-time water pressure judgment subunit is used to determine whether the real-time water pressure exceeds a preset stable water pressure threshold. If so, it obtains a jet parameter adjustment command, wherein the preset stable water pressure threshold is determined based on the type of high-pressure nozzle and the jet stability requirements. The jet parameter acquisition subunit is used to adjust the real-time water pump flow rate and the real-time nozzle diameter based on the real-time water pressure according to the jet parameter adjustment command to obtain real-time jet parameters.

[0065] Specifically, when operating an intelligent cleaning drone system based on high-pressure water jets, the real-time water pressure in the water tank is first monitored and assessed. A preset stable water pressure threshold is determined by those skilled in the art based on the type of high-pressure nozzle (e.g., fan-shaped nozzles, cone nozzles, etc., the structure and design of different nozzle types determine the applicable water pressure range) and the requirements for stable spraying (e.g., requirements for water jet intensity and coverage stability). When the real-time water pressure exceeds this threshold, the system determines that adjustment is needed and thus obtains spray parameter adjustment instructions.

[0066] Then, the injection parameter acquisition subunit adjusts the real-time water pump flow rate and real-time nozzle diameter according to the adjustment command and in combination with the specific value of real-time water pressure, and obtains the real-time injection parameters. The specific steps are explained in detail in the regulator construction microunit and the injection parameter output microunit.

[0067] By determining the preset water pressure threshold based on the nozzle type and spray stability requirements, the system can judge the water pressure in real time and obtain instructions when it exceeds the threshold. Based on this, the water pump flow and nozzle diameter can be adjusted. This achieves the effect of dynamically adjusting the spray parameters according to the real-time water pressure and improving the spray stability of the high-pressure nozzle, ensuring that the cleaning operation is carried out efficiently and stably.

[0068] In one possible implementation, the injection parameter acquisition subunit further includes: A regulator construction micro-unit is used to pre-build a dynamic regulator, wherein the dynamic regulator is built based on a feedforward neural network and includes an input layer, a convolutional layer, and an output layer; a jet parameter output micro-unit is used to transmit the real-time water pressure, real-time water pump flow rate, and real-time nozzle diameter to the dynamic regulator through the input layer, analyze them in the convolutional layer, and output the real-time jet parameters through the output layer.

[0069] In this application embodiment, the dynamic regulator is a device or system that can adjust relevant parameters or outputs in real time according to the system operating status, environmental changes or set requirements, so as to achieve specific goals such as stable performance and optimized efficiency.

[0070] Specifically, in constructing a dynamic regulator based on a feedforward neural network, the network architecture and parameter initialization are performed first. The input layer has three neurons, corresponding to the three key physical quantities: real-time water pressure, real-time pump flow rate, and real-time nozzle diameter. Each input neuron is accompanied by a normalization function to map the raw data to the [0,1] interval to improve model stability. The convolutional layer adopts a 1D convolutional structure, containing three convolutional kernels (1×1, 1×2, and 1×3 in size). Each convolutional kernel is followed by a ReLU activation function to introduce nonlinear characteristics. A sliding window mechanism is used to extract local correlation features between input parameters (such as the time correlation between the rate of change of water pressure and the flow rate adjustment). The output layer is designed with two neurons, corresponding to the pump flow rate adjustment coefficient and the nozzle diameter adjustment command to be adjusted. The tanh activation function is used to constrain the output value to the [-1,1] interval for easy subsequent transformation. The network weight parameters use the Xavier initialization method (a technique used in deep learning to randomly initialize neural network weights), and the bias term is initialized to 0.01 to avoid the gradient vanishing problem.

[0071] Next, model training and optimization were performed. The training dataset was collected from 500 cleaning experiments under different working conditions. Each experiment recorded 100 sets of time-series data (including water pressure, flow rate, nozzle diameter, and corresponding cleaning effect scores), which were divided into training and validation sets in an 8:2 ratio. The Adam optimizer was used during training, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. The loss function was designed as a weighted sum of MSE (mean squared error) and a trend consistency penalty term, where the trend consistency penalty term was calculated by the cosine similarity between the predicted parameter change direction and the actual optimal direction (additional loss was added if the directions were opposite). To prevent overfitting, L2 regularization (weight decay coefficient 0.0001) and Dropout (dropout rate 0.2) strategies were introduced. Every 10 training epochs, the model performance was evaluated on the validation set. When the validation loss no longer decreased for 5 consecutive epochs, an early stopping mechanism was triggered, and the current optimal model parameters were saved. The final trained dynamic regulator has a parameter prediction error of less than 3% on the test set and can accurately capture the nonlinear mapping relationship between water pressure, flow rate and nozzle diameter.

[0072] Next, key data such as real-time water pressure, real-time pump flow rate, and real-time nozzle diameter are input into the dynamic controller. After receiving the data, the input layer performs in-depth analysis to uncover the implicit correlation features between various parameters (such as the optimal matching pattern between pump flow rate and nozzle diameter under different water pressures). After feature extraction and computation by the convolutional layer, the output layer outputs real-time spray parameters based on the analysis results, completing the intelligent derivation from input data to target parameters.

[0073] By constructing a dynamic regulator based on a feedforward neural network, and utilizing the data reception of its input layer, the feature analysis of its convolutional layer, and the result generation process of its output layer, the regulator achieves the effect of accurately adjusting and outputting real-time spray parameters based on real-time water pressure, water pump flow rate, and nozzle diameter. This improves the intelligence and accuracy of parameter adjustment, enabling it to better adapt to complex and ever-changing cleaning operation scenarios and ensuring the stability of high-pressure nozzle spray.

[0074] In one possible implementation, the water flow cleaning execution module 5 further includes: The water tank adopts a centrally located structure.

[0075] Optionally, the water tank can be integrated into the central symmetrical plane of the drone's fuselage, aligning its center of gravity with the drone's center of mass, thereby significantly reducing attitude compensation energy consumption. The water tank is made of lightweight, high-strength composite materials (such as lightweight carbon fiber reinforced epoxy resin), with a wall thickness controlled at 2-3mm to meet the 40L capacity requirement. The water tank's interior features a cross-shaped baffle with 5mm diameter damping holes. When the water flows and sloshes, the baffle generates viscous resistance through the damping holes, converting the sloshing energy into heat dissipation and effectively suppressing liquid sloshing.

[0076] Then, a center of gravity compensation module is embedded in the flight control unit. The attitude angular acceleration change rate is monitored in real time through the IMU (Inertial Measurement Unit) sensor. When a small attitude disturbance caused by water flow sloshing is detected (such as pitch angular acceleration > 0.2° / s²), the output of the quadcopter motor is automatically adjusted (such as increasing the thrust of the front rotor by 3%) to achieve active stability compensation.

[0077] By combining a centrally located water tank layout, an internal damping structure, and active center of gravity compensation, the problem of center of gravity shift and liquid sloshing caused by traditional water tank arrangements has been overcome. This allows the UAV to maintain good stability and maneuverability under different load conditions, significantly reducing flight energy consumption and improving operational safety and efficiency in complex environments.

[0078] In one possible implementation, the water flow cleaning execution module 5 further includes: The water and electricity acquisition unit is used to monitor the battery power and water tank volume of the drone in real time to obtain the real-time battery power and water tank volume; the return command acquisition unit is used to acquire a return command when the real-time battery power is lower than a preset power threshold and / or the real-time water tank volume is lower than a preset water volume threshold, and the flight control unit receives the return command to control the drone to return for charging and / or water filling.

[0079] Specifically, firstly, the water and electricity acquisition unit employs multi-sensor fusion technology to monitor energy status in real time. Battery power monitoring uses the MAX17043 fuel gauge chip, which collects battery voltage, current, and temperature data every 50ms via the I²C interface. Combining coulomb counting with a neural network algorithm (the algorithm process is the same as described above and will not be repeated here due to space limitations), the power estimation error is controlled within ±2%. Water tank level monitoring uses an ultrasonic level sensor (such as the HC-SR04), installed on the top of the tank. It calculates the water level by emitting 40kHz ultrasound and measuring the echo time. After temperature compensation, the accuracy reaches ±3mm, which is then converted into a water volume value based on the tank's cross-sectional area.

[0080] When the real-time battery power is lower than a preset power threshold (e.g., 30%) or the real-time water tank water level is lower than a preset water level threshold (e.g., 15%), the return command acquisition unit triggers a decision-making process. First, the water and power data are smoothed using a Kalman filter to eliminate occasional noise interference (e.g., voltage fluctuations at the moment of motor startup). Then, an evaluation is performed based on the current work progress (e.g., the percentage of area already cleaned) and the energy consumption of the return path (dynamically calculated based on parameters such as real-time wind speed and load). If continuing the work would result in insufficient remaining power to support the return to base (e.g., calculations show that a 25% power consumption is required for the return, while the current power is only 28%), a return command is immediately generated. The command includes the optimal return path (prioritizing downwind routes to reduce energy consumption) and charging / watering priorities (e.g., prioritizing charging when the power is lower than the water level threshold).

[0081] After receiving the return command, the flight control unit first uses GPS to determine its current location and base coordinates, and calculates the optimal return route (considering terrain obstacles, no-fly zones, etc.). During the return process, an adaptive PID control algorithm dynamically adjusts flight parameters. By calculating the deviation between the target value (such as preset flight speed and roll angle) and the actual value in real time, the deviation is quickly responded to by the proportional (P) stage, the cumulative error is eliminated by the integral (I) stage, and the deviation change trend is predicted by the derivative (D) stage. The outputs of these three are linearly combined into a control quantity to dynamically adjust the actuators (such as motor speed and servo angle). In the flight control return scenario, the adaptive PID also automatically adjusts the proportional coefficient, integral time constant, and derivative time constant according to real-time parameters such as battery level and wind speed. This allows the UAV to maintain stable flight parameters through precise adjustments even when the battery is low and the speed is reduced or when there is crosswind interference, thus achieving a safe return. For example, when the battery level drops below 15%, the flight speed is automatically reduced and non-essential equipment (such as auxiliary lighting) is turned off; when a crosswind speed greater than 5 m / s is detected, roll angle compensation is increased (0.5° roll angle for every m / s wind speed) to maintain straight flight. Simultaneously, real-time status information (such as remaining battery power and estimated arrival time) is sent to the ground station, which can then manually intervene to adjust the return strategy.

[0082] Through the above steps, a complete intelligent endurance support system has been built, enabling drones to reliably complete tasks and return safely even in complex environments, significantly improving the continuity of cleaning operations and equipment reliability.

[0083] In summary, the intelligent cleaning drone system based on high-pressure water jet provided in this application has the following technical effects: This application's embodiments comprehensively improve the stability, intelligence, and safety of drone cleaning operations by employing a centrally located water tank structure design for a high-pressure water jet-based intelligent cleaning drone system. This is achieved through dynamic characteristic analysis and feedback adjustment of the cleaning process, adaptive control of water pressure and jet parameters, and intelligent monitoring and return-to-home strategies for water and electricity consumption. It effectively solves problems in traditional designs such as high flight energy consumption due to center-of-gravity shift, delayed evaluation of cleaning effects, fluctuations in jet parameters, and the risk of energy depletion. The system accurately captures changes in water pressure, stain residue trends, and equipment status during the cleaning process, adjusting nozzle parameters and flight strategies in real time to reduce resource consumption while ensuring cleaning quality. Furthermore, intelligent endurance management and dynamic stability control ensure reliable operation and safe return of the drone in complex environments, significantly improving the automation level of urban glass curtain wall cleaning and infrastructure maintenance efficiency.

[0084] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a method for intelligent cleaning drones based on high-pressure water jets, the method comprising: Step A100: Receive and parse the task instructions to obtain the cleaning area and initial cleaning path.

[0085] Step A200: The UAV flies to the cleaning area according to the initial cleaning path, and uses millimeter-wave radar and binocular vision sensors to optimize the obstacle avoidance perception of the initial cleaning path to obtain an optimized cleaning path.

[0086] Step A300: Real-time positioning of the drone is performed through the PTK mobile terminal and PTK base station to obtain real-time positioning data.

[0087] Step A400: When the real-time positioning data reaches the cleaning position of the optimized cleaning path, the drone is controlled by the flight control unit to hover, and the camera component is used to acquire images of the glass curtain wall. The images are then transmitted to the image transmission terminal for analysis to determine the area of ​​the glass curtain wall to be cleaned.

[0088] Step A500: Start the drone's water tank and high-pressure nozzle to clean the glass curtain wall area with high-pressure water jets, and adjust the high-pressure nozzles for cleaning through a two-degree-of-freedom swing mechanism.

[0089] Furthermore, step A200 in the embodiments of this application also includes: While the UAV flies along the initial cleaning path, it uses a binocular vision sensor to perceive the surrounding image and obtain an obstacle perception image; it uses millimeter-wave radar to perceive the obstacle distance and obtain the obstacle radar perception distance; it performs obstacle type and location perception on the obstacle perception image to obtain the obstacle type and the binocular obstacle perception distance; it performs averaging processing on the obstacle binocular perception distance and the obstacle radar perception distance to obtain the target obstacle perception distance; and it optimizes the initial cleaning path based on the obstacle type and the target obstacle perception distance to obtain an optimized cleaning path.

[0090] Furthermore, step A500 in this embodiment of the application also includes: The camera module continuously acquires images of the glass curtain wall area to obtain a cleaning image sequence; features are extracted from the cleaning image sequence according to preset cleaning indicators to obtain a cleaning feature sequence; the cleaning feature sequence is used to identify the cleaning trend, and the high-pressure nozzles are adjusted based on the target trend cleaning features.

[0091] Furthermore, step A500 in this embodiment of the application also includes: First and second cleaning features are extracted from the cleaning feature sequence; the similarity between the first and second cleaning features is calculated using a cosine similarity function, and the calculation result is normalized to construct a first trend interaction matrix; the second cleaning feature is enhanced using the first trend interaction matrix to obtain a first trend cleaning feature; and so on, the remaining cleaning features in the cleaning feature sequence are enhanced sequentially based on the first trend cleaning feature to obtain a target trend cleaning feature; it is determined whether the target trend cleaning feature meets the preset cleaning requirements, and if not, the high-pressure nozzle is adjusted based on the target trend cleaning feature.

[0092] Furthermore, step A500 in this embodiment of the application also includes: The system obtains the real-time water pressure of the water tank; with the goal of improving the spray stability of the high-pressure nozzle, it dynamically adjusts the spray parameters of the high-pressure nozzle based on the real-time water pressure to obtain real-time spray parameters; the real-time spray parameters are transmitted to the microcontroller, and the microcontroller controls the micro servo motor to drive the two-degree-of-freedom swing mechanism to adjust the high-pressure nozzle.

[0093] Furthermore, step A500 in this embodiment of the application also includes: Determine whether the real-time water pressure exceeds a preset stable water pressure threshold. If so, obtain a jet parameter adjustment command, wherein the preset stable water pressure threshold is determined based on the type of high-pressure nozzle and jet stability requirements. Based on the real-time water pressure, adjust the real-time water pump flow rate and the real-time nozzle diameter according to the jet parameter adjustment command to obtain real-time jet parameters.

[0094] Furthermore, step A500 in this embodiment of the application also includes: A pre-built dynamic regulator is constructed based on a feedforward neural network, comprising an input layer, a convolutional layer, and an output layer. The real-time water pressure, real-time water pump flow rate, and real-time nozzle diameter are transmitted to the dynamic regulator through the input layer, analyzed in the convolutional layer, and the real-time injection parameters are output through the output layer.

[0095] Furthermore, step A500 in this embodiment of the application also includes: the water tank adopts a centrally located structure.

[0096] Furthermore, step A500 in this embodiment of the application also includes: The system monitors the battery level and water tank level of the drone in real time to obtain these data. When the real-time battery level is lower than a preset power threshold and / or the real-time water tank level is lower than a preset water level threshold, a return command is obtained. The flight control unit receives the return command and controls the drone to return for charging and / or water replenishment.

[0097] Through the foregoing detailed description of the intelligent cleaning drone system based on high-pressure water jet, those skilled in the art can clearly understand that the intelligent cleaning drone method based on high-pressure water jet in this embodiment corresponds to the system disclosed in Embodiment 1, and has corresponding execution steps and technical effects. For relevant details, please refer to the system section description.

[0098] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart cleaning drone system based on high-pressure water jet, characterized in that, The system includes: The cleaning area acquisition module is used to receive and parse task instructions to obtain the cleaning area and the initial cleaning path; The optimized path acquisition module is used to allow the UAV to fly to the cleaning area according to the initial cleaning path, and to perform obstacle avoidance perception optimization on the initial cleaning path using millimeter-wave radar and binocular vision sensors to obtain an optimized cleaning path. The real-time location acquisition module is used to perform real-time location acquisition of the UAV through PTK mobile terminal and PTK base station terminal, and obtain real-time location data; The cleaning area determination module is used to control the drone to hover through the flight control unit when the real-time positioning data reaches the cleaning position of the optimized cleaning path, and to collect images of the glass curtain wall through the camera component, transmit them to the image transmission sky terminal for image analysis, and determine the area of ​​the glass curtain wall to be cleaned. The water jet cleaning execution module is used to activate the drone's water tank and high-pressure nozzle to perform high-pressure water jet cleaning on the glass curtain wall area, and to adjust the high-pressure nozzle for cleaning through a two-degree-of-freedom swing mechanism.

2. The intelligent cleaning drone system based on high-pressure water jet as described in claim 1, characterized in that, The drone flies to the cleaning area according to the initial cleaning path, and uses millimeter-wave radar and binocular vision sensors to optimize the obstacle avoidance perception of the initial cleaning path to obtain an optimized cleaning path, including: The obstacle image acquisition unit is used to obtain obstacle perception images by using a binocular vision sensor to perceive the surrounding images while the UAV is flying along the initial cleaning path. The obstacle distance acquisition unit uses millimeter-wave radar to perceive the obstacle distance and obtain the obstacle radar sensing distance. An obstacle type acquisition unit is used to perform obstacle type and location perception on the obstacle perception image to obtain the obstacle type and the obstacle binocular perception distance; The target distance acquisition unit is used to perform average processing on the obstacle binocular perception distance and the obstacle radar perception distance to obtain the target obstacle perception distance; The path acquisition optimization unit optimizes the initial cleaning path based on the obstacle type and the target obstacle perception distance to obtain an optimized cleaning path.

3. The intelligent cleaning drone system based on high-pressure water jet as described in claim 1, characterized in that, The system activates the drone's water tank and high-pressure nozzles to perform high-pressure water jet cleaning on the glass curtain wall area, and adjusts the high-pressure nozzles using a two-degree-of-freedom swing mechanism for cleaning. It also includes: An image sequence acquisition unit is used to continuously acquire images of the cleaned glass curtain wall area through a camera component, and acquire a clean image sequence. The feature sequence acquisition unit is used to extract features from the cleaning image sequence according to a preset cleaning index to obtain a cleaning feature sequence; The cleaning trend identification unit is used to identify the cleaning trend of the cleaning feature sequence and make feedback adjustments to the high-pressure nozzle based on the target trend cleaning feature.

4. The intelligent cleaning drone system based on high-pressure water jet as described in claim 3, characterized in that, Cleaning trend identification is performed on the cleaning feature sequence, and the high-pressure nozzle is adjusted based on the target trend cleaning features, including: A cleaning feature extraction subunit is used to extract a first cleaning feature and a second cleaning feature from the cleaning feature sequence; The trend matrix is ​​constructed into a sub-unit. The similarity between the first clean feature and the second clean feature is calculated using the cosine similarity function, and the calculation results are normalized to construct the first trend interaction matrix. The first cleaning feature acquisition subunit uses the first trend interaction matrix to enhance the trend of the second cleaning feature to obtain the first trend cleaning feature; The target cleaning feature acquisition sub-unit is used, and so on. Based on the first trend cleaning feature, the remaining cleaning features in the cleaning feature sequence are sequentially enhanced to obtain the target trend cleaning feature. The target cleaning feature judgment subunit is used to determine whether the target trend cleaning feature meets the preset cleaning requirements. If not, the high-pressure nozzle is adjusted based on the target trend cleaning feature.

5. The intelligent cleaning drone system based on high-pressure water jet as described in claim 1, characterized in that, The drone's water tank and high-pressure nozzles are activated to perform high-pressure water jet cleaning on the glass curtain wall area. The high-pressure nozzles are adjusted via a two-degree-of-freedom swing mechanism to complete the cleaning process, including: A real-time water pressure acquisition unit is used to acquire the real-time water pressure of the water tank; The real-time injection parameter acquisition unit is used to dynamically adjust the injection parameters of the high-pressure nozzle based on the real-time water pressure with the goal of improving the injection stability of the high-pressure nozzle, and obtain the real-time injection parameters. The real-time injection parameter transmission unit is used to transmit the real-time injection parameters to the microcontroller, and the microcontroller controls the micro servo motor to drive the two-degree-of-freedom swing mechanism to adjust the high-pressure nozzle.

6. The intelligent cleaning drone system based on high-pressure water jet as described in claim 5, characterized in that, With the goal of improving the spray stability of the high-pressure nozzle, the spray parameters of the high-pressure nozzle are dynamically adjusted based on the real-time water pressure to obtain real-time spray parameters, including: The real-time water pressure judgment subunit is used to determine whether the real-time water pressure exceeds the preset stable water pressure threshold. If so, it obtains the injection parameter adjustment command. The preset stable water pressure threshold is determined based on the type of high-pressure nozzle and the injection stability requirements. The injection parameter acquisition subunit is used to adjust the real-time water pump flow rate and the real-time nozzle diameter based on the real-time water pressure according to the injection parameter adjustment command, so as to obtain the real-time injection parameters.

7. The intelligent cleaning drone system based on high-pressure water jet as described in claim 6, characterized in that, Based on the real-time water pressure, the real-time water pump flow rate and the real-time nozzle diameter are adjusted according to the injection parameter adjustment command to obtain real-time injection parameters, including: The regulator constructs micro-units for pre-constructing a dynamic regulator, wherein the dynamic regulator is constructed based on a feedforward neural network and includes an input layer, a convolutional layer, and an output layer; The injection parameter output micro-unit is used to transmit the real-time water pressure, real-time water pump flow rate, and real-time nozzle diameter to the dynamic controller through the input layer, analyze them in the convolutional layer, and output the real-time injection parameters through the output layer.

8. The intelligent cleaning drone system based on high-pressure water jet as described in claim 1, characterized in that, The water tank adopts a centrally located structure.

9. The intelligent cleaning drone system based on high-pressure water jet as described in claim 1, characterized in that, Also includes: The water and electricity acquisition unit is used to monitor the battery power and water tank volume of the drone in real time, and obtain the real-time battery power and water tank volume. The return command acquisition unit is used to acquire a return command when the real-time battery power is lower than a preset power threshold and / or the real-time water tank water level is lower than a preset water level threshold. The flight control unit receives the return command and controls the drone to return to charge and / or add water.

10. A method for intelligent cleaning drones based on high-pressure water jets, characterized in that: The method is implemented by the intelligent cleaning drone system based on high-pressure water jet as described in any one of claims 1-9, and the method includes: Receive and parse task instructions to obtain the cleaning area and initial cleaning path; The drone flies to the cleaning area according to the initial cleaning path, and uses millimeter-wave radar and binocular vision sensors to optimize the obstacle avoidance perception of the initial cleaning path to obtain an optimized cleaning path. Real-time positioning data of the drone is obtained by using PTK mobile terminal and PTK base station terminal; When the real-time positioning data reaches the cleaning position of the optimized cleaning path, the drone is controlled by the flight control unit to hover, and the camera component is used to collect images of the glass curtain wall. The images are then transmitted to the image transmission terminal for analysis to determine the area of ​​the glass curtain wall to be cleaned. The drone's water tank and high-pressure nozzles are activated to clean the glass curtain wall area with high-pressure water jets, and the high-pressure nozzles are adjusted through a two-degree-of-freedom swing mechanism for cleaning.

Images