Neural network-based cleaning unmanned aerial vehicle adaptive control method and system

Abstract

The application provides a neural network-based cleaning unmanned aerial vehicle adaptive control method and system, relates to the cleaning technical field, obtains multi-source sensing information of the cleaning unmanned aerial vehicle through a sensor component; activates a visual saliency perception center to process and perceive a target image to be cleaned in the multi-source sensing information, obtains a point to be cleaned; and cooperates with unmanned aerial vehicle posture data and cleaning environment data in the multi-source sensing information to generate an adaptive cleaning path; takes a cleaning coefficient of the point to be cleaned as a constraint, adjusts an initial cleaning strategy to obtain a target cleaning strategy, and combines the adaptive cleaning path to perform cleaning adaptive control on the cleaning unmanned aerial vehicle. The application solves the technical problem that, due to lack of effective information fusion and adaptive decision mechanism, the cleaning unmanned aerial vehicle has identification deviation, path rigidity and cleaning strategy imbalance in a complex task environment, and achieves the technical effect of realizing precise and intelligent cleaning operation.

Description

Technical Field

[0001] This application relates to the field of cleaning technology, specifically to an adaptive control method and system for cleaning drones based on neural networks. Background Technology

[0002] As urban building facades become increasingly complex, traditional manual cleaning methods face significant limitations in terms of safety, efficiency, and job quality. Cleaning drones, as an emerging intelligent cleaning equipment, are increasingly being used for cleaning and maintaining building surfaces such as glass curtain walls and tiled exteriors due to their advantages, including high-altitude operation capabilities, remote control, and automated cleaning. Existing cleaning drone control methods primarily rely on preset paths and fixed cleaning strategies, with some systems using simple image recognition or GPS assistance for target positioning and path planning. However, in practical applications, due to significant differences in building facade structures, complex stain distribution, and variable environmental factors, these methods often suffer from inaccurate target identification, path rigidity failing to adapt to complex surface changes, and a mismatch between cleaning intensity and the actual degree of staining. These shortcomings result in low cleaning coverage, energy waste, and unstable cleaning effects, limiting the application of cleaning drones in high-precision, high-efficiency operation scenarios. Summary of the Invention

[0003] This application provides an adaptive control method and system for cleaning drones based on neural networks. It solves the technical problems of recognition bias, path rigidity and cleaning strategy mismatch caused by the lack of effective information fusion and adaptive decision-making mechanism in the prior art when cleaning drones perform tasks in complex environments. It achieves the technical effect of improving the cleaning accuracy, operation intelligence level and operation stability of cleaning drones in complex environments.

[0004] In view of the above problems, this application provides an adaptive control method for a cleaning drone based on a neural network. The method includes: performing multi-dimensional sensing and monitoring of the cleaning drone through a sensor component to obtain multi-source sensing information, wherein the sensor component is mounted on the cleaning drone; activating the visual saliency perception center to process and perceive the image of the target to be cleaned in the multi-source sensing information to obtain the cleaning point; generating an adaptive cleaning path based on the cleaning point and in conjunction with the drone attitude data and cleaning environment data in the multi-source sensing information; obtaining the cleaning coefficient of the cleaning point and adjusting the initial cleaning strategy of the cleaning drone with the cleaning coefficient as a constraint to obtain a target cleaning strategy; and performing adaptive cleaning control of the cleaning drone at the cleaning point by combining the adaptive cleaning path and the target cleaning strategy.

[0005] On the other hand, this application also provides an adaptive control system for a cleaning drone based on a neural network. The system includes: a multi-dimensional sensing and monitoring module, used to perform multi-dimensional sensing and monitoring of the cleaning drone through sensor components to obtain multi-source sensing information, wherein the sensor components are mounted on the cleaning drone; a target image perception module, used to activate the visual saliency perception center to process and perceive the target image to be cleaned in the multi-source sensing information to obtain the cleaning point; a cleaning path generation module, used to generate an adaptive cleaning path based on the cleaning point and in conjunction with the drone attitude data and cleaning environment data in the multi-source sensing information; a cleaning strategy adjustment module, used to obtain the cleaning coefficient of the cleaning point and adjust the initial cleaning strategy of the cleaning drone with the cleaning coefficient as a constraint to obtain a target cleaning strategy; and a cleaning control module, used to combine the adaptive cleaning path and the target cleaning strategy to perform adaptive cleaning control of the cleaning drone at the cleaning point.

[0006] One or more technical solutions provided in this application have at least the following beneficial effects:

[0007] Multi-dimensional sensing and monitoring of the cleaning drone by sensor components acquires multi-source sensor information, enabling the drone to comprehensively perceive the external cleaning environment and providing multi-dimensional data support for subsequent identification, path planning, and control decisions. By activating the visual saliency perception center, the drone processes and perceives the target image to be cleaned from the multi-source sensor information, simulating the human visual attention mechanism to accurately extract concentrated areas of stains or contaminants in the image, thus obtaining the cleaning points and improving the accuracy and automation of target identification. Based on the cleaning points, and in conjunction with the drone attitude data and cleaning environment data from the multi-source sensor information, an adaptive cleaning path is generated, realizing dynamic path planning and allowing the cleaning operation to be flexibly adjusted according to changes in the building facade structure and flight status. The cleaning coefficient of the cleaning points is obtained to quantify the degree of staining, and the initial cleaning strategy of the cleaning drone is adjusted using the cleaning coefficient as a constraint to obtain the target cleaning strategy, achieving differentiated and precise cleaning. Combining the adaptive cleaning path and the target cleaning strategy, the cleaning drone performs adaptive cleaning control of the cleaning points, achieving highly consistent and efficient automated cleaning operations at the target points.

[0008] In summary, this application achieves full-process adaptive control of cleaning drones for complex building facade cleaning tasks by constructing a multi-source sensor data-driven intelligent control process. First, comprehensive environmental and status data is acquired through multi-dimensional sensors, and a visual saliency perception mechanism is introduced to accurately identify the areas to be cleaned. Based on this, attitude information and environmental parameters are fused to dynamically generate an adaptive cleaning path, improving the flexibility and environmental adaptability of path planning. Simultaneously, a cleaning coefficient is introduced to specifically optimize the cleaning strategy, achieving a precise match between cleaning intensity and the degree of target stains. Finally, the path and strategy are output collaboratively to form a closed-loop control system. Overall, this application significantly improves the cleaning accuracy, operational intelligence, and operational stability of drones in complex environments.

[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] Figure 1 This is a flowchart illustrating the adaptive control method for a cleaning drone based on a neural network, as provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram illustrating the process of obtaining the cleaning point in the neural network-based adaptive control method for cleaning drones provided in this application embodiment.

[0012] Figure 3 This is a schematic diagram of the structure of an adaptive control system for a cleaning drone based on a neural network, provided in an embodiment of this application.

[0013] Figure labeling: Multi-dimensional sensing and monitoring module 10, target image perception module 20, cleaning path generation module 30, cleaning strategy adjustment module 40, cleaning control module 50. Detailed Implementation

[0014] This application provides an adaptive control method and system for cleaning drones based on neural networks. This solves the technical problems in the prior art, such as recognition bias, path rigidity, and cleaning strategy mismatch, caused by the lack of effective information fusion and adaptive decision-making mechanisms when cleaning drones perform tasks in complex environments. This achieves the technical effect of improving the cleaning accuracy, operational intelligence level, and operational stability of cleaning drones in complex environments.

[0015] Example 1, as Figure 1 As shown in the figure, this application provides an adaptive control method for a cleaning drone based on a neural network, the method comprising:

[0016] Step S100: Multi-dimensional sensing and monitoring of the cleaning drone is performed through sensor components to obtain multi-source sensing information, wherein the sensor components are mounted on the cleaning drone.

[0017] Furthermore, the sensor assembly includes a vision sensor, an inertial measurement unit, and an environmental sensor.

[0018] Specifically, the sensor assembly is a group of sensors integrated on the cleaning drone, including visual sensors, inertial measurement units (IMUs), and environmental sensors (including wind speed and direction sensors). Multi-source sensing information is the data set collected by the sensor assembly, reflecting comprehensive status information related to the cleaning task.

[0019] The cleaning drone utilizes its onboard sensor suite to simultaneously perceive its environment and its own status. Visual sensors capture images of the cleaning target (such as building facades), providing the basis for subsequent visual analysis. The inertial measurement unit (IMU) acquires the drone's current attitude information, such as pitch, roll, yaw, and three-axis acceleration and angular velocity data, to support path planning and stability control. Environmental sensors monitor key physical variables in the external environment in real time, such as wind speed and direction; this data plays a crucial role in path correction and interference mitigation. These different types of sensor data collectively constitute multi-source sensing information, providing comprehensive data support for subsequent steps.

[0020] Step S200: Activate the visual saliency perception center to process and perceive the target image to be cleaned in the multi-source sensing information to obtain the points to be cleaned.

[0021] Specifically, the visual saliency perception center is an image analysis module based on a neural network design. It simulates human perception of "prominent" areas in an image to extract the target area to be cleaned. The point to be cleaned is the location coordinate of the target area to be cleaned.

[0022] The pre-trained visual saliency perception center is activated to extract features from the target image to be cleaned from the input multi-source sensor information. First, the saliency feature parameters stored within the visual saliency perception center are called to analyze dimensions such as color contrast, texture variation, and edge density in the image, forming a multi-dimensional feature parameter set. Then, a feature map is generated using the center-periphery contrast operator, and further fused with the multi-channel feature maps to form a target saliency map. In this map, the intensity value of salient regions is significantly higher than that of surrounding regions. Based on the target saliency map, points with high cleaning priority, i.e., points to be cleaned, are extracted, providing a coordinate basis for subsequent paths and strategies.

[0023] This step enables the automatic identification of the areas in the image that require the most urgent cleaning, replacing manual annotation or blind global scanning, greatly improving cleaning efficiency and reducing energy and time waste, while ensuring priority cleaning of key areas.

[0024] Step S300: Based on the points to be cleaned, and in conjunction with the UAV attitude data and cleaning environment data in the multi-source sensor information, generate an adaptive cleaning path.

[0025] Specifically, the adaptive cleaning path is a dynamically generated path based on operational environment conditions, UAV status, and other factors, allowing for flexible adjustments to adapt to changing environments. First, a preliminary flight path (initial path) is constructed based on the location of the points to be cleaned and the current UAV attitude data. This path covers all target points. Then, the current environmental data is further analyzed to calculate environmental impact coefficients (such as offset angle compensation parameters). These coefficients are then used to spatially offset and angularly correct the initial path, resulting in a final cleaning path with higher stability and safety during actual flight. This avoids issues such as wind disturbance and attitude instability, achieving adaptive path optimization.

[0026] By incorporating attitude and environmental factors for adaptive path adjustment, the cleaning path not only achieves complete coverage but also possesses dynamic adaptability, maintaining stable flight and precise cleaning in complex external environments, thereby improving the success rate and task completion of cleaning operations.

[0027] Step S400: Obtain the cleaning coefficient of the point to be cleaned, and adjust the initial cleaning strategy of the cleaning drone with the cleaning coefficient as a constraint to obtain the target cleaning strategy.

[0028] Specifically, the cleaning coefficient is a comprehensive quantitative parameter that measures the cleaning difficulty and demand intensity of a specific cleaning point. The cleaning strategy is a combination of parameters encompassing cleaning operations performed by the drone, including spraying, water pressure control, and nozzle movement. First, the area of ​​each cleaning point is analyzed in relation to its neighboring areas to calculate its visual deviation index, and the cleaning coefficient is constructed accordingly. Then, this coefficient is compared to a preset benchmark cleaning coefficient to obtain the cleaning intensity ratio. Based on this ratio, key indicator parameters in the initial cleaning strategy (such as nozzle outlet pressure, water-to-ash ratio, nozzle oscillation frequency, and nozzle oscillation amplitude) are retrieved and fine-tuned proportionally. For example, if the cleaning coefficient is higher than the benchmark value, the nozzle pressure is increased or the cleaning time is extended, ultimately forming a target cleaning strategy adapted to the specific cleaning point. The initial cleaning strategy is a pre-set strategy for the cleaning drone to perform cleaning operations.

[0029] Step S500: Combining the adaptive cleaning path with the target cleaning strategy, the cleaning drone performs adaptive cleaning control of the points to be cleaned.

[0030] Specifically, the adaptive cleaning path generated in the aforementioned steps is integrated with the target cleaning strategy into the control command set to drive the UAV to execute the task. The UAV control system adjusts its heading and speed based on the position nodes and flight attitude in the adaptive cleaning path, while dynamically adjusting the nozzle action according to the cleaning parameters in the target cleaning strategy. During execution, feedback modules, such as pressure feedback and attitude error correction modules, are also integrated to achieve closed-loop adjustment of the cleaning action. When external disturbances (such as wind changes) are detected that affect the actual recoil force, the cleaning parameters are adjusted in real time to maintain cleaning accuracy and UAV stability.

[0031] Furthermore, step S100 includes:

[0032] Step S110: The visual sensor acquires an image of the building facade to be cleaned by the cleaning drone, which is recorded as the target image to be cleaned.

[0033] Step S120: The attitude data of the cleaning drone is obtained by dynamic monitoring through the inertial measurement unit and recorded as the drone attitude data.

[0034] Step S130: Monitor the operating environment of the cleaning drone using the environmental sensors to obtain the operating environment wind speed and wind direction, and assemble the cleaning environment data.

[0035] Step S140: The target image to be cleaned, the UAV attitude data, and the cleaning environment data constitute the multi-source sensing information.

[0036] Specifically, before the cleaning task begins, the vision sensors are activated to acquire images of the building facade at specific angles and frame rates. These images will serve as the foundational data for subsequent saliency analysis and location identification. During the acquisition process, trajectory stabilization mechanisms and image de-shaking algorithms are typically employed to ensure clear and identifiable images. Image data is recorded with timestamps for easy synchronization with subsequent sensor information.

[0037] During flight, the inertial measurement unit (IMU) collects real-time attitude data of the UAV, including changes in its three-dimensional orientation and acceleration. This data reflects whether the current flight attitude is stable, whether there are any oscillations, rolls, or yawing, and provides support for the controller to correct the path and nozzle attitude.

[0038] During cleaning operations, changes in wind force directly affect the spraying trajectory and the flight path of the drone. By using environmental sensors on the cleaning drone, the working environment can be monitored to obtain cleaning environment data such as wind speed and wind direction, so as to dynamically adjust the control parameters.

[0039] By fusing the three data sources mentioned above at the same timestamp, a structured multi-source sensor information set is formed, including: the original building facade image frame; IMU attitude information synchronized with the building facade image frame in time; and wind speed and direction values. Timestamp synchronization ensures that the image frame, attitude, and wind force measurements correspond to the same state time, achieving standardized integration of heterogeneous data. This provides a comprehensive and accurate perception foundation for subsequent task identification, path planning, and strategy formulation, thus establishing the prerequisites for intelligent cleaning control.

[0040] Furthermore, such as Figure 2 As shown, step S200 includes:

[0041] Step S210: Extract the saliency features from the visual saliency perception center memory.

[0042] Step S220: Extract the multidimensional feature parameter set of the target image to be cleaned based on the saliency features.

[0043] Step S230: Draw the first saliency map of the first feature in the saliency features according to the multidimensional feature parameter set.

[0044] Step S240: Based on the visual saliency perception center, fuse the first saliency map to obtain a target saliency map, and obtain the visually salient region in the target saliency map.

[0045] Step S250: The location of the visually significant area is taken as the point to be cleaned.

[0046] Specifically, the visual saliency perception center is constructed using an improved ResNeXt101 neural network to identify salient regions in the target image to be cleaned and determine the locations to be cleaned. When processing the target image from multi-source sensor information, the saliency features built into the visual saliency perception center are first invoked to extract a preset set of saliency features related to the target image to be cleaned. These saliency features include dimensions such as edge gradient distribution, texture contrast, shape sparsity, and color deviation.

[0047] Next, the visual saliency perception center uses an improved ResNeXt101 network to perform forward inference on the input target image to be cleaned, extracting a multi-dimensional feature parameter set of the target image. The backbone structure of this network includes multiple residual blocks, employs grouped convolution to enhance feature extraction capabilities, and generates multi-scale semantic representations at different network layers.

[0048] To enhance the ability to identify visually salient regions in complex environments, a large-scale contextual feature learning module is further integrated into the visual saliency perception center. This module introduces a hollow spatial pyramid pooling structure to extract contextual information under different receptive fields in parallel using multiple convolutional channels with different dilation rates. Furthermore, by fusing channel attention and spatial attention mechanisms, it suppresses interference from factors such as glass curtain wall reflections and decorative obstructions.

[0049] Based on the extracted multidimensional feature parameter set, a first saliency map associated with the first feature in the saliency features is drawn, and the salient regions are visualized in the form of a heatmap, with each pixel corresponding to a saliency score. Here, the first feature refers to any feature dimension in the saliency features, including edge gradient distribution, texture contrast, shape sparsity, color deviation, etc. The first saliency map reflects regions in the target image with high local differences in the first feature. Combining the global contextual awareness capability of the visual saliency perception center, the first saliency maps at multiple scales are weighted and fused to generate the target saliency map. The fusion strategy employs feature concatenation and convolutional normalization to ensure the accuracy of the saliency response in the image space. Regions with saliency response intensity higher than a preset threshold are extracted from the target saliency map, and the locations of these regions are output as points to be cleaned. These points have stable repeatability and spatial positioning reference value, and can be used for subsequent path planning and control strategy adjustments.

[0050] Furthermore, step S230 includes:

[0051] Step S231: Match the first parameter group corresponding to the first feature in the multidimensional feature parameter set.

[0052] Step S232: Obtain the first feature map based on the first parameter group, and retrieve the center-surrounding operator to analyze the first feature map to obtain the first saliency map.

[0053] Specifically, the first parameter set is the parameter set corresponding to the first feature in the multidimensional feature parameter set, containing the specific value and distribution information of the first feature. The first feature map is an image generated based on the first parameter set, highlighting the distribution of the first feature in the image to be cleaned. The center-periphery operator is an image processing operator used to analyze the relationship between each pixel in an image and its surrounding pixels.

[0054] First, determine the type of the first feature (e.g., color, texture, or edge). Then, based on the feature identifier or parameter label, quickly index and retrieve the first parameter set corresponding to the first feature from the multidimensional feature parameter set. The multidimensional feature parameter set is a structured collection of feature data, where each parameter corresponds to the response tensor of a different level in the neural network. It has a unique identifier and index structure, so the corresponding first parameter set can be quickly found through indexing or labeling.

[0055] The first feature map is generated by mapping the values ​​in the first parameter group to pixel intensity values ​​in the image domain. The mapping process includes: firstly, preprocessing the data in the first parameter group, including normalization and feature scaling, to adapt it to the range of image pixel values. Taking color features as an example, the RGB values ​​are normalized to the [0, 1] interval. For non-color features, such as shape features, the feature value corresponding to each pixel can be determined by combining interpolation algorithms. Next, a blank image matrix with the same size as the target image to be cleaned is initialized as the first feature map. Then, the pixel values ​​in the matrix are filled sequentially according to the feature values ​​in the first parameter group to obtain the first feature map. Then, the center-periphery operator is invoked to perform saliency analysis on the first feature map to obtain the first saliency map. The center-periphery operator is used to simulate the human eye's "center-sensitive, edge-suppressed" perception mechanism in saliency recognition. It adopts a multi-scale Gaussian difference or Laplacian pyramid structure and performs local response enhancement through the following function to calculate the saliency value of each pixel: In this saliency model, S(x,y) represents the saliency value of pixel (x,y), indicating the contrast difference intensity between the pixel and its larger neighborhood in the local region. F(x,y) is the first feature value of pixel (x,y), and G(σ) is the standard Gaussian kernel, where σ1 < σ2, used to extract the contrast difference between the local region and its neighborhood. The resulting first saliency value indicates the initial region of interest in the target image to be cleaned, providing input for subsequent saliency fusion and target location determination.

[0056] Furthermore, step S300 includes:

[0057] Step S310: Generate an initial cleaning path based on the points to be cleaned and the UAV attitude data.

[0058] Step S320: Analyze the wind speed and wind direction in the work environment to obtain the environmental impact coefficient.

[0059] Step S330: Adjust the initial cleaning path according to the environmental impact coefficient to obtain the adaptive cleaning path.

[0060] Specifically, firstly, based on multiple cleaning points output by the visual saliency perception module, and combined with the cleaning UAV attitude data (including position coordinates, yaw angle, pitch angle, roll angle, etc.) collected by the inertial measurement unit, spatial trajectory fitting methods (such as 3D spline interpolation algorithm, Bézier curve, or shortest path planning) are used to sort and connect the points to generate an initial cleaning path. During the path generation process, the UAV attitude is calculated as a position vector in a 3D Euler angle and coordinate system, and the airframe heading limit and maximum maneuvering angle threshold are introduced to ensure the flightability of the generated path and match it with the attitude of the nozzles mounted on the cleaning equipment.

[0061] Wind speed and direction in the cleaning environment directly affect the stability and spraying accuracy of the drone. Therefore, it is necessary to model and evaluate these factors, calculating environmental impact coefficients based on wind speed and direction to quantify the environment's influence on drone flight. The wind speed vector V and wind direction angle θ in the working environment are obtained through environmental sensors. ω The environmental impact coefficient E is calculated by combining the relative wind angle Δθ and the degree of wind field disturbance with the current flight direction of the UAV. f The calculation formula is as follows: Where ||V|| represents the magnitude of the wind speed vector V in the operating environment, and ω1 and ω2 are empirical weighting coefficients, which are used to characterize the contributions of wind speed and relative wind direction to flight stability, respectively.

[0062] Based on the above analysis, the environmental impact coefficient E is obtained. f The initial cleaning path is dynamically corrected. Correction methods include: path offset compensation: the original path points are adjusted laterally according to the wind direction offset to maintain the perpendicular angle between the cleaning nozzles and the target points; path sparse / dense adjustment: when E... f When the wind speed is high (above a preset threshold), the path points are appropriately densified to improve control accuracy; conversely, they are kept sparse to improve efficiency. Flight speed adjustment: In areas with high wind speed, the flight speed is reduced to improve path tracking accuracy. Dynamic attitude matching: By fitting the UAV dynamics model, the flight attitude is adjusted in real time to adapt to path changes. Finally, an adaptive cleaning path is obtained, which fully considers the spatial distribution of the points to be cleaned, the UAV's flight attitude, and the disturbance effects of the operating environment, and has good safety, stability, and cleaning accuracy.

[0063] Furthermore, step S400 includes:

[0064] Step S410: Obtain the arbitrary comparison deviation index between the point to be cleaned and any neighboring points based on the target saliency map.

[0065] Step S420: Obtain the area of ​​the point to be cleaned, and combine it with the arbitrary comparison deviation index to obtain the cleaning coefficient.

[0066] Specifically, the arbitrary contrast deviation index is an indicator that measures the degree of difference in saliency between a point to be cleaned and its arbitrary neighboring points, reflecting the relative saliency of the point compared to its surrounding area. Here, "arbitrary neighboring points" refers to any one of the points in the vicinity of the point to be cleaned. First, based on the target saliency map output by the visual saliency perception center, for each point to be cleaned, neighboring points within a preset radius (e.g., 10-20 pixels) are selected as a reference group. The saliency values ​​of the point to be cleaned and its arbitrary neighboring points are extracted, and the difference between the two is calculated to obtain the arbitrary contrast deviation index. A larger value indicates a more significant visual deviation and a more urgent need for cleaning.

[0067] Count the pixels of the points to be cleaned in the target saliency map to obtain the physical projected area of ​​the points to be cleaned. Combine the area with an arbitrary contrast deviation index to calculate the cleaning coefficient that comprehensively represents the cleaning priority and time evaluation. An example of the calculation formula is as follows: , where: C i Let A be the cleaning coefficient for the i-th point to be cleaned, and λ1 and λ2 be the fusion weights; i Let N be the area of ​​the i-th point to be cleaned, and N be the number of neighboring reference points; D i,j Let be the arbitrary comparison deviation index between the i-th point to be cleaned and any j-th arbitrary neighboring point. The obtained cleaning coefficient will be used in the next step to dynamically adjust the UAV cleaning strategy, ensuring efficient cleaning under the optimal path in a limited resource and complex environment.

[0068] Furthermore, step S400 also includes:

[0069] Step S430: Match the baseline cleaning coefficient corresponding to the initial cleaning strategy, and calculate the cleaning ratio between the coefficient to be cleaned and the baseline cleaning coefficient.

[0070] Step S440: Read the predetermined cleaning index and match the initial index parameter corresponding to the predetermined cleaning index in the initial cleaning strategy.

[0071] Step S450: Adjust the initial index parameters based on the ratio to be cleaned to obtain the target index parameters.

[0072] Step S460: Construct the target cleaning strategy based on the target index parameters.

[0073] Furthermore, the predetermined cleaning index refers to any one index in the predetermined index set, which includes nozzle outlet pressure, water-ash ratio, nozzle oscillation frequency, and nozzle oscillation amplitude.

[0074] Specifically, the baseline cleaning coefficient corresponding to the initial cleaning strategy is matched, which is the reference pollution level coefficient corresponding to the initial strategy under the current operating conditions. Then, based on the cleaning coefficient obtained in step S420 above, the ratio between it and the baseline cleaning coefficient is calculated to obtain the cleaning ratio. If the cleaning ratio is greater than 1, it means that the cleaning intensity needs to be increased based on the initial strategy; if the cleaning ratio is less than 1, it means that the cleaning intensity can be appropriately reduced based on the initial strategy.

[0075] Read the pre-set predetermined cleaning index and extract the corresponding initial index parameters from the initial cleaning strategy. The predetermined cleaning index is any item in the predetermined index set, which includes, but is not limited to: nozzle outlet pressure (unit: MPa), water-cement ratio (i.e., cleaning fluid concentration), nozzle oscillation frequency (unit: Hz), and nozzle oscillation amplitude (unit: °).

[0076] The calculated cleaning ratio is used as an adjustment factor to quantitatively adjust the initial index parameters to obtain target index parameters that match the current pollution intensity. This is achieved by multiplying each initial index parameter by the cleaning ratio. For example, if the cleaning ratio is 1.3, the initial index parameter for the nozzle oscillation frequency is 2Hz, and the initial index parameter for the nozzle outlet pressure is 0.3MPa, after adjustment, the target index parameter for the nozzle oscillation frequency increases to 2.6Hz, and the target index parameter for the nozzle outlet pressure increases to 0.39MPa. These adjusted target index parameters are then combined to generate a complete target cleaning strategy for adaptive control and scheduling of the cleaning drone in subsequent steps.

[0077] Furthermore, step S500 also includes:

[0078] Step S610: Establish a backwash force model and combine it with the target cleaning strategy to obtain the predicted backwash force.

[0079] Step S620: Adjust the predicted backlash force according to the environmental impact coefficient to obtain the actual backlash force.

[0080] Step S630: Perform anti-interference adjustment on the cleaning adaptive control based on the actual backflow force.

[0081] Specifically, the backwash force model is a mathematical model used to describe the relationship between backwash force and parameters such as nozzle outlet pressure and water-cement ratio of the cleaning equipment during the cleaning process. Based on fluid mechanics principles, considering the physical parameters and operating conditions of the cleaning equipment, a backwash force model is established. Then, the nozzle outlet pressure, water-cement ratio, and other parameters in the target cleaning strategy are substituted into the backwash force model to predict the backwash disturbance force experienced by the UAV during the cleaning task at the designated cleaning point, i.e., the predicted backwash force. The backwash force model can be constructed based on the following mechanical approximation models: Where Fr represents the predicted backflow force, P represents the nozzle outlet pressure, A represents the jet action area (which can be estimated from the nozzle oscillation amplitude and distance), and μ is the backflow correction coefficient, which can be set empirically based on the nozzle installation position and tilt angle. This model can be generated in real time by the cleaning drone control system during mission initialization and serves as a disturbance prediction reference.

[0082] The environmental impact coefficient obtained in step S300 is applied to the predicted backlash force to correct the actual backlash force, which can be calculated in the following form: Where Fr′ represents the actual recoil force, and E f δ is the environmental impact coefficient, and δ is the backflash disturbance sensitivity coefficient, which is determined experimentally.

[0083] Based on the actual backlash force, a fuzzy adaptive PID controller is used to perform anti-interference regulation on the cleaning adaptive control execution process in step S500. The fuzzy adaptive PID controller takes the disturbance error and its rate of change between the actual backlash force and the predicted backlash force as input variables and outputs the cleaning control compensation quantity to dynamically adjust the control command of the cleaning execution unit, so as to ensure that the fine cleaning task can still be stably executed in complex wind fields and strong disturbance environments.

[0084] In summary, the neural network-based adaptive control method for cleaning drones provided in this application has the following beneficial effects:

[0085] This application embodiment utilizes a visual sensor, inertial measurement unit, and environmental sensor mounted on a cleaning drone to acquire images of the target to be cleaned, drone attitude data, and environmental wind speed and direction information. This data is fused into multi-source sensor information, providing necessary input for subsequent path planning and cleaning control. Next, an improved ResNext101 neural network is used as the visual saliency perception center. A large-scale contextual feature learning module extracts local target features from the target image and fuses global contextual information to overcome complex background interference. Through saliency feature extraction and processing, a target saliency map is formed, and the cleaning points are identified, ensuring efficient identification of the target area. Based on the cleaning points and drone attitude data, an initial cleaning path is generated. Combining environmental wind speed and direction, an environmental impact coefficient is analyzed, and the path is further adjusted to obtain an adaptive cleaning path, ensuring that the cleaning path can cope with dynamic changes in complex environments. By calculating the target saliency map and the contrast deviation index, a cleaning coefficient is obtained, and the initial cleaning strategy is adjusted based on this coefficient to obtain the target cleaning strategy. By matching the baseline cleaning coefficient and adjusting the initial index parameters, the accuracy and efficiency of the cleaning process are ensured. During the cleaning process, a backlash force model is established, and the backlash force is predicted in conjunction with the target cleaning strategy. The predicted backlash force is then adjusted based on the environmental impact coefficient. Combined with the actual backlash force, anti-interference control is achieved to ensure the stability of the UAV and the cleaning effect. Under the influence of disturbances, PID fuzzy control is used to adjust the cleaning strategy in real time. By feeding back the disturbance error and rate of change, the PID control quantity is dynamically adjusted to effectively suppress the impact of external disturbances such as wind speed changes on the cleaning path and quality. Adaptive adjustment gives the cleaning process higher anti-interference capability and improves cleaning stability and accuracy. Overall, the embodiments of this application significantly improve the cleaning accuracy, operational intelligence level, and operational stability of UAVs in complex environments.

[0086] Example 2, as Figure 3 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides an adaptive control system for a cleaning drone based on a neural network, the system comprising:

[0087] The multi-dimensional sensing and monitoring module 10 is used to perform multi-dimensional sensing and monitoring of the cleaning drone through sensor components to obtain multi-source sensing information, wherein the sensor components are mounted on the cleaning drone.

[0088] The target image perception module 20 is used to activate the visual saliency perception center to process and perceive the target image to be cleaned in the multi-source sensing information, and obtain the points to be cleaned.

[0089] The cleaning path generation module 30 is used to generate an adaptive cleaning path based on the points to be cleaned and in conjunction with the UAV attitude data and cleaning environment data in the multi-source sensor information.

[0090] The cleaning strategy adjustment module 40 is used to obtain the cleaning coefficient of the point to be cleaned, and adjust the initial cleaning strategy of the cleaning drone with the cleaning coefficient as a constraint to obtain the target cleaning strategy.

[0091] The cleaning control module 50 is used to combine the adaptive cleaning path with the target cleaning strategy to perform adaptive cleaning control of the cleaning drone at the point to be cleaned.

[0092] Furthermore, the sensor assembly includes a vision sensor, an inertial measurement unit, and an environmental sensor. In this embodiment, the multi-dimensional sensing and monitoring module 10 is also used to perform the following steps:

[0093] The visual sensor acquires an image of the building facade to be cleaned by the cleaning drone, which is recorded as the target image to be cleaned; the inertial measurement unit dynamically monitors the drone's own attitude data, which is recorded as the drone attitude data; the environmental sensor monitors the drone's operating environment to obtain the operating environment wind speed and wind direction, which are then combined to form the cleaning environment data; the target image to be cleaned, the drone attitude data, and the cleaning environment data constitute the multi-source sensing information.

[0094] Furthermore, in this embodiment of the application, the target image perception module 20 is also used to perform the following steps:

[0095] Extract salient features from the memory of the visual saliency perception center; extract a multidimensional feature parameter set of the target image to be cleaned based on the salient features; draw a first saliency map of the first feature in the salient features based on the multidimensional feature parameter set; fuse the first saliency map based on the visual saliency perception center to obtain a target saliency map, and obtain the visually salient region in the target saliency map; use the position of the visually salient region as the point to be cleaned.

[0096] Furthermore, in this embodiment of the application, the target image perception module 20 is also used to perform the following steps:

[0097] Match the first parameter group corresponding to the first feature in the multidimensional feature parameter set; obtain the first feature map based on the first parameter group, and call the center-surrounding operator to analyze the first feature map to obtain the first saliency map.

[0098] Furthermore, in this embodiment of the application, the cleaning path generation module 30 is also used to perform the following steps:

[0099] An initial cleaning path is generated based on the points to be cleaned and the attitude data of the UAV; the wind speed and wind direction in the working environment are analyzed to obtain an environmental impact coefficient; the initial cleaning path is adjusted based on the environmental impact coefficient to obtain an adaptive cleaning path.

[0100] Furthermore, in this embodiment of the application, the cleaning strategy adjustment module 40 is also used to perform the following steps:

[0101] Based on the target saliency map, obtain the arbitrary comparison deviation index between the point to be cleaned and any neighboring points; obtain the area of ​​the point to be cleaned, and combine it with the arbitrary comparison deviation index to obtain the cleaning coefficient.

[0102] Furthermore, in this embodiment of the application, the cleaning strategy adjustment module 40 is also used to perform the following steps:

[0103] Match the baseline cleaning coefficient corresponding to the initial cleaning strategy, and calculate the cleaning ratio between the coefficient to be cleaned and the baseline cleaning coefficient; read the predetermined cleaning index, and match the initial index parameter corresponding to the predetermined cleaning index in the initial cleaning strategy; adjust the initial index parameter based on the cleaning ratio to obtain the target index parameter; construct the target cleaning strategy based on the target index parameter.

[0104] Furthermore, the predetermined cleaning index refers to any one index in the predetermined index set, which includes nozzle outlet pressure, water-ash ratio, nozzle oscillation frequency, and nozzle oscillation amplitude.

[0105] Furthermore, the system described in this application embodiment also includes an anti-interference control module, which is used to perform the following steps:

[0106] A backwash force model is established, and a predicted backwash force is obtained by combining it with the target cleaning strategy; the predicted backwash force is adjusted according to the environmental impact coefficient to obtain the actual backwash force; the actual backwash force is used to perform anti-interference regulation on the cleaning adaptive control.

[0107] Through the foregoing detailed description of the neural network-based adaptive control method for cleaning drones, those skilled in the art can clearly understand that the neural network-based adaptive control system for cleaning drones in this embodiment, as corresponding to the system disclosed in Embodiment 2, has corresponding functional modules and beneficial effects as it corresponds to the method disclosed in Embodiment 1. For relevant details, please refer to the method section.

[0108] 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 neural network-based adaptive control method for cleaning unmanned aerial vehicles, characterized in that, The application relates to a cleaning unmanned aerial vehicle and a cleaning method thereof. The cleaning unmanned aerial vehicle is subjected to multi-dimensional sensing monitoring through a sensor assembly, and multi-source sensing information is obtained, wherein the sensor assembly is mounted on the cleaning unmanned aerial vehicle; A visual saliency perception center is activated to process and perceive a target image to be cleaned in the multi-source sensing information, and a point to be cleaned is obtained, including: Saliency features in a memory of the visual saliency perception center are extracted; Multi-dimensional feature parameter sets of the target image to be cleaned are obtained according to the saliency features; A first saliency map of a first feature in the saliency features is drawn according to the multi-dimensional feature parameter sets; According to the visual saliency perception center, the first saliency map is fused to obtain a target saliency map, and a visual saliency region in the target saliency map is acquired; The position of the visual saliency region is taken as the point to be cleaned; Based on the point to be cleaned, unmanned aerial vehicle attitude data and cleaning environment data in the multi-source sensing information are cooperated to generate an adaptive cleaning path, including: According to the point to be cleaned and the unmanned aerial vehicle attitude data, an initial cleaning path is generated; The environmental influence coefficient is obtained by analyzing the working environment wind speed and the working environment wind direction; According to the environmental influence coefficient, the initial cleaning path is adjusted to obtain the adaptive cleaning path; A cleaning coefficient of the point to be cleaned is acquired, and an initial cleaning strategy of the cleaning unmanned aerial vehicle is adjusted by taking the cleaning coefficient as a constraint to obtain a target cleaning strategy. The cleaning coefficient is a comprehensive quantitative parameter for measuring the cleaning difficulty and demand intensity of a cleaning point. The cleaning strategy is a parameter combination including unmanned aerial vehicle spraying, water pressure control and nozzle moving cleaning operation, including: According to the target saliency map, any comparative deviation index of the point to be cleaned and any neighborhood point is obtained. The any comparative deviation index is an index for measuring the difference degree of the saliency features of the point to be cleaned and any neighborhood point, and reflects the relative saliency of the point and the surrounding area; The area of the point to be cleaned is acquired, and the cleaning coefficient is obtained by combining the any comparative deviation index; The reference cleaning coefficient corresponding to the initial cleaning strategy is matched, and the cleaning ratio of the cleaning coefficient to the reference cleaning coefficient is calculated; A predetermined cleaning index is read, and an initial index parameter corresponding to the predetermined cleaning index in the initial cleaning strategy is matched; The initial index parameter is adjusted based on the cleaning ratio to obtain a target index parameter; The target cleaning strategy is established based on the target index parameter; The cleaning unmanned aerial vehicle is subjected to cleaning adaptive control of the point to be cleaned in combination with the adaptive cleaning path and the target cleaning strategy.

2. The neural network-based adaptive control method for cleaning drones according to claim 1, wherein, The sensor assembly includes a visual sensor, an inertial measurement unit and an environment sensor. The cleaning unmanned aerial vehicle is subjected to multi-dimensional sensing monitoring through the sensor assembly, and multi-source sensing information is obtained, including: The visual sensor is used to collect an image of a building facade to be cleaned by the cleaning unmanned aerial vehicle, which is recorded as the target image to be cleaned; The self-attitude data of the cleaning unmanned aerial vehicle is obtained by dynamic monitoring of the inertial measurement unit, and is denoted as unmanned aerial vehicle attitude data; The working environment of the cleaning unmanned aerial vehicle is monitored by the environmental sensor to obtain working environment wind speed and working environment wind direction, and the cleaning environment data is formed; The target image to be cleaned, the unmanned aerial vehicle attitude data and the cleaning environment data form the multi-source sensing information.

3. The method of claim 1, wherein the method further comprises: The first saliency map of the first feature in the saliency feature is drawn according to the multi-dimensional feature parameter set, including: Matching the first parameter group corresponding to the first feature in the multi-dimensional feature parameter set; According to the first parameter group, a first feature map is obtained, and a central-peripheral operator is called to analyze the first feature map to obtain the first saliency map.

4. The neural network-based adaptive control method for cleaning drones according to claim 1, wherein, The predetermined cleaning index refers to any one index in the predetermined index set, and the predetermined index set includes nozzle outlet pressure, water-cement ratio, nozzle swing frequency and nozzle swing amplitude.

5. The neural network-based adaptive control method for cleaning drones according to claim 4, wherein, After combining the adaptive cleaning path and the target cleaning strategy, the cleaning adaptive control of the cleaning unmanned aerial vehicle on the target cleaning point is further included, and the method further includes: A backwash force model is established, and a predicted backwash force is obtained by combining the target cleaning strategy; According to the environmental influence coefficient, the predicted backwash force is adjusted to obtain an actual backwash force; The actual backwash force is combined to perform anti-interference regulation and control on the cleaning adaptive control.

6. A neural network-based cleaning drone adaptive control system, characterized in that, The system is used to execute the neural network-based cleaning unmanned aerial vehicle adaptive control method of any one of claims 1-5, including: A multi-dimensional sensing monitoring module is used to monitor the cleaning unmanned aerial vehicle by a sensor assembly to obtain multi-source sensing information, wherein the sensor assembly is mounted on the cleaning unmanned aerial vehicle; A target image perception module is used to activate a visual saliency perception center to process and perceive the target image to be cleaned in the multi-source sensing information to obtain a cleaning point; A cleaning path generation module is used to generate an adaptive cleaning path based on the cleaning point to be cleaned and in cooperation with the unmanned aerial vehicle attitude data and the cleaning environment data in the multi-source sensing information; A cleaning strategy adjustment module is used to obtain a cleaning coefficient of the cleaning point to be cleaned, and to adjust an initial cleaning strategy of the cleaning unmanned aerial vehicle by taking the cleaning coefficient as a constraint to obtain a target cleaning strategy; A cleaning control module is used to combine the adaptive cleaning path and the target cleaning strategy to perform cleaning adaptive control of the cleaning unmanned aerial vehicle on the cleaning point to be cleaned.

Images