A photovoltaic panel cleaning nozzle self-adaptive control method and system

Abstract

The application relates to the technical field of electric digital data processing, and discloses a photovoltaic panel cleaning nozzle adaptive control method and system, which realizes real-time acquisition of a scene image below a cleaning device, obtains a photovoltaic panel effective area binary mask through semantic segmentation processing, and extracts left and right boundary pixel coordinates of the area in an image coordinate system; through a pixel angle mapping model pre-calibrated, nozzle swing angle boundaries are solved, a final control angle range is obtained in combination with a safety margin, a swing center and an amplitude are calculated to generate an expected angle trajectory; a double-loop framework is adopted to execute servo control, a low-frequency decision loop with a vision sampling frame rate is used to process scene judgment and trajectory updating, and a high-frequency servo loop with real-time angle feedback is used to track the expected trajectory; when a photovoltaic panel boundary or a gap is identified, a high-priority intelligent reflection instruction is triggered to cooperatively control a water pump and nozzle swing start and stop. The application solves the problem of non-adaptability and serious waste in the prior art and achieves the purposes of high precision and adaptability.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to an adaptive control method and system for cleaning nozzles of photovoltaic panels. Background Technology

[0002] Current photovoltaic cleaning devices (such as cleaning robots and drone-borne cleaning systems) mostly use open-loop control with preset programs for nozzle control when performing cleaning tasks. This method has the following prominent problems: low level of intelligence: unable to adaptively adjust according to changes in the real-time position, shape, and size of the photovoltaic panels; poor cleaning effect: prone to missed cleaning or repeated cleaning due to discontinuous photovoltaic panel arrays, installation errors, or equipment drift; resource waste: significant waste of cleaning fluid, water, and electricity; reliance on manual intervention: requires high technical skills from operators and poses safety risks.

[0003] For example, Chinese patent CN119477825A discloses an operation method for a drone-based photovoltaic panel contamination identification and cleaning system, providing the following technical solution: using a high-definition camera mounted on the drone to photograph the photovoltaic panel and preprocessing the image data; using a mask generator encoder to extract image features and constructing a feature vector from the extracted effective data; improving the Osprey optimization algorithm using a multi-difference Cauchy mutation strategy and a refraction inverse learning strategy to obtain the IOOA algorithm; inputting the feature vector into an enhanced YOLOv8 model for training, optimizing the enhanced YOLOv8 model parameters using the IOOA algorithm, and using the trained and optimized enhanced YOLOv8 model to identify photovoltaic panel contamination; based on the identification results of the enhanced YOLOv8 model, controlling the drone to approach the cleaning target and perform cleaning actions. This invention can significantly improve the intelligence level of the drone-based photovoltaic panel contamination identification and cleaning system, achieving more efficient and accurate operation and maintenance management. However, the above-mentioned method for operating a photovoltaic panel pollution identification and cleaning system based on drones is limited to pollution identification and location. The cleaning execution is still a preset open-loop control, which cannot dynamically adjust the nozzle trajectory according to the real-time position of the photovoltaic panel, resulting in problems such as missed spraying, repeated spraying, and waste of resources. Summary of the Invention

[0004] This invention solves the problems of lack of self-adaptation, serious waste, and fixed cleaning mode in the prior art, and proposes an adaptive control method and system for photovoltaic panel cleaning nozzles, achieving the goals of high precision, adaptability, and resource conservation.

[0005] Furthermore, the present invention aims to achieve precise dynamic following of the nozzle to the boundary of the photovoltaic panel through real-time visual recognition and closed-loop control, thereby improving cleaning accuracy and adaptability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An adaptive control method for photovoltaic panel cleaning nozzles includes: Real-time acquisition of scene images below the cleaning device; semantic segmentation processing to obtain a binary mask of the effective area of ​​the photovoltaic panel; extraction of the left and right boundary pixel coordinates of the area in the image coordinate system. By using a pre-calibrated pixel angle mapping model, the nozzle swing angle boundary is solved, and the final control angle range is obtained by combining the safety margin. The swing center and amplitude are calculated to generate the desired angle trajectory. A dual-loop architecture is used to perform servo control, with a low-frequency decision loop that matches the visual sampling frame rate to handle scene analysis and trajectory updates, and a high-frequency servo loop with real-time angle feedback to track the desired trajectory. When the boundary or gap of the photovoltaic panel is detected, a high-priority intelligent reflection command is triggered to coordinate the start and stop of the water pump and nozzle swing.

[0007] An adaptive control system for a photovoltaic panel cleaning nozzle includes: a visual perception unit that acquires images of the photovoltaic panel in real time and outputs a binary mask of the effective area, with the output end connected to the input end of a control decision unit; the control decision unit establishes a real-time mapping between pixel coordinates and nozzle swing angle and generates a desired control trajectory, with the output end connected to an execution control unit; the execution control unit adopts a dual-loop control architecture.

[0008] By using image perception, coordinate mapping, and servo control real-time closed-loop control logic, the dynamic and precise alignment of the nozzle swing trajectory with the actual physical boundary of the photovoltaic panel is achieved. At the same time, through intelligent reflection commands and coordinated control of the water pump, the precise dispensing of cleaning fluid is achieved, which greatly improves cleaning accuracy, environmental adaptability and resource utilization efficiency, and reduces reliance on manual intervention.

[0009] Preferably, the step of obtaining the effective region binary mask of the photovoltaic panel through semantic segmentation includes scaling the acquired image to a fixed size, performing pixel value normalization preprocessing, and then inputting it into a pre-trained U-Net semantic segmentation model. The model outputs a probability map of the same size as the input image. The value of each pixel in the probability map represents the confidence level of belonging to the photovoltaic panel region. The probability map is binarized by a preset confidence threshold, and morphological opening and closing operations are performed to obtain the effective region binary mask.

[0010] The standardized preprocessing process ensures the consistency of model input. Based on the U-Net semantic segmentation model, pixel-level accurate recognition of photovoltaic panel areas is achieved. It has strong robustness to interference from complex scenes such as light changes, shadows, reflections, and stains. With binarization and morphological post-processing, a stable, complete, and noise-free effective area mask of the photovoltaic panel can be output, providing a highly reliable visual input benchmark for subsequent servo control.

[0011] Preferably, the step of extracting the left and right boundary pixel coordinates of the region in the image coordinate system includes scanning the binary mask line by line, extracting the minimum and maximum column coordinates of the photovoltaic panel region in the horizontal direction of the image as the left and right boundary pixel coordinates of the photovoltaic panel region, and determining the pixel coordinates corresponding to the image center line as the reference center coordinates.

[0012] By using line-by-line scanning, the true boundary of the photovoltaic panel in the horizontal direction of the image is accurately located, providing a coordinate reference for subsequent pixel angle mapping. At the same time, using the image center line as the reference center coordinate ensures the precise correspondence between the nozzle swing center and the center of the photovoltaic panel area, avoiding offset and omission of the cleaning range.

[0013] Preferably, the calculation of the nozzle swing angle boundary includes: determining the geometric correspondence between the image pixel coordinate system and the world coordinate system with the nozzle rotation center as the origin; determining the proportional coefficient of the change in nozzle swing angle corresponding to the pixel offset by the camera's field of view and image width parameters; mapping the pixel coordinates corresponding to the image center line to the mechanical median angle of the nozzle; and converting the left and right boundary pixel coordinates of the photovoltaic panel area into the left and right limit physical angles of the nozzle swing based on the calibrated proportional coefficient.

[0014] A precise linear mapping relationship was established from the image pixel space to the nozzle physical execution space. The pixel coordinates to the swing angle were quickly calculated through a pre-calibrated scaling factor. Sufficient control accuracy can be guaranteed in normal operation scenarios with stable camera installation parameters, providing the core conversion basis for the adaptive dynamic adjustment of the nozzle swing range.

[0015] Preferably, the calculation of the nozzle swing angle boundary also includes collecting the three-axis attitude angle and current ground clearance data of the camera on the cleaning device; when a deviation is detected between the actual ground clearance of the camera and the calibrated height, the pre-calibrated proportional coefficient is corrected accordingly according to the proportion of the height deviation; when a tilt state is detected where the camera optical axis is not perpendicular to the cleaning surface, the perspective projection model is switched, the pixel coordinates are converted into normalized camera coordinates through the camera intrinsic parameter matrix, and then combined with the extrinsic parameter rotation matrix and translation vector corresponding to the attitude data acquired in real time, the camera coordinates are converted into the world coordinate system with the nozzle rotation center as the origin, and the expected nozzle swing angle of the corresponding pixel coordinates is calculated.

[0016] It achieves dynamic adaptive compensation for changes in the attitude and height of the cleaning carrier, solving the mapping deviation problem caused by bumps, slopes, wind deviation, and installation misalignment during carrier movement. It can still ensure the accuracy of pixel angle mapping under complex working conditions, greatly improving the system's adaptability to different working scenarios and different mobile carriers, and expanding the application boundaries of the system.

[0017] Preferably, the dual-loop architecture for servo control includes two control loops operating in different frequency domains. The adaptive decision loop operates at a frequency matching the visual sampling frame rate to perform photovoltaic panel boundary recognition, photovoltaic panel gap judgment, cleaning mode decision, and desired angle trajectory update. The high-bandwidth servo loop operates at a frequency much higher than the adaptive decision loop to perform high-precision position tracking of the nozzle oscillation. The core control law uses a PID controller, with the input being the angle error between the desired angle trajectory and the actual oscillation angle of the nozzle, and the output driving a servo motor to drive the nozzle to complete the oscillation action.

[0018] The dual-ring architecture, which is decoupled in the frequency domain, enables deep collaboration between intelligent decision-making and high-speed execution. The low-frequency decision ring focuses on scene understanding and trajectory planning, avoiding the waste of hardware resources caused by high-frequency computing. The high-frequency servo ring ensures high-precision tracking of nozzle oscillation and strong anti-disturbance capability, and can effectively suppress tracking errors caused by mechanical transmission gaps and high-frequency vibration of the carrier.

[0019] Preferably, the triggering high-priority intelligent reflection command includes: when the edge or gap of the photovoltaic panel is detected, immediately generating a boundary event signal, and sending a preset emergency stop or reverse motion command directly to the driver of the servo motor through a high-priority interrupt channel.

[0020] By bypassing the calculation cycle of conventional controllers, the response delay of photovoltaic panel boundary events is greatly reduced, achieving zero boundary crossing of nozzles at the photovoltaic panel boundary. At the same time, it can coordinate with water pumps to achieve synchronous start and stop, achieving zero dry spraying effect, greatly reducing the ineffective consumption of cleaning fluid and electricity, and improving the reliability and resource utilization efficiency of cleaning operations.

[0021] Preferably, the dual-loop architecture for servo control further includes the encoder in the high-bandwidth servo loop reporting the actual motion state of the nozzle, motor load data, and external disturbance information to the adaptive decision loop in real time via a state feedback stream; the adaptive decision loop sending the photovoltaic panel boundary change trend predicted based on continuous frame images as a feedforward information stream to the controller of the high-bandwidth servo loop in advance.

[0022] The bidirectional information flow enables deep collaboration and forward compensation of dual-loop control. The state feedback flow allows the decision loop to optimize scenario analysis and decision logic based on the actual working conditions of the execution end. The feedforward information flow allows the servo loop to adapt to upcoming trajectory changes in advance, achieving smoother trajectory tracking and significantly improving the tracking smoothness and overall robustness of the system when facing continuous boundary changes or sudden disturbances.

[0023] Preferably, the method of obtaining the final control angle range by combining the safety margin includes introducing a fixed safety margin, correcting the calculated left and right limit physical angles, and obtaining the final swing angle range used for control; the safety margin is contracted inward based on the left and right boundary angles calculated by the pixel angle mapping model to form the final nozzle swing control range.

[0024] It can effectively offset the control deviation caused by image segmentation edge positioning error, camera calibration error, and mechanical transmission gap. While ensuring full coverage cleaning of the photovoltaic panel area, it tightens the effective swing range of the nozzle, avoids ineffective spraying caused by the nozzle going out of bounds, and improves cleaning accuracy and resource utilization efficiency.

[0025] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention significantly improves the accuracy and reliability of photovoltaic panel cleaning by constructing a real-time closed-loop system of visual perception, coordinate mapping, and servo control. It uses a deep learning model to identify the photovoltaic panel boundary in real time and maps the pixel coordinates to the nozzle swing angle, ensuring that the nozzle trajectory always matches the actual position of the photovoltaic panel. This solves the problem of missed cleaning and re-cleaning caused by position mismatch in traditional open-loop control, and greatly improves the cleaning quality and operation stability.

[0026] 2. This invention has environmental adaptability and can respond in real time to complex working conditions such as photovoltaic panel installation errors, carrier movement disturbances and changes in light. By introducing an inertial measurement unit and dynamic attitude compensation, the system can automatically adjust control parameters when the drone or robot platform is bumpy or tilted, ensuring that the nozzle always accurately covers the target area. This demonstrates a high degree of intelligence and robustness and significantly reduces the reliance on manual intervention.

[0027] 3. This invention achieves efficient resource utilization through precise nozzle control and water pump coordination. The system activates spraying only when the photovoltaic panel area is detected and immediately shuts off during the return stroke or interval, avoiding waste of cleaning fluid and electricity. The intelligent reflection path mechanism can respond to boundary events in milliseconds, achieving zero boundary crossings and zero dry spraying, bringing significant economic benefits and environmental value to photovoltaic cleaning operations. Attached Figure Description

[0028] Figure 1 This is a flowchart of the dual-loop fusion adaptive servo control method for an adaptive control method of photovoltaic panel cleaning nozzles according to the present invention.

[0029] Figure 2 This is a schematic diagram of the coordinate mapping from image space to actuator space for an adaptive control method for cleaning nozzles of photovoltaic panels according to the present invention.

[0030] Figure 3This is a timing diagram comparing the absolute values ​​of tracking errors in the open-loop and closed-loop systems of an adaptive control method for cleaning nozzles on photovoltaic panels according to the present invention.

[0031] Figure 4 This is a time-series comparison diagram of the actual angle response under boundary events for the adaptive control method of photovoltaic panel cleaning nozzles according to the present invention. Detailed Implementation

[0032] See Figures 1 to 4 As shown, an adaptive control method for cleaning nozzles on photovoltaic panels includes: Real-time acquisition of scene images below the cleaning device; semantic segmentation processing to obtain a binary mask of the effective area of ​​the photovoltaic panel; extraction of the left and right boundary pixel coordinates of the area in the image coordinate system. By using a pre-calibrated pixel angle mapping model, the nozzle swing angle boundary is solved, and the final control angle range is obtained by combining the safety margin. The swing center and amplitude are calculated to generate the desired angle trajectory. A dual-loop architecture is used to perform servo control, with a low-frequency decision loop that matches the visual sampling frame rate to handle scene analysis and trajectory updates, and a high-frequency servo loop with real-time angle feedback to track the desired trajectory. When the boundary or gap of the photovoltaic panel is detected, a high-priority intelligent reflection command is triggered to coordinate the start and stop of the water pump and nozzle swing.

[0033] An adaptive control system for a photovoltaic panel cleaning nozzle includes: a visual perception unit that acquires images of the photovoltaic panel in real time and outputs a binary mask of the effective area, with the output end connected to the input end of a control decision unit; the control decision unit establishes a real-time mapping between pixel coordinates and nozzle swing angle and generates a desired control trajectory, with the output end connected to an execution control unit; the execution control unit adopts a dual-loop control architecture.

[0034] The purpose of this invention is to provide a true closed-loop visual servo control system that integrates real-time image acquisition, photovoltaic panel area recognition, nozzle oscillation parameter calculation, and servo control to achieve a real-time closed loop of "perception-decision-execution". This enables the nozzle oscillation range to match the photovoltaic panel boundary in real time, adapts to photovoltaic panels of different sizes and layouts, reduces cleaning resource waste, and improves cleaning quality and efficiency.

[0035] This invention employs a frequency-domain decoupled dual-loop control architecture as its core execution framework. This architecture consists of two control loops operating at different frequencies, working in deep collaboration. The low-frequency decision loop uses an adaptive decision loop to operate at a low frequency matching the visual sampling frame rate. Its core functions include photovoltaic panel boundary recognition, inter-panel gap determination, cleaning operation mode decision-making, and dynamic updating of the nozzle's desired angle trajectory. The high-frequency servo loop uses a high-bandwidth servo loop to complete closed-loop operation at a frequency significantly higher than the adaptive decision loop. It focuses on achieving high-precision position tracking of the nozzle's oscillation. Its core control unit uses a PID controller. The controller's input signal is the angle error between the nozzle's desired angle trajectory and the actual oscillation angle. The output control signal directly drives the servo motor, which in turn drives the nozzle to complete the corresponding oscillation action.

[0036] The dual-ring architecture achieves control performance optimization and coordination through bidirectional information flow: the encoder built into the high-bandwidth servo ring reports the actual operating status of the nozzle, the load data of the servo motor, and the external disturbance information during operation to the adaptive decision ring in real time through the status feedback link; at the same time, the adaptive decision ring converts the photovoltaic panel boundary change trend predicted by continuous frame images into feedforward control information and sends it to the controller of the high-bandwidth servo ring in advance, thereby improving the smoothness of trajectory tracking and the ability to resist disturbances.

[0037] In the front-end perception stage of the control process, the cleaning device collects scene images below the device in real time during operation. The collected images are first standardized and preprocessed, and then the pixel values ​​are normalized after scaling the images to a fixed size. The preprocessed images are then input into a pre-trained U-Net semantic segmentation model, which outputs a probability map with the same size as the input image. The value of each pixel in the probability map represents the confidence that the corresponding pixel belongs to the photovoltaic panel area. Subsequently, the probability map is binarized by a preset confidence threshold, and the binarization result is optimized by morphological opening and closing operations to finally obtain a binary mask of the effective area of ​​the photovoltaic panel.

[0038] After generating the binary mask, the mask is scanned line by line to extract the minimum and maximum column coordinates of the photovoltaic panel region in the horizontal direction of the image. The two sets of coordinates are used as the left and right boundary pixel coordinates of the photovoltaic panel region, respectively. At the same time, the pixel coordinates corresponding to the center line of the image are determined as the reference center coordinates for subsequent coordinate mapping calculations.

[0039] Based on the extracted boundary pixel coordinates, the nozzle swing angle boundary is calculated using a pre-calibrated pixel angle mapping model. This mapping model pre-establishes the geometric correspondence between the image pixel coordinate system and the world coordinate system with the nozzle rotation center as the origin. Using the camera field of view and image width parameters, the proportional coefficient corresponding to the change in nozzle swing angle is calculated. The pixel coordinates corresponding to the image centerline are mapped to the mechanical median angle of the nozzle. Then, based on the calibrated proportional coefficient, the left and right boundary pixel coordinates of the photovoltaic panel area are converted into the left and right limit physical angles of the nozzle swing, respectively.

[0040] During the angle calculation process, the three-axis attitude angles and current ground clearance data of the camera mounted on the cleaning device are collected simultaneously to achieve dynamic adaptive adjustment of the mapping model: when a deviation between the actual ground clearance of the camera and the calibration height is detected, the pre-calibrated scaling factor is corrected according to the corresponding ratio of the height deviation; when the camera optical axis is detected to be tilted in a non-perpendicular state with the cleaning surface, the perspective projection model is automatically switched. First, the pixel coordinates are converted into normalized camera coordinates through the camera intrinsic parameter matrix, and then the camera coordinates are converted into the world coordinate system with the nozzle rotation center as the origin by combining the extrinsic parameter rotation matrix and translation vector corresponding to the real-time attitude data, and the expected nozzle swing angle under the corresponding pixel coordinates is calculated.

[0041] After calculating the nozzle swing limit angle, a fixed safety margin is introduced to correct the calculation result. This safety margin is reduced inward based on the calculated left and right limit physical angles to obtain the final swing angle range for nozzle control. Based on the final control angle range, the swing center and swing amplitude of the nozzle are calculated, the desired angle trajectory of the nozzle swing is generated, and sent to the dual-loop control architecture for execution.

[0042] Throughout the entire control operation, a high-priority emergency response mechanism is set up: when the adaptive decision loop identifies the edge of the photovoltaic panel or the gap between panels through real-time image recognition, it will immediately generate a boundary event signal. This signal will send a preset emergency stop or reverse motion command directly to the servo motor driver through a high-priority interrupt channel, triggering intelligent reflection control. At the same time, the intelligent reflection command will synchronously link the water pump control unit to achieve coordinated matching between the start and stop of the water pump and the oscillation action of the nozzle. Spraying will only be turned on when the nozzle oscillation covers the effective photovoltaic panel area, and spraying will be turned off when a panel-free area is identified or during the boundary return phase.

[0043] In another embodiment, the core of the invention lies in constructing a real-time, closed-loop "vision-servo" control system, enabling the nozzle oscillation of the cleaning device to strictly follow the real-time physical boundary of the photovoltaic panel below it. This system is physically integrated onto a mobile carrier (such as a robot or drone) carrying the cleaning device; however, the carrier is not the focus of this discussion. Essentially, the invention is a position servo loop that uses vision as a sensor and the nozzle as an actuator, which can be divided into a vision perception unit, a control decision unit, and an execution control unit.

[0044] After the system starts up, the cleaning device begins to move along the photovoltaic array, entering the main control loop. The visual perception unit starts working first, which is designed to convert the raw video stream from the camera into a precise "photovoltaic panel position signal" that can be understood by the control program.

[0045] In this embodiment, image acquisition is performed by using a top-down high-definition camera to capture real-time images of the scene below the device at a frequency of 10Hz. The resolution is 1920×1080, ensuring that the image remains clear even at a device operating speed of 0.1m / s and providing sufficient control response bandwidth. A 12mm focal length lens is selected, with the lens height controlled at 1.5 meters, ensuring that a single frame image covers a field of view slightly wider than that of a single photovoltaic panel, providing a buffer area for boundary recognition.

[0046] In this embodiment, during image preprocessing and segmentation, the acquired RGB images undergo a preprocessing process, including scaling to a fixed size (512×512 pixels) required by the segmentation network and normalizing pixel values. The images are then input into a lightweight and specially trained U-Net semantic segmentation model. This model is enhanced and trained on a photovoltaic panel image dataset, effectively distinguishing dark photovoltaic cell regions from the background (such as borders, supports, ground, and vegetation). The model output is a probability map of the same size as the input, where the value (0-1) of each pixel represents its confidence level in belonging to the photovoltaic panel region.

[0047] This invention uses an improved U-Net network as a photovoltaic panel region segmentation model, and its specific structure is shown in Table 1 below.

[0048] Table 1:

[0049] In the table, all convolutional layers (except the output layer) are followed by batch normalization (BN) and ReLU activation functions, and 3×3 convolutions with padding=1 are used to maintain the feature map size. Upsampling uses 2×2 transposed convolutions with a stride of 2. Skip connections are achieved by concatenating the channel dimensions. The model input is a 3-channel RGB image with a size of 512×512, and the output is a single-channel probability map with a size of 512×512.

[0050] The training dataset was constructed using 10,000 images collected from multiple photovoltaic power plants, including aerial views by drones and close-up views by robots, covering various lighting conditions, shadows, stains, dust, and reflections. The LabelMe tool was used to perform pixel-level fine-grained annotations on the photovoltaic panel regions in each image, generating a binary mask (0 for background). The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. During training, random horizontal flipping, random brightness / contrast adjustments (±20%), random Gaussian noise (σ≤0.05), and random affine transformations (rotation ±10°, scaling ±15%) were applied online to improve the model's generalization ability.

[0051] Training parameters are as follows: Loss function: Dice Loss + Binary Cross Entropy joint loss, weight ratio 1:1; Optimizer: Adam, initial learning rate 1e-3, decayed to 1e-5 using cosine annealing strategy; Batch size: 16; Epochs: 200, early stopping mechanism (if the Dice coefficient on the validation set does not increase for 20 consecutive rounds, the process stops). Hardware: Single NVIDIA RTX 3090 GPU, training time approximately 12 hours; The final model achieved an mIoU of 92.3% and a Dice coefficient of 94.7% on the test set.

[0052] During deployment, the model is accelerated using TensorRT or ONNX Runtime, with a single-frame inference time of <30ms (512×512 input), meeting the 10Hz real-time processing requirement. The input image undergoes the same normalization as during training (mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]), and the output probability map is directly used for post-processing after sigmoid activation.

[0053] The probabilistic map is binarized using a fixed threshold (0.5) to obtain the initial mask. Subsequently, morphological opening and closing operations are used to smooth the edges, followed by connected component filtering and a series of post-processing steps. This allows the system to stably and robustly extract an accurate, complete, and singular binary mask M_t representing the final effective region of the photovoltaic panel from the original image, which contains complex backgrounds and interference (such as reflections and similar objects). This mask represents the precise image projection of the target photovoltaic region to be cleaned, as observed from the camera's perspective at the current moment. White pixels (value 255) represent the photovoltaic panel, and black pixels (value 0) represent the background. The mask M_t serves as the sole input for visual servo control, providing a highly reliable guarantee for subsequent coordinate mapping.

[0054] Reference Figure 2 As shown, Figure 2 This diagram illustrates the coordinate mapping from image space to actuator space in the adaptive control method for photovoltaic panel cleaning nozzles according to the present invention. It is responsible for converting the photovoltaic panel pixel boundaries identified by the vision system into physical angle control commands for the nozzle oscillation mechanism. Figure 2 The left side shows the image plane (dashed rectangle). The solid rectangle represents a single frame image captured by the camera. The white solid block inside represents the binary mask M_t of the effective area of ​​the photovoltaic panel obtained after U-Net semantic segmentation and post-processing. The system scans the mask M_t and extracts the left and right boundary pixel coordinates of this area in the horizontal direction (X-axis) of the image: the minimum column coordinate X_left and the maximum column coordinate X_right. These two values ​​define the pixel width of the photovoltaic panel in the image. The pixel coordinates of the image centerline (corresponding to the camera's optical axis) are defined as X_center.

[0055] Before system deployment, camera calibration is required to determine the geometric relationship between the image pixel coordinate system and the world coordinate system with the nozzle rotation center as the origin. When the cleaning carrier moves smoothly, in actual deployment, taking a camera installation height of 1.5 meters and a top-down field of view of 60° as an example, the coverage width of a single frame image on the ground is approximately: 2 × 1.5 × tan(30°) ≈ 1.73 meters. This ensures that a standard-sized photovoltaic panel (approximately 1 meter wide) can appear completely in the image, with sufficient boundary detection buffer area. In this embodiment, the invention uses a simplified scale-offset model, which has sufficient accuracy under the condition that the camera optical axis is approximately perpendicular to the cleaning surface and the installation is stable. The scale factor k (unit: degrees / pixel) can be obtained through calibration experiments. Its physical meaning is the change in the nozzle swing angle corresponding to the offset of one pixel in the image. This value is determined by the camera focal length and installation height, and its calculation formula is: k equals camera field of view / image width, resulting in k ≈ 0.03125° / pixel. Furthermore, if the cleaning carrier is significantly disturbed in a complex environment, in order to achieve its self-adaptation, the present invention preferably uses an inertial measurement unit and a laser rangefinder to acquire the attitude and height information of the cleaning carrier in real time, and dynamically adjusts the coordinate mapping model accordingly to compensate for the pixel-angle mapping deviation caused by the carrier's large tilt or height change.

[0056] The nozzle oscillates with a single degree of freedom around its center of rotation. The angle when the nozzle points directly downwards is defined as the mechanical center angle θ_center (set to 0°). The limiting angle for the nozzle to oscillate to the left is θ_min, and the limiting angle for oscillation to the right is θ_max. Therefore, the mapping relationship from the image plane to the nozzle oscillation space can be described as: Mapping the image center pixel coordinate X_center to the mechanical midpoint angle θ_center of the nozzle, the formula for calculating the physical angle θ corresponding to any pixel coordinate X in the image is: θ equals k * (X - X_center) plus θ_center.

[0057] Therefore, the physical angular boundaries that the nozzle should cover are calculated in real time: θ_min equals k * (X_left minus X_center) plus θ_center; θ_max equals k * (X_right minus X_center) plus θ_center.

[0058] To further enhance the system's adaptability to complex environments on mobile cleaning vehicles (such as drones and robots), the preferred embodiment of this invention introduces an inertial measurement unit (IMU) and a laser rangefinder to collect the camera's three-axis attitude angles (pitch angle θp, roll angle θr, and yaw angle θy) and ground clearance H in real time. The system dynamically corrects the coordinate mapping model based on the IMU data, ensuring that nozzle control is always based on the photovoltaic panel boundary as seen from the vehicle's current actual perspective.

[0059] Height Adaptive: If the actual height H deviates from the calibrated height H0 (1.5m), the proportional correction coefficient k is equal to (H0 / H) * k0, where k0 is the proportional coefficient at the calibrated height.

[0060] Tilt Compensation: When the camera optical axis is not perpendicular to the ground, pixel offset is no longer linearly related to the physical offset of the ground. In this case, the simplified scale model is abandoned, and the complete perspective projection model is used: First, the pixel coordinates are converted into normalized camera coordinates through the camera intrinsic parameter matrix. Then, combined with the extrinsic parameter rotation matrix R and translation vector t obtained from the IMU, the camera coordinates are transformed into the world coordinate system with the nozzle rotation center as the origin. Finally, the desired angle θ_desired is calculated through inverse kinematics.

[0061] To enhance system robustness, considering edge localization errors in the image segmentation algorithm, camera calibration errors, and mechanical transmission clearances, a fixed safety margin δ (set to 1° in this embodiment) is introduced. This yields the final swing angle range used for control: [θ_min′, θ_max′] = [θ_min + δ, θ_max - δ]. The system further calculates the swing center θ_center' as (θ_min′ plus θ_max′) / 2 and the swing amplitude A as (θ_max′ minus θ_min′) / 2, used to generate a smooth, periodic swing trajectory. Based on the calculated swing angle range [θ_min′, θ_max′], the system generates the desired angle trajectory θ_desired(t) of the nozzle.

[0062] like Figure 1 As shown, Figure 1 This invention presents a flowchart of a dual-loop fusion adaptive servo control method for photovoltaic panel cleaning nozzles. The intelligent execution terminal, constituting the entire visual servo closed loop, is specifically designed to address the challenges of dynamic boundary tracking, resistance to environmental disturbances, and precise cleaning fluid dispensing in photovoltaic panel cleaning. It constructs a dual-layer control architecture with deep fusion of an "adaptive decision loop" and a "high-bandwidth servo loop," and possesses an "intelligent reflection path," achieving high-performance adaptive conversion from planning instructions to physical actions. The system's physical input is the desired angle trajectory θ_desired(t) from the preceding control decision unit, and the physical output is the actual oscillation angle θ_actual(t) of the nozzle. The detailed workflow of this control system is as follows: The control system in this invention consists of two loops operating in different frequency domains, each with its own function yet deeply collaborative. The adaptive decision loop operates at a visual frame rate (10Hz) and is responsible for handling low-frequency intelligent tasks such as photovoltaic panel boundary recognition, gap judgment, and cleaning mode decision-making. The high-bandwidth servo loop operates in a closed loop at a frequency of 500Hz, specifically responsible for high-precision position tracking. The core control law of the high-bandwidth servo loop is implemented by a digital PID (proportional-integral-derivative) controller, whose input is the error e(t) consisting of the desired trajectory θ_desired(t) and the actual angle θ_actual(t) fed back in real time by a high-precision absolute encoder (resolution better than 0.1°). The controller output drives a brushless DC servo motor, which in turn drives the nozzle to perform single-degree-of-freedom reciprocating oscillation via a linkage mechanism. The two loops are decoupled yet collaborative in terms of frequency and function, jointly ensuring rapid response and stable execution in dynamic scenarios. The high-bandwidth servo loop (500Hz) not only tracks the desired trajectory, but its closed-loop feedback itself has a good ability to suppress internal disturbances such as mechanical transmission gaps and high-frequency vibrations. The encoder provides real-time feedback θ_actual, and the PID controller immediately corrects deviations, enabling the system to maintain good static accuracy even in the presence of small, continuous vibrations (such as drone rotor vibration). For large shock disturbances, the fast response of the servo loop can also limit their impact to a single control cycle.

[0063] After receiving the θ_desired(t) command upon system startup, it performs scene analysis based on visual perception. In normal tracking mode, the command is processed step-by-step by the PID controller, driving the motor to execute and forming the main control loop. Its innovation lies in the fact that when the adaptive decision loop identifies the edge or gap of the photovoltaic panel based on real-time image recognition, it immediately generates a 'boundary event' signal. This signal sends a preset emergency stop or reverse motion command directly to the servo motor driver through a high-priority interrupt channel. This high-priority control path bypasses the conventional error calculation cycle of the PID controller in the high-bandwidth servo loop, achieving millisecond-level boundary response. In this invention, this is defined as the "intelligent reflection path," which ensures that the nozzle can achieve zero boundary crossing and zero empty spray when encountering a boundary. This is a core adaptive characteristic for the irregular structure of the photovoltaic array.

[0064] The system achieves deep collaboration and forward-looking compensation through bidirectional information flow. On one hand, the high-precision encoder in the inner loop reports the actual motion state, load, and external disturbances (such as vibrations caused by uneven ground) to the adaptive decision loop in real time via a state feedback stream, assisting it in making more accurate judgments and decisions regarding photovoltaic cleaning scenarios. On the other hand, the adaptive decision loop sends the predicted boundary change trends as feedforward information to the PID controller in the high-bandwidth servo loop in advance, enabling it to adjust its output in advance and achieve smoother, overshoot-free trajectory tracking. This information fusion mechanism significantly improves the system's tracking smoothness and overall robustness when facing continuous changes or sudden disturbances.

[0065] In addition, an independent water pump control module works in conjunction with the aforementioned motion control loop. It receives instructions from the main control unit and activates the spray only when the nozzle swings to cover the effective photovoltaic panel area, and immediately shuts off when returning or when an area without panels is detected, thereby achieving synchronous and precise spraying of the cleaning fluid.

[0066] In another embodiment, only the difference in angle tracking accuracy between the open-loop preset program and the closed-loop system of the present invention is compared; all other test conditions remain the same and consistent with the foregoing description of the present invention. The photovoltaic panel is set to be continuous and gapless, the water pump is running continuously, and the nozzle performs a complete cleaning cycle at a period of 2.0 seconds. The visual sampling frequency is set to 10Hz. The open-loop system has a preset fixed amplitude of ±15°, generating a fixed triangular wave desired trajectory θ_desired_fixed(t), while the closed-loop system dynamically adjusts the amplitude based on visual detection, dynamically generating the desired trajectory θ_desired_dynamic(t). The calculation method and the values ​​of the safety margin δ and proportional coefficient k in the process are consistent with the foregoing description of the present invention. The actual angles θ_actual_open and θ_actual_closed of the open-loop and closed-loop systems are collected by the angle encoder, respectively.

[0067] After the system has entered a stable operating state, a segment from the continuous working state is extracted for performance analysis. Table 2 below shows the performance data of a complete nozzle cleaning cycle (2.0s).

[0068] Table 2:

[0069] Table 2 reveals the fundamental differences between the traditional open-loop system and the closed-loop system of this invention in terms of tracking accuracy, adaptability, and long-term stability: At the first peak of motion (t=0.4s), the actual angle of the open-loop system (+12.12°) lagged significantly behind the expected value (+15.00°), with an error of -2.88°. This indicates that the open-loop system, lacking feedback, cannot overcome inherent delays such as transmission backlash and inertia, resulting in an inherent systematic tracking error.

[0070] At the same moment, the actual angle of the closed-loop system of this invention (+14.65°) closely matches the expected value (+14.86°) calculated based on the real-time board width, with an error of only -0.21°. This proves that the closed-loop feedback mechanism can compensate for system delays and disturbances in real time, achieving high-precision tracking.

[0071] The peak value of the desired trajectory θ_desired_dynamic in the closed-loop system of this invention is not a fixed value, but is slightly adjusted around 14.86° depending on the width of the visually detected board (e.g., the peak value difference between t=0.4s and t=1.4s). This demonstrates the system's "adaptive" capability in optimizing control commands in real time based on perceived information. In contrast, the commands in the open-loop system are fixed throughout the entire process, completely unaware of fluctuations in the actual board width, and lack any adaptability.

[0072] The open-loop system exhibits a significant error accumulation effect. At t=1.9s, when the expected command returns to the midpoint (0.00°), its actual angle remains at -3.53°, resulting in a large steady-state position error. At the start of the next cycle (t=2.0s), it is unable to keep up with the new command (actual -0.88° vs. expected +3.00°). This demonstrates that the error in the open-loop system accumulates and worsens over time and cannot self-recover.

[0073] The closed-loop system of this invention exhibits excellent stability. Throughout the entire 2.0s cycle, its actual angle consistently and closely tracks the dynamically changing command. At t=1.9s and t=2.0s, the system accurately returns to the position required by the command (with errors of -0.04° and +0.11°, respectively), without any error accumulation. This is thanks to the inherent "deviation detection and real-time correction" mechanism of closed-loop control.

[0074] To most intuitively reveal the essential differences between the two systems in terms of accuracy, stability, and non-cumulative nature, the following time series diagrams comparing the absolute values ​​of tracking errors |e(t)| of the open-loop and closed-loop systems are drawn based on the data in Table 2. Figure 3 In this diagram, the horizontal axis (X) represents time t (unit: s, 0.0s - 2.0s), and the vertical axis (Y) represents the absolute value of the tracking error |e(t)|, which equals |θ_actual(t) minus θ_desired(t)| (unit: °). The dark curve represents the open-loop system |e_open(t)|, which equals |θ_actual_open minus θ_desired_fixed|, and the light curve represents the closed-loop system of this invention |e_closed(t)|, which equals |θ_actual_closed minus θ_desired_dynamic|.

[0075] As shown in the figure, at t=0.4s (the positive peak point), the open-loop system represented by the dark curve has θ_desired_fixed reaching its peak value of +15.0° and is about to begin decelerating and reversing downwards (speed is zero). Due to the inertia of the motor and load, θ_actual_open cannot instantaneously follow the speed change of the command; its actual speed is still positive and is still "rushing" towards the peak value, leading to the first large error peak (2.88°). From t=0.5s onwards, the desired command has already turned downwards (speed is negative), and the actual angle, under the influence of inertia, has just reached its physical peak value and begins to decelerate and reverse. At this time, although the actual position is still lagging (position error exists), the directions of motion of both begin to converge (both are moving in the negative direction). Therefore, the absolute value of the position error |e(t)| will temporarily decrease. However, the error of the open-loop system does not return to zero after t=1.0s and continues to shift in the negative direction in the new cycle, forming a large steady-state error at t=2.0s. The error curve (dark line) of an open-loop system visually illustrates how the error fluctuates with changes in instructions, but it also demonstrates the fundamental flaw of open-loop systems—the lack of feedback, uncontrollable error, and the accumulation of errors.

[0076] In the figure, the error curve (light-colored line) of the closed-loop system of this invention remains close to the zero-error baseline throughout the entire 2.0s period, with minimal fluctuation (<0.3°) and no long-term deviation trend. In contrast, the dark-colored curve (open-loop) shows a peak error of nearly -3° at 0.4s, a difference in accuracy by an order of magnitude. This demonstrates the real-time correction capability of the closed-loop feedback of this invention. Figure 3 This demonstrates that the closed-loop system of this invention solves the fundamental problem of "error accumulation" that traditional open-loop systems cannot overcome, and achieves long-term stable and high-precision operation.

[0077] In another embodiment, the aim is to quantitatively compare the response performance of the traditional single-ring architecture and the dual-ring architecture of this invention when encountering photovoltaic panel boundaries using detailed millisecond-level time-series data. The core objective is to verify how the "intelligent reflection path" transforms the structural advantages of the dual-ring system into deterministic safety and energy-saving benefits.

[0078] In this embodiment, the time period data is the complete critical stage (0-50ms) captured from the occurrence of the boundary event to the system's initial response, as extracted from the experiment. Time 0 is a custom starting point, representing the 10ms preceding the boundary event. That is, visual recognition occurs at t=10ms, when the vision system completes processing the current frame and identifies the photovoltaic panel boundary (safety boundary θ_max' = 14.5°). The adaptive decision loop, with a frequency of 10Hz, is responsible for processing visual information, generating and updating the θ_desired(t) instruction, with an update cycle of 100ms. In the data segment selected in this example, the instruction is updated at t=32ms. The servo loop, with a frequency of 500Hz and a control cycle of 2ms, is responsible for tracking the current θ_desired(t) instruction with high precision.

[0079] In the comparative system of this example, the traditional single-loop system contains only an equivalent "decision + servo" hybrid single-loop, and visual events must wait until the next decision cycle (10Hz frequency) to update the θ_desired(t) instruction. In contrast, the dual-loop system structure of this invention includes an independent adaptive decision loop (10Hz) and a high-bandwidth servo loop (500Hz). After recognizing an event, the decision loop can send an emergency instruction to the driver within 2ms via a "smart reflection path," while simultaneously updating the regular θ_desired(t) instruction at a normal pace (10Hz frequency).

[0080] like Figure 4 As shown, after visual recognition of the boundary at t=10ms, the actual angle trajectory of the system of the present invention (light curve) with intelligent reflection path rapidly deviates from that of the traditional single-loop system (dark curve) after t=12ms. The traditional single-loop system (dark curve) must wait for the instruction cycle (frequency 10Hz) occurring at t=32ms in this example to update, resulting in an uncontrollable delay window of 22ms, causing its actual angle to continuously veer towards the danger zone, with a maximum deviation of 0.60°. In contrast, the system of the present invention (light curve), through intelligent reflection path, triggers direct braking within 2ms after visual recognition, strictly limiting the actual angle to the safe peak +14.39°. Furthermore, the water pump of the system of the present invention (light curve) begins to shut down approximately 2ms after recognition, while the water pump of the traditional single-loop system (dark curve) may not begin to shut down until at least 22ms later (after instruction update). This demonstrates that the reflection path essentially achieves the energy-saving benefits of "zero air jetting".

[0081] In summary, the system of this invention reduces the event response delay from 22ms to 2ms, thereby reducing the maximum out-of-bounds angle from 0.60° to 0° and the out-of-bounds duration from 22ms to 0ms. This demonstrates that the "intelligent reflection path" transforms the theoretical advantage of milliseconds into deterministic engineering benefits of "zero out-of-bounds" and "zero air spray" in photovoltaic cleaning operations.

[0082] In summary, the advantages of this invention are: 1. Significantly improved cleaning precision and reliability of photovoltaic panels: This invention constructs a real-time closed loop of "image perception - coordinate mapping - servo control". Because the system continuously identifies the photovoltaic panel boundary at a frequency of 10Hz and calculates control commands in real time, the nozzle's oscillation trajectory can dynamically and accurately conform to the actual physical boundary of the photovoltaic panel. This fundamentally solves the problem of missed cleaning and re-cleaning caused by position mismatch in open-loop control. Furthermore, the intelligent reflection path can respond in milliseconds when the boundary is detected, achieving physical-level boundary crossing protection and ensuring the absolute reliability of the cleaning action.

[0083] 2. Strong adaptability to complex field environments: The system does not rely on preset, fixed photovoltaic panel size and position data, but instead cleans what it sees in real-time. Therefore, regardless of the unevenness of the photovoltaic panels due to installation errors, or changes in perspective and relative position caused by carrier movement (such as robots climbing slopes or drones experiencing bumps), the system can automatically adapt. The deep learning-based (U-Net) visual segmentation model is robust to interference such as shadows and reflections, and can stably output effective area masks, providing reliable front-end perception for adaptive control.

[0084] 3. High utilization rate of resources such as water, cleaning fluid, and electricity: The control unit synchronously and precisely controls the start and stop of the water pump based on the real-time photovoltaic panel area mask. Spraying occurs only when the nozzles swing to cover the effective area, and immediately shuts off upon return or when no panel gap is detected. This avoids ineffective spraying of the support structure, ground, or gap areas as in traditional methods. Furthermore, the "safety margin δ" introduced in the coordinate mapping further tightens the effective cleaning range, minimizing redundant movements while ensuring full coverage and reducing energy consumption from idling.

[0085] 4. The system has a high degree of intelligence, reducing reliance on manual labor: This invention integrates identification, decision-making, and execution, achieving full-process autonomy. Operators only need to start the system and monitor its status, eliminating the need for precise path programming or frequent manual intervention, thus reducing operational difficulty and safety risks. The adaptive decision loop can handle scenarios such as determining the gaps between photovoltaic panels, making reasonable cleaning mode decisions and demonstrating the system's scenario understanding and decision-making capabilities.

[0086] All data collection and extraction in this invention are carried out under compliant and legal conditions.

Claims

1. An adaptive control method for cleaning nozzles on photovoltaic panels, characterized in that, include: Real-time acquisition of scene images below the cleaning device; semantic segmentation processing to obtain a binary mask of the effective area of ​​the photovoltaic panel; extraction of the left and right boundary pixel coordinates of the area in the image coordinate system. By using a pre-calibrated pixel angle mapping model, the nozzle swing angle boundary is solved, and the final control angle range is obtained by combining the safety margin. The swing center and amplitude are calculated to generate the desired angle trajectory. A dual-loop architecture is adopted to perform servo control. A low-frequency decision loop that matches the visual sampling frame rate handles scene analysis and trajectory updates, while a high-frequency servo loop with real-time angle feedback tracks the desired trajectory. The dual-loop architecture achieves collaboration through bidirectional information flow: the encoder of the high-bandwidth servo loop reports the actual movement status of the nozzle, motor load data, and external disturbance information during operation to the adaptive decision loop in real time through the status feedback link. The adaptive decision loop, based on the photovoltaic panel boundary change trend predicted by continuous frame images, transforms it into feedforward control information and sends it to the high-bandwidth servo loop in advance. When the boundary or gap of the photovoltaic panel is detected, a high-priority intelligent reflection command is triggered to coordinate the start and stop of the water pump and nozzle swing.

2. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 1, characterized in that, The process of obtaining the effective region binary mask of the photovoltaic panel through semantic segmentation includes scaling the acquired image to a fixed size, performing pixel value normalization preprocessing, and then inputting it into a pre-trained U-Net semantic segmentation model. The model outputs a probability map of the same size as the input image. The value of each pixel in the probability map represents the confidence level of belonging to the photovoltaic panel region. The probability map is binarized by setting a pre-set confidence threshold, and morphological opening and closing operations are performed to obtain the effective region binary mask.

3. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 2, characterized in that, The step of extracting the left and right boundary pixel coordinates of the region in the image coordinate system includes scanning the binary mask line by line, extracting the minimum and maximum column coordinates of the photovoltaic panel region in the horizontal direction of the image as the left and right boundary pixel coordinates of the photovoltaic panel region, and determining the pixel coordinates corresponding to the image center line as the reference center coordinates.

4. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 3, characterized in that, The calculation of the nozzle swing angle boundary includes determining the geometric correspondence between the image pixel coordinate system and the world coordinate system with the nozzle rotation center as the origin, determining the proportional coefficient of the change in nozzle swing angle corresponding to the pixel offset by the camera's field of view and image width parameters, mapping the pixel coordinates corresponding to the image center line to the mechanical median angle of the nozzle, and converting the left and right boundary pixel coordinates of the photovoltaic panel area into the left and right limit physical angles of the nozzle swing based on the calibrated proportional coefficient.

5. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 4, characterized in that, The calculation of the nozzle swing angle boundary also includes collecting the three-axis attitude angle and current ground clearance data of the camera on the cleaning device. When a deviation between the actual ground clearance of the camera and the calibrated height is detected, the pre-calibrated proportional coefficient is corrected according to the proportion of the height deviation. When a tilted state where the camera optical axis is not perpendicular to the cleaning surface is detected, the perspective projection model is switched, and the pixel coordinates are converted into normalized camera coordinates through the camera intrinsic parameter matrix. Then, combined with the extrinsic parameter rotation matrix and translation vector corresponding to the attitude data acquired in real time, the camera coordinates are converted into the world coordinate system with the nozzle rotation center as the origin, and the expected nozzle swing angle of the corresponding pixel coordinates is calculated.

6. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 3 or 5, characterized in that, The dual-loop architecture for servo control includes two control loops operating in different frequency domains. The adaptive decision loop operates at a frequency matching the visual sampling frame rate to perform photovoltaic panel boundary recognition, photovoltaic panel gap judgment, cleaning mode decision, and desired angle trajectory update. The high-bandwidth servo loop operates at a frequency much higher than the adaptive decision loop to perform high-precision position tracking of the nozzle oscillation. The core control law uses a PID controller, with the input being the angle error between the desired angle trajectory and the actual oscillation angle of the nozzle, and the output driving the servo motor to drive the nozzle to complete the oscillation action.

7. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 6, characterized in that, The triggering of the high-priority intelligent reflection command includes: when the edge or gap of the photovoltaic panel is detected, immediately generating a boundary event signal, and sending a preset emergency stop or reverse motion command directly to the driver of the servo motor through the high-priority interrupt channel.

8. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 6, characterized in that, The dual-loop architecture for servo control also includes the encoder in the high-bandwidth servo loop reporting the actual motion state of the nozzle, motor load data, and external disturbance information to the adaptive decision loop in real time through a state feedback stream. The adaptive decision loop will send the photovoltaic panel boundary change trend predicted by continuous frame images as a feedforward information stream to the controller of the high-bandwidth servo loop in advance.

9. The adaptive control method for photovoltaic panel cleaning nozzles according to claim 1 or 4, characterized in that, The process of obtaining the final control angle range by combining the safety margin includes introducing a fixed safety margin, correcting the calculated left and right limit physical angles to obtain the final swing angle range used for control; the safety margin is contracted inward based on the left and right boundary angles calculated by the pixel angle mapping model to form the final nozzle swing control range.

10. An adaptive control system for photovoltaic panel cleaning nozzles, employing the adaptive control method for photovoltaic panel cleaning nozzles as described in any one of claims 1-9, characterized in that, include: The visual perception unit acquires images of the photovoltaic panel in real time and outputs a binary mask of the effective area. The output end is connected to the input end of the control decision unit. The control decision unit establishes a real-time mapping between pixel coordinates and nozzle swing angle and generates the desired control trajectory. The output end is connected to the execution control unit. The execution control unit adopts a dual-loop control architecture.

Images