Detachable photovoltaic module cleaning robot, control method, intelligent carrier

Abstract

The application provides a detachable photovoltaic module cleaning robot, a control method and an intelligent carrying device. The method collects the state of a cleaning head in real time through a distance and inclination sensor, judges and records the motion inertia parameters and spatial pose at the moment of separation, combines three-dimensional geographic information of a photovoltaic array, and constructs geometric prior constraints of a target panel. By using kinematics simulation and multi-dimensional constraints, a smooth transition trajectory is generated to ensure the continuity of acceleration and the safety of the envelope, and the shoulder and elbow joints are driven in segments to realize the suspension pre-adjustment and precise contact closed-loop adjustment. After contact, the modeling data is adaptively optimized according to the real-time error to realize the smooth transition closed-loop control of multi-axis cooperation, and the switching accuracy and system robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of cleaning equipment, and more particularly to a detachable photovoltaic module cleaning robot, a control method, and an intelligent transport device. Background Technology

[0002] With the widespread deployment of distributed photovoltaic power stations, the autonomous operation capability of intelligent photovoltaic cleaning robots has become a key factor in improving cleaning efficiency and reducing operation and maintenance costs. In existing technologies, multi-axis robotic arm photovoltaic cleaning robots are gradually becoming mainstream. They achieve efficient coverage and automatic cleaning of large-scale photovoltaic array surfaces through multi-degree-of-freedom coordinated drive of shoulder and elbow joints. Common control methods in the industry mainly rely on algorithms such as closed-loop PID feedback, forward and inverse kinematics, and trajectory planning, combined with basic sensor information such as distance and tilt angle, to adjust the cleaning head's posture and height in real time, balancing panel adhesion quality and motion trajectory smoothness. As the need to adapt to complex photovoltaic array scenarios (such as staggered heights, intermittent spatial installations, and multi-row spacing) continues to increase, the industry is also beginning to explore technologies such as trajectory curve interpolation and end-effector servo servo to improve operational flexibility.

[0003] Typical applications of this technology are often seen in scenarios such as: a photovoltaic module cleaning robot based on multi-axis collaborative control, which uses motors to drive the shoulder and elbow joints, adjusting the height of the cleaning brush relative to the module surface based on real-time ranging signals to ensure adhesion and cleaning quality. Some solutions achieve pose alignment and path correction in complex geometric environments by adding module surface information acquisition points or upper-level visual navigation. In recent years, some high-end systems have begun to adopt dynamic trajectory replanning and multi-degree-of-freedom simulation, but these typically rely on high-performance industrial cameras, multi-point radar, or the addition of environmental sensors, which are costly and lack integrability in large-scale applications in actual photovoltaic power plants.

[0004] However, for photovoltaic panel arrays that are arranged discontinuously or laid in segments, existing multi-axis cleaning robots still face the following prominent problems when switching to the next adjacent or discontinuous panel after cleaning one panel: During the transition of the cleaning head from the current board surface to the target board surface, the coordinated movements of the robotic arm's shoulder and elbow joints are often out of sync due to robot inertia, nonlinear dynamics, and kinematic constraints. This can lead to sudden angle jumps, velocity reversals, or momentary jitters in the cleaning head's trajectory. Especially in real-world scenarios with significant variations in board surface height, tilt angle, and curvature, traditional closed-loop control can only respond to current position deviations and lacks prediction and smooth optimization of the suspended motion path. This makes the end effector prone to "impact," "head-on collisions," or localized overloads when landing or contacting the target board surface, affecting board safety and cleaning continuity.

[0005] To ensure a smooth transition in complex scenarios, some solutions attempt to introduce visual positioning or environmental 3D scanning to identify target panels. However, these solutions often significantly increase the number of sensors, computing resources, and the difficulty of hardware and software integration, limiting their applicability and reducing reliability in harsh outdoor environments.

[0006] Existing trajectory planning methods for multi-axis cooperative control mostly employ traditional offline interpolation or static constraint optimization, lacking trajectory smoothing algorithms based on dynamic memory and inertial continuity. The prediction and control of the next target area based on the motion state of the previous stage cannot achieve adaptive optimization, easily leading to accumulated errors and equipment fatigue during multiple switching operations, affecting the robot's service life and maintenance costs.

[0007] In summary, current photovoltaic cleaning robots suffer from structural technical deficiencies when performing multi-axis spatial height switching control between intermittent photovoltaic panels: First, they lack a control mechanism that relies on dynamic memory of the previous motion state and combines prior knowledge of the target panel surface to complete a seamless transition without sudden jumps; second, they lack lightweight embedded algorithm frameworks that do not require external vision to reduce hardware complexity, improve trajectory smoothness, safety, and adaptability; and third, they lack a feedback-optimization-memory adaptive iterative closed loop, resulting in insufficient self-learning ability during panel switching and difficulty in coping with the continuous self-optimization requirements of variable and complex photovoltaic scenarios.

[0008] Therefore, there is an urgent need to develop a smooth transition control method that can achieve multi-axis coordination of the shoulder and elbow joints and is based on a memory-prediction-optimization self-learning mechanism in complex intermittent photovoltaic scenarios without adding hardware structure or relying on expensive sensors and environmental recognition modules. This method would eliminate the risk of trajectory jumps, improve cleaning continuity, safety and adaptability, and thus make up for the shortcomings of existing technologies. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, this application provides a detachable photovoltaic module cleaning robot, comprising: A multi-axis robotic arm includes an upper arm and a forearm; the upper arm is driven by a shoulder joint pitch motor, and the proximal end of the forearm is hinged to the distal end of the upper arm and driven by an elbow joint pitch motor; both the shoulder joint pitch motor and the elbow joint pitch motor have corresponding servo drivers and encoders. A cleaning head is installed at the end of the forearm, and a cleaning brush is provided on the bottom surface of the cleaning head; the cleaning head is also provided with a distance sensor and a tilt angle sensor; The controller is electrically connected to the distance sensor and the tilt angle sensor respectively, and is electrically connected to the servo drivers corresponding to the shoulder joint pitch motor and the elbow joint pitch motor through a communication interface.

[0010] This application provides an intelligent transport device, including the aforementioned detachable photovoltaic module cleaning robot, and also includes a mobile lift; the mobile lift has a lifting platform, and a turntable structure is provided on the lifting platform; the base end of the multi-axis robotic arm is mounted on the turntable structure, and rotates relative to the lifting platform through the turntable structure.

[0011] This application provides a cleaning control method for use in a detachable photovoltaic module cleaning robot or an intelligent transport device, comprising the following steps: S1: Obtain the time-series data of the vertical distance between the bottom surface of the cleaning head and the current photovoltaic panel surface collected by the distance sensor, and determine whether the vertical distance continues to increase and exceeds the preset detachment threshold. If the condition is met, generate a detachment trigger signal. S2: Based on the detachment trigger signal, freeze and record the instantaneous spatial pose vector of the cleaning head, the shoulder joint pitch angle, the elbow joint pitch angle, and the instantaneous angular velocity and angular acceleration of the shoulder and elbow joints, and construct detachment memory modeling data containing the offset trend direction and rate. S3: Retrieve the installation tilt angle, edge position, and local curvature characteristic parameters of the next target photovoltaic panel from the pre-stored 3D geographic information database of the photovoltaic array, and generate the geometric prior constraint set of the target panel surface by combining the motion inertia parameters in the memory-free modeling data; S4: Using a kinematic simulation engine, the memory-free modeling data and the prior geometric constraint set of the target plate are input into the kinematic solution model to generate a non-jump transition trajectory sequence driven by the joint of the shoulder and elbow joints under the condition of satisfying the constraints of the joint motion range and acceleration continuity of the robotic arm. S5: Extract the first segment of the fine-tuning action segment of the non-jump transition trajectory sequence as the suspension pre-adjustment execution command, and inject the suspension pre-adjustment execution command into the servo driver of the shoulder joint pitch motor and the elbow joint pitch motor when the cleaning head is in the suspension transition period to complete the initial posture alignment and coarse height positioning. S6: Monitor the contact signal fed back by the distance sensor in real time. When the bottom surface of the cleaning head is detected to be in contact with the target photovoltaic panel surface again, switch to the closed-loop pose adjustment logic and execute the remaining fine-tuning segment of the non-jump transition trajectory sequence to achieve bonding. S7: Dynamically optimize the data update strategy for detached modeling during subsequent panel switching process based on the calculation results of real-time distance deviation and tilt angle deviation after contact.

[0012] Compared with the prior art, the beneficial effects of this application are: (1) By constructing a three-stage collaborative adjustment mechanism of "detachment modeling - target pre-simulation driving - suspended pre-adjustment execution", this application effectively overcomes the problems of response lag, trajectory jitter and contact impact caused by traditional multi-joint cleaning robots relying solely on real-time feedback when switching between discontinuous photovoltaic panel arrays. Existing technologies generally adopt closed-loop control strategies based on the current distance deviation, which are difficult to cope with sudden changes in the panel surface or height transitions. This often results in phenomena such as delayed adjustment of the robotic arm posture, impact of the end effector on the panel surface or unstable adhesion, which seriously affects the cleaning quality and structural life. This solution innovatively freezes and stores complete motion state parameters immediately when a detachment signal is detected, realizing the retention of dynamic characteristics such as pose, angular velocity and acceleration at the moment of detachment. It also combines the spatial geometric information of the next target panel surface to perform forward trajectory pre-simulation, enabling the control system to have memory capability in the time dimension and prediction capability in the spatial dimension. Based on this, a shoulder-elbow joint transition trajectory that satisfies multiple physical constraints is generated using a kinematic simulation engine. This allows for large-scale attitude alignment and coarse height adjustment during the air phase, significantly advancing the adjustment window and avoiding the passive response mode of "falling first and then rising" in traditional methods. This greatly reduces the risk of dynamic shock and vibration during the switching process between plates, and improves the continuity of operations and the reliability of the mechanism.

[0013] (2) This method, without adding extra sensors, modifying the mechanical structure, or introducing a complex visual recognition system, fully leverages the time-series characteristics of existing sensor data streams and the computational potential of embedded models, achieving a paradigm upgrade at the control strategy level. It possesses extremely high engineering practicality and deployment compatibility. Compared to schemes that rely on high-cost 3D vision or laser scanning for environmental perception, this application constructs a low-resource-consumption state pre-simulation framework by integrating a pre-stored 3D geographic information database of photovoltaic arrays with real-time motion inertial parameters. This framework can complete the localized generation and injection control of safe trajectories within milliseconds, ensuring that the adjustment process meets multiple constraints such as joint angle limits, angular acceleration continuity, and end-effector motion envelope boundaries, thus preventing trajectory singularities or motion abrupt changes. In particular, by pre-executing the first 60% of the pre-simulated trajectory in a high-fidelity manner during the suspension period, the robotic arm completes the main posture matching before contacting the target plate, greatly shortening the fine-tuning stroke and response time after contact, resulting in a smoother and more precise final bonding process. This "pre-judgment-pre-adjustment-micro-calibration" control logic not only significantly improves the system's adaptability to discontinuous and unevenly tilted photovoltaic panel groups, but also effectively reduces the wear and energy consumption caused by frequent start-stop and reverse adjustment of the servo system, thereby extending the service life of key transmission components and improving the overall operating efficiency and maintenance economy.

[0014] The combined effects of these technologies have formed a joint height adjustment and optimization system that is forward-looking, stable, and lightweight. This system represents a shift in control philosophy from "passive response" to "active pre-planning," significantly enhancing the autonomous operation capability and environmental adaptability of multi-joint cleaning robots in complex photovoltaic power station environments. Furthermore, this method possesses excellent scalability and general transfer potential, making it suitable for various mobile operation robot systems that require continuous operation between discrete surfaces. It is particularly well-suited for mountainous or rooftop photovoltaic power station scenarios with large terrain undulations, diverse installation angles, and irregular panel spacing, providing reliable technical support for improving the intelligent operation and maintenance level of clean energy. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of the structure of the detachable photovoltaic module cleaning robot is shown. Figure 2 A schematic diagram of the intelligent transportation device is shown. Figure 3 The main flowchart of the cleaning control method is shown; Figure 4 A sub-flowchart of the cleaning control method is shown; Figure 5 Another sub-flowchart of the cleaning control method is shown.

[0017] Explanation of key component symbols: 100 - Multi-axis robotic arm; 110 - Upper arm; 120 - Forearm; 130 - Shoulder joint pitch motor; 140 - Elbow joint pitch motor; 200 - Cleaning head; 210 - Housing; 220 - Cleaning brush; 230 - Distance sensor; 240 - Tilt angle sensor; 300 - Mobile lifting platform; 400 - Turntable structure. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] Example 1 When traditional photovoltaic module cleaning robots switch between intermittently arranged or segmented photovoltaic panel arrays, the coordinated movements of the shoulder and elbow joints of the multi-axis robotic arm are often asynchronous due to robot inertia, nonlinear dynamics, and kinematic constraints. This results in sudden angle jumps, speed reversals, or momentary jitters in the cleaning head trajectory. Especially in real-world scenarios with significant variations in panel height, tilt angle, and curvature, the end effector is prone to impacts, collisions, or localized overloads when landing on or contacting the target panel, affecting panel safety and cleaning continuity.

[0020] Please see Figure 1 To address this issue, this application proposes a detachable photovoltaic module cleaning robot, comprising a multi-axis robotic arm. The upper arm is driven by a shoulder joint pitch motor, and the proximal end of the forearm is hinged to the distal end of the upper arm and driven by an elbow joint pitch motor. Both the shoulder and elbow joint pitch motors have corresponding servo drivers and encoders. A cleaning head is mounted at the end of the forearm, with a cleaning brush on its bottom surface. The cleaning head also includes a distance sensor and a tilt angle sensor. A controller is electrically connected to the distance and tilt angle sensors, and electrically connected to the corresponding servo drivers of the shoulder and elbow joint pitch motors via a communication interface. This robot can perceive the attitude and position of the cleaning head in real time and precisely drive the robotic arm to solve the problems of trajectory instability and impact during the inter-module switching process.

[0021] The robot includes a multi-axis robotic arm designed as a combination of an upper arm and a forearm. The movement of the upper arm is driven by a shoulder joint pitch motor, while the proximal end of the forearm is hinged to the distal end of the upper arm and driven by an elbow joint pitch motor. This two-section robotic arm structure, through the coordinated movement of the two pitch joints, allows the cleaning head to perform flexible spatial positioning within a large working range.

[0022] Both the shoulder and elbow pitch motors have corresponding servo drivers and encoders. The servo driver receives motion commands from the controller and precisely controls the motor's speed, torque, and position. The encoder detects the motor's rotation angle or the actual position of the joint in real time and transmits these feedback signals to the servo driver and controller.

[0023] The cleaning head is mounted at the end of the forearm, and a cleaning brush is installed on its bottom surface. The cleaning brush is the component that directly contacts the surface of the photovoltaic module to perform the cleaning operation. For example, the cleaning brush can be fixedly installed, and the cleaning area can be covered by the overall movement of the robotic arm. The cleaning head is also equipped with a distance sensor and a tilt angle sensor. The distance sensor is used to measure the vertical distance between the bottom surface of the cleaning head and the surface of the photovoltaic module in real time. The tilt angle sensor is used to detect the pitch angle and roll angle of the cleaning head relative to the horizontal plane.

[0024] The controller is electrically connected to the distance sensor and tilt angle sensor, and also electrically connected to the servo drives corresponding to the shoulder and elbow joint pitch motors via a communication interface. The controller is the core of the entire system, responsible for receiving and processing sensor data, generating motion commands based on preset control logic, and sending these commands to the servo drives via a communication interface (e.g., industrial fieldbuses such as CAN bus or EtherCAT), thereby driving the robotic arm's motors to perform precise movements. For example, the controller can be an embedded microprocessor running pre-programmed motion control algorithms. In another implementation, the controller can be an industrial PC with greater computing power to handle complex kinematic and dynamic models. Through this connection method, the controller can achieve real-time monitoring and precise adjustment of the cleaning head's position and attitude, ensuring the continuity and safety of the cleaning operation.

[0025] This application constructs a robotic arm system with multi-axis collaborative capabilities through the aforementioned technical solution, integrating distance and tilt angle sensors, and cooperating with a controller to precisely control the servo drive. Thus, without adding additional complex vision hardware, real-time perception and dynamic adjustment of the cleaning head's posture and position are achieved. This robot effectively solves the problems of sudden jumps in the cleaning head's trajectory, reversed speed, or momentary jitter when switching between intermittent photovoltaic panel arrays, and avoids potential impacts, collisions, or localized overloads during landing or contact with the target panel surface, thereby improving the continuity, safety, and adaptability of the cleaning operation.

[0026] In some embodiments, the cleaning head includes a housing, and a cleaning brush is mounted on the side of the housing facing the photovoltaic module; a cleaning drive is provided inside the housing, and the output end of the cleaning drive is connected to the rotating shaft of the cleaning brush through a transmission component to drive the cleaning brush to rotate around its axis.

[0027] Specifically, the housing is the outer structure of the cleaning head. Its main function is to house and protect the cleaning drive components, transmission components, and other electronic components inside the cleaning head, while providing a stable mounting base for the cleaning brush. This housing can be made from various materials and molding processes. For example, lightweight, high-strength engineering plastics such as ABS or polycarbonate can be used, manufactured in one piece through injection molding to effectively reduce the overall weight of the cleaning head, lower the load on the robotic arm, and provide good corrosion resistance and electrical insulation.

[0028] The cleaning brush is mounted on the side of the housing facing the photovoltaic module. This mounting method ensures that the cleaning brush can directly and effectively contact the surface of the photovoltaic module to perform the cleaning task. The cleaning brush can be fixed in various ways, such as by bolts, clips, or by a design with a quick-release mechanism to securely install it on the bottom or side of the housing for quick replacement and routine maintenance after wear.

[0029] The housing houses a cleaning drive unit, which is the core component providing rotational power to the cleaning brush. This drive unit can utilize various types of motors. For example, a brushless DC motor (BLDC) can be selected, offering significant advantages such as compact size, high efficiency, long service life, and high control precision, making it ideal for use within the limited space inside the cleaning head.

[0030] The output end of the cleaning drive is connected to the rotating shaft of the cleaning brush via a transmission assembly. In this design, the transmission assembly plays a crucial role in efficiently and stably transmitting the rotational power generated by the cleaning drive to the cleaning brush, and can reduce speed, increase torque, or change the transmission direction as needed. For example, a gear transmission mechanism, such as spur gears, helical gears, or worm gears, can be used. By rationally designing the gear ratio, speed reduction and increased torque can be achieved, ensuring that the cleaning brush receives sufficient cleaning torque when contacting the photovoltaic panel surface. Another approach is to use a synchronous belt pulley transmission mechanism, which features smooth transmission, low operating noise, and no lubrication required, making it particularly suitable for environments with high cleanliness requirements or noise sensitivity. In certain specific cases, if the rotating shafts of the cleaning drive and the cleaning brush are coaxial and their speeds are matched, they can also be directly connected via a coupling.

[0031] By integrating independent cleaning drive components and transmission assemblies within the cleaning head, the cleaning brush achieves active self-rotation, thus decoupling the cleaning action from the overall macroscopic motion of the multi-axis robotic arm. This design transforms the cleaning head from a passive end effector that follows the robotic arm's posture adjustments into an active cleaning unit. When the multi-axis robotic arm moves the cleaning head to the photovoltaic module surface and maintains a relatively stable contact posture, the cleaning drive component inside the cleaning head drives the cleaning brush to rotate around its own axis via the transmission assemblies. This rotational motion, combined with the brush bristle structure, generates continuous and stable frictional and shearing forces, effectively loosening and removing stubborn stains such as dust, bird droppings, and limescale from the photovoltaic panel surface.

[0032] Example 2 Please see Figure 2 This embodiment provides an intelligent transport device, including the above-mentioned detachable photovoltaic module cleaning robot and a mobile lift; the mobile lift has a positioning and heading angle sensing module and is electrically connected to the controller; the mobile lift has a lifting platform, and a turntable structure is provided on the lifting platform; the base end of the multi-axis robotic arm is mounted on the turntable structure and rotates relative to the lifting platform through the turntable structure.

[0033] A mobile lifting platform is a mechanical device capable of autonomous or passive movement, providing vertical lifting capabilities to change the working height of the equipment mounted on it. For example, it can employ a scissor lift mechanism, driven by a hydraulic cylinder or electric screw, to achieve smooth vertical lifting.

[0034] The positioning and heading angle sensing module is a set of sensors used to acquire the precise position information of the transport device in the global coordinate system and its current orientation angle. It can map the target panel edge position from a pre-stored 3D geographic information database of the photovoltaic array to the coordinate system of the movable elevator, while providing the elevator's current orientation angle to correct coordinate system rotation deviations. For example, this module may include a GNSS (Global Navigation Satellite System) receiver for providing high-precision positioning, and an electronic compass or inertial measurement unit (IMU) for providing heading angle information. The data from these sensors, after fusion processing, can be used for coordinate system mapping and deviation correction.

[0035] A lifting platform is a planar structure on top of a movable lifting platform used to support a cleaning robot and move vertically with the lifting platform.

[0036] The turntable structure is mounted on a lifting platform and is a mechanical device capable of horizontal rotation, used to adjust the orientation of the equipment (multi-axis robotic arm base) mounted on it. For example, the turntable structure can use worm gear or rack and pinion transmission, driven by a servo motor, to achieve high-precision, wide-range rotation; or it can use a slewing bearing structure, driven by a hydraulic motor or electric motor, to provide greater load-bearing capacity and rotational torque.

[0037] The base of the multi-axis robotic arm is mounted on a turntable structure, which is designed to securely fix the multi-axis robotic arm on the rotatable turntable structure, thereby giving the robotic arm additional horizontal rotational freedom and expanding its working range.

[0038] By rotating the turntable relative to the lifting platform, a multi-axis robotic arm can change its working direction and coverage area without moving the entire lifting platform, improving operational flexibility and efficiency. This can be achieved by integrating a rotary drive mechanism (such as a motor or reducer) within the turntable structure, which, upon receiving commands from the controller, drives the turntable to rotate precisely around its vertical axis; or by connecting the turntable structure to a fixed drive source on the lifting platform via an external drive device (such as gears, chains, or belts) to achieve relative rotation.

[0039] The specific working principle of this solution is as follows: First, it should be noted that in large-scale photovoltaic power plants, the arrangement of photovoltaic module arrays is not completely continuous, but consists of multiple rows of modules with a certain spacing between each row, and the installation height or tilt angle of some rows is different.

[0040] This intelligent transport device includes a mobile lift with a lifting platform. A turntable structure is mounted on the lifting platform. A detachable photovoltaic module cleaning robot is mounted on this turntable structure; specifically, the base end of the cleaning robot's multi-axis robotic arm is mounted on the turntable structure. The mobile lift allows the cleaning robot to be transported between different rows of photovoltaic modules and its overall height can be adjusted to accommodate modules of varying heights. The turntable structure allows the multi-axis robotic arm to rotate horizontally relative to the lifting platform, thereby expanding the cleaning robot's working range and reducing the need for frequent lift movements.

[0041] The core of this cleaning robot is a multi-axis robotic arm, comprising a main arm and a forearm. The main arm is driven by a shoulder joint pitch motor, and the proximal end of the forearm is hinged to the distal end of the main arm and driven by an elbow joint pitch motor. Both the shoulder and elbow joint pitch motors are equipped with corresponding servo drivers and encoders for precise control and feedback of motor motion. The cleaning head is mounted at the end of the forearm. A cleaning brush is located on the bottom surface of the cleaning head. Inside the cleaning head is a housing, with the cleaning brush mounted on the side of the housing facing the photovoltaic modules. A cleaning drive unit is also located within the housing; its output is connected to the rotation shaft of the cleaning brush via a transmission assembly, driving the brush to rotate around its axis, thereby improving cleaning efficiency. The cleaning head is also equipped with distance and tilt sensors for real-time sensing of the relative position and orientation of the cleaning head to the photovoltaic module surface. A controller is electrically connected to the distance and tilt sensors, and also electrically connected to the servo drivers corresponding to the shoulder and elbow joint pitch motors via a communication interface.

[0042] In actual operation, the cleaning robot first cleans the first row of photovoltaic modules. The controller receives real-time data from distance and tilt sensors to determine the distance and tilt angle between the cleaning head and the module surface. Based on this data, the controller calculates the required motion commands for the shoulder and elbow joint pitch motors and sends them to the corresponding servo drives via the communication interface. The servo drives precisely adjust the posture of the upper arm and forearm, ensuring the cleaning brushes are always in optimal contact with the module surface at the optimal distance and angle. The encoder provides real-time feedback on the precise position of the motors, ensuring the accuracy of the closed-loop control.

[0043] When the cleaning robot finishes cleaning the current row of components and needs to transition to the adjacent next row of components, which are taller and have a different tilt angle, this system demonstrates its advantages. Unlike existing technologies that rely solely on feedback control based on the current position deviation, which can easily lead to sudden angle jumps, unstable speeds, or momentary jitters during the transition, this cleaning robot employs a memory-prediction-optimization control strategy. The controller first "memorizes" the motion state of the cleaning head at the end of the current component, including its speed and acceleration. Then, combining this with prior information such as the position, height, and tilt angle of the next row of target components (either preset or obtained through other means), the controller "predicts" and plans a smooth transition trajectory from the edge of the current component to the starting point of the target component. This trajectory is designed to avoid abrupt changes in speed or direction during the transition, ensuring the continuity and stability of the motion. Subsequently, the controller "optimizes" the motion commands of the multi-axis robotic arm, decomposing the predicted smooth trajectory into a coordinated motion sequence of shoulder and elbow joint pitch motors. These commands are sent to the servo drivers via a communication interface. Throughout the entire suspended transition, distance and tilt sensors continuously monitor the relative state of the cleaning head to the environment. Based on this real-time feedback, the controller fine-tunes the predicted trajectory to ensure the accuracy of the cleaning head's movement in the air. Servo drives and encoders work together to precisely execute the motor's movement, allowing the boom and forearm to move smoothly along the optimized trajectory. The cleaning head traces a smooth curve in the air, avoiding the "heavy impact," "head-on collision," or localized overload phenomena commonly found in traditional solutions.

[0044] As the cleaning head approaches the target component surface, distance and tilt sensors provide high-precision proximity information. The controller then fine-tunes the motor motion to ensure the cleaning head descends smoothly and controllably onto the new component surface. Once on contact, the cleaning process continues, with the controller continuously adjusting the robotic arm's posture based on sensor feedback to ensure cleaning quality. This feedback-optimization-memory adaptive iterative closed-loop control mechanism enables the cleaning robot to achieve seamless transitions when dealing with complex, discontinuous photovoltaic module arrays, improving the safety, continuity, and adaptability of cleaning operations while reducing reliance on expensive external vision or 3D scanning modules.

[0045] Example 3 Please see Figures 3-5 This embodiment provides a cleaning control method applicable to detachable photovoltaic module cleaning robots or intelligent transport devices.

[0046] The cleaning control methods specifically include: S1: Acquire the time-series data of the vertical distance between the bottom surface of the cleaning head and the current photovoltaic panel surface collected by the distance sensor, and determine whether the vertical distance continues to increase and exceeds the preset detachment threshold. If the condition is met, generate a detachment trigger signal.

[0047] S2: Based on the detachment trigger signal, freeze and record the instantaneous spatial pose vector of the cleaning head, the shoulder joint pitch angle, the elbow joint pitch angle, and the instantaneous angular velocity and angular acceleration of the two joints to construct detachment memory modeling data containing the offset trend direction and rate.

[0048] S3: Retrieve the installation tilt angle, edge position, and local curvature characteristic parameters of the next target photovoltaic panel from the pre-stored 3D geographic information database of the photovoltaic array, and generate the geometric prior constraint set of the target panel surface by combining the motion inertia parameters in the memory-free modeling data.

[0049] S4: Using a kinematic simulation engine, the memory-free modeling data and the prior geometric constraints of the target plate are input into the kinematic solution model. Under the condition of satisfying the constraints of the range of motion and acceleration continuity of the robotic arm joints, a non-jump transition trajectory sequence jointly driven by the shoulder and elbow joints is generated.

[0050] S5: Extract the first segment of the fine-tuning motion segment of the non-jump transition trajectory sequence as the suspension pre-adjustment execution command, and inject the suspension pre-adjustment execution command into the servo drivers of the shoulder joint pitch motor and elbow joint pitch motor when the cleaning head is in the suspension transition period to complete the initial posture alignment and coarse height positioning.

[0051] S6: Real-time monitoring of the contact signal fed back by the distance sensor. When the bottom surface of the cleaning head is detected to be in contact with the target photovoltaic panel surface again, it seamlessly switches to the closed-loop pose adjustment logic and executes the remaining fine-tuning segment of the non-jump transition trajectory sequence to achieve bonding.

[0052] S7: Based on the calculation results of real-time distance deviation and tilt angle deviation after contact, dynamically optimize the data update strategy of memory-free modeling in the subsequent plate switching process to form a multi-axis collaborative smooth transition control closed loop with adaptive capabilities.

[0053] Step S1 specifically includes: S1.1: Perform analog-to-digital conversion and noise filtering on the raw voltage signal output by the distance sensor at a fixed sampling frequency to obtain a discrete sampling sequence of the vertical distance between the bottom surface of the cleaning head and the current photovoltaic panel surface, which includes timestamp information, as the raw input data for subsequent trend analysis.

[0054] S1.2: Construct a fixed-length sliding time window based on the vertical spacing discrete sampling sequence, and perform a first-order difference operation on the vertical spacing discrete sampling sequence within the window to generate an instantaneous velocity vector sequence characterizing the rate of change of the vertical spacing.

[0055] S1.3: Perform continuous positive judgment logic processing on the instantaneous velocity vector sequence, count the number of velocity components greater than zero in the instantaneous velocity vector sequence within the preset observation period, and generate a continuity flag bit that represents the continuous increasing trend of vertical spacing.

[0056] S1.4: Compare the latest sampled value in the vertical spacing discrete sampling sequence with the pre-stored detachment threshold constant, and perform joint condition judgment in combination with the logic state of the continuity flag bit to generate a detachment judgment result that indicates that the cleaning head has completely detached from the current plate surface.

[0057] Obtain the continuous flag state value generated by S1.3, and the latest sampled data of the vertical spacing discrete sampling sequence in S1.1. This latest sampled value represents the instantaneous physical distance between the bottom surface of the cleaning head and the surface of the photovoltaic panel during the current system clock cycle.

[0058] The preset detachment threshold constant is read from the parameter configuration area stored in the controller. This constant is set according to the maximum allowable suspension height of the brush body and the dynamic response delay characteristics of the robotic arm, and is used to define the critical boundary between effective cleaning contact and complete detachment.

[0059] A numerical comparison operation is performed to determine whether the latest sampled value is strictly greater than the departure threshold constant. If the condition is met, it is determined that the cleaning head has met the departure distance condition in spatial position, and a departure intermediate determination signal is generated.

[0060] Read the logic level of the continuity flag bit to verify whether the rate of change of vertical spacing continues to increase positively within the preset observation period. This step aims to eliminate false judgments of instantaneous distance jumps caused by sensor noise or local protrusions on the board surface.

[0061] A logical AND operation is performed on the distance separation intermediate determination signal and the continuity flag state. Only when both are true simultaneously is it confirmed that the cleaning head is not only in a long-distance state, but also in a stable separation process, thereby generating a separation determination result indicating that the cleaning head has completely separated from the current plate surface.

[0062] Through the aforementioned joint condition judgment mechanism, the trend analysis results of the previous step are transformed into a definite state switching command, which realizes high-reliability detection of the instantaneous switching of intermittent photovoltaic panels and effectively avoids the false triggering of control logic caused by signal jitter.

[0063] For example, the detachment threshold constant is set to 50mm, corresponding to the distance at which the brush body completely leaves the board surface while retaining a safety margin. The preset observation period is 200ms, and the sampling frequency is 100Hz. When the latest sampled value is 52.5mm, and the continuity flag shows that the velocity components of the past 20 sampling points are all greater than 0, the logical AND operation result is true, a high-level detachment judgment result is output, and the subsequent pose freeze process is triggered.

[0064] S1.5: Based on the state-triggered interrupt response mechanism where the disengagement determination result is true, immediately lock the current system clock and output a high-level pulse signal to generate a disengagement trigger signal for freezing subsequent joint state data.

[0065] The logic state signal indicating that the exit determination result is true is received from step S1.4 and used as the initial input condition for triggering the interrupt response mechanism.

[0066] The interrupt service routine inside the controller detects this logical state transition, immediately reads the current count value of the system's high-precision timer, and locks the precise timestamp of the escape event.

[0067] Based on the locked timestamp, the controller sends a synchronization freeze command to the data acquisition bus, forcibly pausing the encoder data refresh buffers of the shoulder pitch motor and elbow pitch motor.

[0068] Meanwhile, the controller outputs a high-level pulse signal with a fixed pulse width to the servo drive module through the general purpose input / output interface (GPIO). This pulse signal serves as a hardware-level trigger-off flag.

[0069] After receiving a high-level pulse, the servo driver immediately latches the current value of the motor rotor position register and the instantaneous angular velocity value output by the speed observer.

[0070] Through the aforementioned interrupt response and hardware latching mechanism, the detachment determination result of the previous step is transformed into a detachment trigger signal containing precise time reference and frozen joint state data, thereby achieving lossless capture of the motion state of the cleaning head at the moment of detachment and providing definite initial boundary conditions for subsequent construction of detachment memory modeling data.

[0071] For example, when the distance sensor detects that the vertical spacing continuously increases within 50ms and the latest sampled value exceeds 20mm, S1.4 outputs the true value. The controller responds to the interrupt within 10μs and locks the timestamp t=1024.5ms. The GPIO port outputs a high-level pulse with a pulse width of 100μs and an amplitude of 3.3V. The shoulder joint encoder latches an angle of 45.2 degrees and an angular velocity of 0.5rad / s; the elbow joint encoder latches an angle of -30.1 degrees and an angular velocity of -0.2rad / s. This process ensures the synchronization of state data within a microsecond-level time window, eliminates state estimation errors caused by communication delays, and significantly improves the initial accuracy of trajectory pre-simulation.

[0072] Step S2 specifically includes: S2.1: Obtain the distance sensor timing data and tilt angle sensor reading at the moment the release trigger signal takes effect. Use the coordinate transformation matrix to map the vertical distance of the bottom surface of the cleaning head and the tilt angle relative to the horizontal plane to the coordinate system of the mobile vehicle base, and generate the instantaneous spatial pose state vector to establish the absolute position reference and attitude reference of the cleaning head in three-dimensional space.

[0073] In response to the interrupt request triggered by the departure signal, the system immediately locks the current clock cycle and synchronously reads the latest vertical distance sampling value output by the distance sensor and the real-time tilt angle reading output by the tilt angle sensor.

[0074] Construct a homogeneous transformation matrix from the local coordinate system of the cleaning head to the coordinate system of the mobile vehicle base. This matrix includes rotational components determined by the angles of the base rotary motor, shoulder joint pitch motor, and elbow joint pitch motor, as well as translational components determined by the length of the robotic arm link.

[0075] The vertical distance measured by the distance sensor is used as the displacement component along the normal of the cleaning head. Combined with the pitch and roll angles relative to the horizontal plane measured by the tilt angle sensor, the three-dimensional position vector of the center point of the bottom surface of the cleaning head in the local coordinate system is calculated.

[0076] By using the homogeneous transformation matrix to perform linear mapping operations on the local position vector, the absolute spatial coordinates of the center point of the bottom surface of the cleaning head in the coordinate system of the mobile vehicle base are calculated, and the position reference is established.

[0077] Simultaneously, the tilt angle sensor data is mapped to the base coordinate system through the Euler angle transformation algorithm to generate a quaternion or rotation matrix representing the attitude of the cleaning head, thus establishing the attitude reference.

[0078] The absolute spatial coordinates are combined and encapsulated with the base attitude matrix to generate an instantaneous spatial pose state vector containing six degrees of freedom information.

[0079] By using coordinate transformation and multi-sensor data fusion processing, the discrete sensor readings from the previous step are transformed into instantaneous spatial pose vector data under a unified reference system, thereby establishing the absolute position and attitude reference at the moment the cleaning head leaves, and providing accurate initial boundary conditions for subsequent trajectory pre-simulation.

[0080] For example, the distance sensor sampling value is set to 0.15 meters, and the tilt angle sensor reading is 30 degrees for pitch and 0 degrees for roll. A homogeneous transformation matrix T containing the current joint angles of the robotic arm is constructed. Multiplying the local coordinates (0, 0, 0.15) on the left by matrix T yields the position vector P in the base coordinate system. base =(1.2, 0.5, 0.8). Convert the tilt angle to a rotation matrix R. base The encapsulation yields the pose vector State=[1.2, 0.5, 0.8, R]. base This process eliminates the influence of sensor installation errors and significantly improves pose memory accuracy.

[0081] S2.2: Based on the instantaneous spatial pose state vector, read the encoder feedback pulses of the shoulder joint pitch motor and the elbow joint pitch motor, and calculate the real-time rotation angle of the two joints through the speed observer algorithm inside the servo driver to generate the shoulder joint pitch angle and the elbow joint pitch angle, so as to quantify the specific configuration parameters of the upper arm and forearm of the robotic arm in the vertical plane.

[0082] The system receives the raw encoder feedback pulse data from the shoulder and elbow joint servo drives at the moment the disengagement trigger signal takes effect, and uses it as the underlying input source for configuration calculation.

[0083] The original encoder pulse sequence is subjected to direction discrimination and counting accumulation processing. Combined with the motor pole pair number and reduction ratio parameters, the pulse increment is mapped to the absolute number of rotations and relative angular displacement of the mechanical shaft system, generating uncompensated original joint angle observation values.

[0084] The servo driver's internal speed observer algorithm module, which is pre-stored in the controller's non-volatile memory area, is invoked. This module is based on the extended Kalman filter principle, uses the original joint angle observation value as the measurement input, and the motor current loop feedback torque as the control input, to construct a state-space equation that includes the moment of inertia and friction coefficient of the robotic arm link.

[0085] By using the prediction steps of the state-space equations and the estimated values ​​of joint angular velocity and angular acceleration from the previous control cycle, the prior state vector at the current moment is calculated, thus forming a forward-looking prediction of the joint motion trend.

[0086] During the update step, the residual between the predicted angle corresponding to the prior state vector and the actual encoder observation angle is calculated. The residual is then weighted and corrected using the Kalman gain matrix to suppress the measurement jitter caused by encoder quantization noise and transmission backlash.

[0087] The optimal state estimate is output after Kalman filtering and smoothing. The shoulder joint pitch angle and elbow joint pitch angle are extracted as quantitative parameters to characterize the precise geometric configuration of the upper arm and forearm in the vertical plane.

[0088] The servo drive uses a speed observer algorithm to denoise and estimate the state of the original encoder data, converting the noisy pulse signal into high-fidelity shoulder and elbow pitch angle data. This enables precise quantification of the robot arm's instantaneous configuration parameters, providing jitter-free initial boundary conditions for subsequent trajectory pre-simulation.

[0089] For example, the shoulder joint servo motor encoder resolution is set to 17 bits, and the reduction ratio is 100:1. At the moment of trigger disengagement, the original pulse count value is read as 1048576, corresponding to an original angle of 180 degrees. The speed observer sampling period is set to 1ms, the process noise covariance Q is set to 0.01, and the measurement noise covariance R is set to 0.1. When a pulse jump is detected that causes a sudden change in the original angle of 0.5 degrees, the observer makes a prediction based on the angular velocity of 0.2 rad / s at the previous moment, and calculates that the predicted angle deviation is only 0.0002 degrees. Combined with the Kalman gain K=0.09, the final output shoulder joint pitch angle correction is 0.045 degrees, effectively filtering out high-frequency noise caused by gear backlash and ensuring the continuity and smoothness of the calculated angle.

[0090] S2.3: Perform time series difference operation on the pitch angle of the shoulder joint and the pitch angle of the elbow joint, calculate the rate of change of angular displacement in combination with the sampling interval of the servo control cycle, and further differentiate the rate of change of angular displacement to obtain the angular acceleration value, generate the instantaneous angular velocity of the shoulder joint, the instantaneous angular velocity of the elbow joint, the instantaneous angular acceleration of the shoulder joint and the instantaneous angular acceleration of the elbow joint, so as to characterize the motion inertia characteristics of the robotic arm end effector at the moment of disengagement.

[0091] Based on the real-time rotation angle sequence of the shoulder and elbow joints obtained by S2.2, the angle sampling points within a preset time window before the moment of departure from the trigger signal are extracted to construct a discrete time series dataset containing historical posture information.

[0092] A first-order backward difference operation is performed on the discrete time series dataset to calculate the angular displacement increment between adjacent sampling times. Combined with the fixed sampling period of the servo control system, the preliminary estimates of the instantaneous angular velocities of the shoulder and elbow joints are obtained.

[0093] The central difference method is used to perform a second differential on the preliminary estimate of the instantaneous angular velocity to eliminate numerical noise interference and accurately obtain the instantaneous angular acceleration of the shoulder joint and the instantaneous angular acceleration of the elbow joint, so as to quantify the dynamic inertial state of the robotic arm at the moment of disengagement.

[0094] The joint angular velocity and angular acceleration at the k-th sampling moment are calculated using the following formulas:

[0095] Where, ω k Let θ be the angular velocity at time k. k Let be the joint angle at time k, and Δt be the sampling period.

[0096]

[0097] Where, α k Let be the angular acceleration at time k.

[0098] The calculated angular velocity and angular acceleration data are subjected to moving average filtering to suppress high-frequency jitter and generate smoothed instantaneous angular velocities of the shoulder joint, elbow joint, shoulder joint, and elbow joint.

[0099] By using the above differential and filtering processing methods, the angle position data from the previous step is transformed into velocity and acceleration technical indicators that characterize motion inertia, thereby achieving precise quantification of the dynamic characteristics of the robotic arm at the moment of disengagement and providing reliable initial dynamic boundary conditions for subsequent trajectory prediction.

[0100] S2.4: The least squares method is used to fit the evolution trend of the instantaneous spatial pose state vector within a preset time window before separation. The displacement direction component and velocity component of the cleaning head relative to the current plate surface normal are extracted to generate the offset trend direction and offset trend rate to describe the dynamic separation trajectory characteristics when the cleaning head leaves the contact.

[0101] S2.5: The instantaneous spatial pose state vector, shoulder joint pitch angle, elbow joint pitch angle, shoulder joint instantaneous angular velocity, elbow joint instantaneous angular velocity, shoulder joint instantaneous angular acceleration, elbow joint instantaneous angular acceleration, offset trend direction and offset trend rate are structured and encapsulated, and written into the controller's non-volatile storage unit to construct memory-free modeling data, so as to form a complete initial state snapshot for driving subsequent trajectory pre-simulation.

[0102] like Figure 4 As shown, step S3 specifically includes: S3.1: Based on the data reading instruction activated by the detachment trigger signal, perform an index query operation on the pre-stored photovoltaic array three-dimensional geographic information database to extract the installation tilt angle scalar, edge position coordinate vector and local curvature feature matrix of the next target photovoltaic panel, and generate the static geometric parameter set of the target panel surface.

[0103] S3.2: Based on the static geometric parameter set of the target plate, the coordinate transformation algorithm is used to map the edge position coordinate vector to the coordinate system of the moving vehicle base, and the normal unit vector of the target plate is calculated by combining the installation tilt angle scalar to generate the spatial pose reference vector of the target plate.

[0104] Read the edge position coordinate vector of the next target photovoltaic panel from the pre-stored 3D geographic information database of the photovoltaic array. This vector is defined in the global geodetic coordinate system and includes longitude, latitude and altitude components.

[0105] Obtain the real-time positioning data and heading angle data of the mobile vehicle in the global geodetic coordinate system, and construct the rotation and translation transformation matrix from the global geodetic coordinate system to the mobile vehicle base coordinate system.

[0106] A linear transformation operation is performed on the edge position coordinate vector of the target photovoltaic panel using a rotation and translation transformation matrix to eliminate the difference between the coordinate system origin offset and axial rotation, thereby generating the relative edge position coordinate vector of the target photovoltaic panel in the coordinate system of the moving vehicle base.

[0107] Extract the installation tilt angle scalar of the target photovoltaic panel, which represents the degree of tilt of the photovoltaic panel plane relative to the horizontal plane, and determine the azimuth deviation of the photovoltaic panel arrangement direction relative to the forward direction of the moving vehicle.

[0108] Based on the relative edge position coordinate vector and the installation tilt angle scalar, a local geometric model of the target photovoltaic panel plane is constructed to determine the spatial extension direction of the photovoltaic panel surface in the coordinate system of the mobile vehicle base.

[0109] Based on the installation tilt angle scalar and azimuth angle deviation, the cosine value of the normal direction on the target photovoltaic panel surface is calculated to construct the initial normal vector.

[0110] The initial normal vector is normalized to eliminate vector magnitude error and generate a standardized target plate surface normal unit vector.

[0111] The relative edge position coordinate vector and the target plate surface normal unit vector are structurally combined to form the target plate surface spatial pose reference vector containing position reference and attitude reference.

[0112] By using coordinate mapping and normal calculation, the static geometric parameters from the previous step are transformed into the target plate spatial pose reference vector in the coordinate system of the moving vehicle base, thereby achieving the expected technical effect of accurate alignment of the target attitude in subsequent trajectory planning.

[0113] For example, the target photovoltaic panel's edge position in the global coordinate system is (100.5, 200.3, 15.2), and the current position of the moving vehicle is (100.0, 200.0, 15.0) with a heading angle of 0 degrees. After constructing the transformation matrix, the relative edge position coordinate vector is calculated to be (0.5, 0.3, 0.2). The installation tilt angle is 30 degrees, and the azimuth deviation is 0 degrees. The calculated normal unit vector is (0, -0.5, 0.866). The final generated target panel spatial pose reference vector includes the position (0.5, 0.3, 0.2) and the normal (0, -0.5, 0.866), significantly improving the initial alignment accuracy of trajectory planning.

[0114] S3.3: Based on the offset trend direction and offset trend rate in the data detached from memory, and combined with the instantaneous angular velocity of the shoulder joint, the instantaneous angular velocity of the elbow joint, the instantaneous angular acceleration of the shoulder joint, and the instantaneous angular acceleration of the elbow joint, a motion inertial parameter vector characterizing the dynamic characteristics of the robotic arm end effector at the moment of separation is constructed.

[0115] S3.4: Based on the target plate's spatial pose reference vector and motion inertial parameter vector, perform multi-dimensional parameter fusion calculation, and superimpose the local curvature feature matrix as a safety margin factor onto the motion envelope boundary to generate a target plate geometric prior constraint set containing position constraints, attitude constraints, and dynamic continuity constraints.

[0116] S3.5: Based on the prior geometric constraints of the target plate, perform data structure serialization and encapsulation processing, convert it into a standard input format that can be recognized by the kinematic simulation engine, and generate a standardized constraint input stream to drive the generation of non-jump transition trajectory sequences.

[0117] Step S4 specifically includes: S4.1: Based on the instantaneous spatial pose state vector and motion inertial parameters in the memory-free modeling data, combined with the installation tilt angle and edge position information in the geometric prior constraint set of the target plate, a multi-dimensional state space mapping model including the initial boundary conditions and the termination boundary conditions is constructed to establish the initial solution interval of the shoulder joint pitch angle and the elbow joint pitch angle in the time domain.

[0118] S4.2: Perform joint motion range verification on the initial solution range output by the multidimensional state space mapping model, eliminate angle combinations that exceed the physical limits of the upper arm and forearm, and generate a set of feasible posture candidates that conform to the safety boundary of the mechanical structure envelope, providing a reasonable search space for subsequent trajectory planning.

[0119] S4.3: Using the kinematic simulation engine, perform acceleration continuity constraint mapping operations on the feasible posture candidate set, calculate the rate of change of angular acceleration of the shoulder joint pitch motor and elbow joint pitch motor at each discrete time step, and screen out smooth motion path segments with angular acceleration change rate less than the preset abrupt change threshold to form a preliminary prototype of a jump-free trajectory.

[0120] Read the joint angle data of two adjacent discrete time steps from the feasible posture candidate set, and denote them as the shoulder joint angle θ at the current time t. s(t) elbow joint angle θ e(t) And the shoulder joint angle θ at the next moment t+Δt s(t+Δt) elbow joint angle θ e(t+Δt) .

[0121] Based on the fixed sampling period Δt of the servo control system, the instantaneous angular velocity of each joint at the midpoint is calculated using the central difference method, where the instantaneous angular velocity ω of the shoulder joint is... s Through formula Obtain the instantaneous angular velocity ω of the elbow joint. e Similarly, calculations are performed to eliminate the influence of numerical noise on speed estimation.

[0122] Perform the difference operation again on the angular velocity sequence to obtain the instantaneous angular acceleration of each joint at the discrete time step, where the instantaneous angular acceleration α of the shoulder joint is... s Through formula This parameter is confirmed to directly reflect the changing trend of motor torque.

[0123] The absolute value of the difference in angular acceleration between adjacent time steps is calculated and defined as the Jerk index, which is used to quantify the impact degree during motion. If the Jerk index of a certain path segment exceeds a preset abrupt change threshold J... max If so, it is determined that the segment has a risk of trajectory jump.

[0124] Iterate through all path segments in the feasible pose candidate set and remove those with a Jerk metric greater than J. max For non-smooth paths, retain path segments that satisfy acceleration continuity constraints and splice them together in time sequence to form a preliminary prototype of abrupt trajectory.

[0125] Through the above acceleration continuity constraint mapping and filtering process, the geometrically feasible posture set from the previous step is transformed into preliminary non-jump trajectory prototype data with dynamic smoothness characteristics, thereby achieving the expected technical effect of avoiding rigid impact during the switching of the robotic arm on the board surface.

[0126] For example, the servo control cycle Δt is set to 0.01s, and the preset mutation threshold J is set. maxThe angular velocity is 50 rad / s³. In a certain candidate path segment, the shoulder joint angular velocity is 2.0 rad / s at t=0.1s, 2.1 rad / s at t=0.11s, and 2.25 rad / s at t=0.12s. The calculated angular acceleration is 10 rad / s² at t=0.105s and 15 rad / s² at t=0.115s, resulting in a rate of change of angular acceleration of 500 rad / s³. Because 500 rad / s³ is much greater than the threshold of 50 rad / s³, this path segment is deemed to have a risk of sudden jumps and is eliminated. The system continues to search other candidate paths until a smooth path sequence with a rate of change of angular acceleration consistently below 50 rad / s³ is found, ensuring a significant improvement in the stability of the washing head's movement during the suspension transition period.

[0127] S4.4: Based on the initial non-jump trajectory prototype, perform multi-segment polynomial interpolation optimization processing, perform high-order smooth fitting on the shoulder joint pitch angle sequence and elbow joint pitch angle sequence, eliminate velocity reversal and jerk abrupt changes at path nodes, and generate the original transition trajectory sequence of shoulder and elbow joint joint drive with strict acceleration continuity.

[0128] S4.5: Verify the end effector motion envelope safety of the original transition trajectory sequence driven by the shoulder and elbow joints, simulate the spatial scanning volume of the cleaning head during the suspended transition period and perform collision detection with the local curvature features of the target plate, and output the final safety-verified non-jump transition trajectory sequence as the direct data source for the suspended pre-adjustment execution command.

[0129] like Figure 5 As shown, step S5 specifically includes: S5.1: Based on the timestamp index in the non-jump transition trajectory sequence, the shoulder joint pitch angle time series data and the elbow joint pitch angle time series data are segmented and processed to obtain the suspended pre-adjustment action segment covering the interval from the beginning of the suspended transition period to 60% of the total duration.

[0130] S5.2: The inverse kinematics algorithm is used to map and transform the pose vector of the end effector in the suspended pre-tuned motion segment to generate a discretized joint angular displacement command set corresponding to the shoulder joint pitch motor and the elbow joint pitch motor.

[0131] S5.3: Based on the discretized joint angular displacement instruction set, the trapezoidal velocity curve planning algorithm is used to smoothly reconstruct the joint angular velocity parameters to output the shoulder joint pitch motor control quantity sequence and elbow joint pitch motor control quantity sequence that satisfy the acceleration continuity constraint.

[0132] S5.4: The shoulder joint pitch motor control sequence and the elbow joint pitch motor control sequence are injected in real time into the position loop input terminals of the shoulder joint pitch motor servo driver and the elbow joint pitch motor servo driver via the fieldbus communication protocol, so as to drive the upper arm and forearm to perform the initial alignment action.

[0133] Receive the discretized joint angular displacement instruction set of the shoulder and elbow joints after reconstruction by trapezoidal velocity curve planning. This instruction set contains timestamp indices that satisfy acceleration continuity constraints and corresponding target angle values.

[0134] The discrete joint angular displacement instruction set is encapsulated using a communication protocol. Based on the CANopen or EtherCAT fieldbus communication standard, the target angles of the shoulder joint and elbow joint in each control cycle are mapped to PDO (Process Data Object) or CoE (CANopen over EtherCAT) service data units.

[0135] During the encapsulation process, a synchronization flag and a cyclic redundancy check code are added to the angle data of each joint to ensure the real-time performance and integrity of data transmission and prevent instruction packet loss or distortion caused by electromagnetic interference.

[0136] The packaged data frames are broadcast to the position loop input registers of the shoulder joint pitch motor servo driver and the elbow joint pitch motor servo driver at a fixed control cycle via a high-speed fieldbus interface.

[0137] The position loop controller inside the servo drive receives the target angle command and, in conjunction with the current actual angle fed back by the encoder, calculates the position deviation error signal.

[0138] Based on the proportional-integral-derivative control algorithm, the position deviation is converted into motor torque command, which drives the servo motor to produce the corresponding rotational motion.

[0139] Driven by servo motors, the upper arm and forearm perform coordinated movements strictly according to the first segment of the pre-programmed trajectory, eliminating the lag effect caused by mechanical transmission gaps.

[0140] The control sequences of the shoulder joint pitch motor and the elbow joint pitch motor are injected in real time into the position loop input terminals of the shoulder joint pitch motor servo driver and the elbow joint pitch motor servo driver via the fieldbus communication protocol, so as to drive the upper arm and forearm to perform initial posture alignment, thereby realizing rapid posture response and coarse spatial positioning of the cleaning head during the suspension stage.

[0141] For example, the fieldbus communication cycle is set to 1ms, the target angle sequence for the shoulder joint is [30.5°, 31.2°, ..., 45.0°], and the target angle sequence for the elbow joint is [-15.0°, -14.5°, ..., 0.0°]. The controller encapsulates the angle value per millisecond into an 8-byte data frame and sends it to the driver via the EtherCAT bus. The servo driver position loop gain is set to Kp=500, Ki=10, and Kd=50. Within a 100ms pre-tuning execution period, the actual angle tracking error of the shoulder joint remains within ±0.05°, the tracking error of the elbow joint remains within ±0.03°, the tilt angle deviation of the bottom surface of the cleaning head relative to the target board surface rapidly converges from the initial 5° to within 0.5°, and the height position deviation is reduced to within 2mm, significantly improving the smoothness and alignment accuracy of the inter-board switching process.

[0142] S5.5: Based on the real-time vertical distance data fed back by the distance sensor, the amplitude gain of the currently executing shoulder joint pitch motor control sequence and elbow joint pitch motor control sequence is dynamically adjusted to complete the coarse positioning of the cleaning head relative to the target photovoltaic panel surface and maintain the stability of the movement in the suspended state.

[0143] The vertical distance between the bottom surface of the cleaning head and the surface of the target photovoltaic panel is sampled in real time by the distance sensor and used as the feedback input signal for coarse height positioning.

[0144] A sliding window mean filter is applied to the vertical spacing sample values ​​to eliminate high-frequency noise interference and extract smooth distance features that characterize the current suspension height.

[0145] Calculate the residual value between the smooth distance feature and the pre-stored ideal contact distance of the target plate surface, and generate a distance deviation scalar for correcting the trajectory amplitude.

[0146] Based on the distance deviation scalar, the amplitude correction coefficients of the shoulder joint pitch motor and the elbow joint pitch motor are calculated using the proportional gain adjustment algorithm, and a dynamic gain matrix is ​​constructed.

[0147] The dynamic gain coefficient is calculated using the following formula:

[0148] Where K is the dynamic gain coefficient, k is the proportional adjustment factor, d is the current smoothing distance feature, and d0 is the ideal contact distance of the target plate surface.

[0149] The dynamic gain coefficient is multiplied by the angle command values ​​in the currently executing shoulder joint pitch motor control sequence and elbow joint pitch motor control sequence to generate a joint angular displacement command set after amplitude correction.

[0150] Perform forward kinematics verification on the joint angular displacement command set after amplitude correction to ensure that the corrected end pose is still within the safe suspended area above the target plate.

[0151] The corrected joint angular displacement command set is sent to the servo driver in real time via fieldbus, driving the boom and forearm to adjust their spatial posture synchronously.

[0152] By using a dynamic gain adjustment processing method, the distance deviation of the previous step is converted into the amplitude correction parameter of the joint control quantity, so as to achieve coarse positioning of the cleaning head during the suspension transition period and maintain the stability of the movement.

[0153] For example, the target contact distance d0 is set to 5mm, and the proportional adjustment factor k is 0.2. When the distance sensor reports a current smoothing distance d of 15mm, the dynamic gain coefficient K is calculated to be 1.4. This coefficient is applied to the original shoulder joint angle command of 30° and the elbow joint angle command of -45°, resulting in corrected commands of 42° and -63°, respectively. The servo driver executes the correction commands, causing the cleaning head to quickly approach the target height while maintaining a stable posture, avoiding violent shaking, and significantly improving the smoothness of the transition between boards.

[0154] Step S6 specifically includes: S6.1: Perform high-frequency sampling and sliding window filtering on the vertical spacing time-series data collected by the distance sensor to extract the vertical spacing change rate feature and generate a contact judgment flag, which serves as the input condition for triggering the control mode switching.

[0155] S6.2: Based on the state transition signal of the contact determination flag, execute the dynamic transfer logic of the servo driver control authority, smoothly transfer the control of the shoulder joint pitch motor and elbow joint pitch motor from the trajectory tracking module to the deviation feedback adjustment module, so as to generate a mode switching completion status signal.

[0156] S6.3: Utilize the mode switching completion status signal to activate the breakpoint continuation mechanism of the non-jump transition trajectory sequence, extract the angle increment instruction set in the remaining fine-tuning segment, and construct the residual trajectory execution queue as the driving basis for precise positioning.

[0157] S6.4: Based on the residual trajectory, execute the angle increment instruction set in the queue, combine it with the real-time calculated distance deviation and tilt angle deviation, and execute the multi-axis collaborative impedance compensation algorithm to generate the final motor torque instruction containing the stiffness adaptive adjustment.

[0158] S6.5: Based on the final motor torque command, drive the shoulder joint pitch motor and elbow joint pitch motor to perform micro-displacement correction actions to eliminate residual pose error at the end of the robotic arm and achieve constant pressure and precise bonding between the bottom surface of the cleaning head and the surface of the target photovoltaic panel.

[0159] Receive the final motor torque command sequence generated by S6.4, which includes stiffness adaptive adjustment, and decompose it into independent torque control components for the shoulder joint pitch motor and the elbow joint pitch motor.

[0160] Feedforward compensation calculations are performed on the torque control components of the shoulder joint pitch motor, and static gravity compensation torque based on the gravity model of the upper arm and forearm is superimposed to counteract the interference of the robot arm's own gravity on the end contact pressure.

[0161] Friction force identification and compensation are performed on the torque control component of the elbow joint pitch motor. Based on the pre-stored Coulomb friction and viscous friction coefficient model, the frictional resistance torque at the current angular velocity is calculated and superimposed in the opposite direction to eliminate the crawling phenomenon caused by nonlinear friction of the transmission chain.

[0162] The compensated net torque command is converted into a three-phase winding drive current through the current loop controller, which drives the shoulder and elbow joint servo motors to produce a small rotational displacement.

[0163] The vertical spacing change rate fed back by the distance sensor is collected in real time. If the rate of decrease in spacing exceeds the preset safety threshold, a virtual damping torque term is immediately introduced to suppress the elastic rebound caused by contact impact.

[0164] The equivalent stiffness of the joint is dynamically adjusted by using an impedance control law, so that the cleaning head exhibits compliant characteristics when it comes into contact with the target panel, thus avoiding rigid collision damage to the photovoltaic glass surface.

[0165] By using a multi-axis collaborative force-position hybrid control method, the final motor torque command is converted into a constant normal contact force between the bottom surface of the cleaning head and the surface of the target photovoltaic panel, achieving a shock-free and high-precision constant pressure bonding effect.

[0166] For example, the shoulder joint gravity compensation torque is set to 15 N·m, and the elbow joint friction compensation coefficient is set to 0.05 N·m / (rad / s). When a contact signal is detected, a virtual damping torque of 2 N·m is applied to limit the contact impact force to within 5 N. Ultimately, the pressure fluctuation range between the cleaning head and the plate surface is controlled within ±0.5 N, significantly improving cleaning uniformity.

[0167] Step S7 specifically includes: S7.1: Obtain the distance deviation residual sequence and tilt angle deviation residual sequence at the moment of contact, and perform weighted moving average filtering on the distance deviation residual sequence and tilt angle deviation residual sequence to extract the steady-state error feature vector characterizing the current plate switching accuracy.

[0168] S7.2: Based on the steady-state error feature vector, the pre-set memory-free correction coefficient matrix is ​​identified and updated online using the recursive least squares algorithm to generate a dynamic correction gain matrix containing historical error compensation factors.

[0169] S7.3: Read the target panel geometric prior constraint set from the pre-stored photovoltaic array 3D geographic information database, apply the dynamic correction gain matrix to the installation tilt angle parameters and edge position parameters of the target panel geometric prior constraint set, and reconstruct an updated version of the target panel geometric prior constraint set with error self-compensation capability.

[0170] S7.4: Call the offset trend direction and rate data in the memory-free modeling data generated in the previous loop, combine it with the updated target plate geometric prior constraint set, and recalculate the expected values ​​of shoulder joint pitch angle and elbow joint pitch angle through the inverse kinematics mapping function to generate the optimized memory-free modeling data update package.

[0171] S7.5: Write the optimized memory-free modeling data update package into the controller's non-volatile storage unit to replace the original memory-free modeling data, so as to complete the adaptive iterative update of the multi-axis collaborative smooth transition control closed loop and ensure that the generation benchmark of the floating pre-adjustment execution command is more accurate when the panel switches next time.

[0172] Receive the optimized memory-free modeling data update packet generated by S7.4. This data packet contains the expected angles, offset trend directions, and rate parameters of the shoulder and elbow joints after calibration with a dynamically corrected gain matrix.

[0173] Perform CRC cyclic redundancy check on the update package to verify the integrity and consistency of data transmission, ensuring that storage write operations are based on an error-free data source.

[0174] Access the designated logical address sector of the non-volatile memory unit through the controller's internal bus interface. This sector is pre-divided into independent storage blocks for storing snapshots of the current board switching state.

[0175] Execute the flash erase command to clear the charge state of the physical storage unit occupied by the original data that was removed from memory modeling, and eliminate the potential interference of historical data residue on the accuracy of subsequent reading.

[0176] The optimized data that passes the verification is serialized according to the preset byte alignment format and written to the erased storage sector to establish a new pose reference index.

[0177] Update the version number and timestamp in the stored metadata to indicate the latest state after the completion of this adaptive iteration, so that it can be called when the trigger signal is generated again next time.

[0178] By using the above data persistence processing method, the error compensation result calculated in the previous step is transformed into reusable initial state reference data, realizing the adaptive iterative update of the multi-axis collaborative smooth transition control closed loop, and ensuring that the generation reference of the suspended pre-adjustment execution command is more accurate when the panel switches next time.

[0179] For example, a dedicated sector 0x0A00-0x0AFF of 256 bytes is partitioned in the non-volatile memory. The optimized data packet contains floating-point numbers such as shoulder joint angle 35.2 degrees, elbow joint angle -12.8 degrees, and offset rate 0.05 m / s, totaling 128 bytes. The system calculates the CRC32 checksum as 0x9F2A3B1C, and after confirming its accuracy, sends a Flash erase command to sector 0x0A00. After the erase is completed, the 128 bytes of data, along with 8 bytes of header information (including version number V2.1 and timestamp T=1024ms), are written sequentially. After writing, the data in this sector is read and compared to confirm that all bytes are consistent. When the cleaning robot detaches from the photovoltaic panel next time, the controller directly reads the data in this sector as the initial value for trajectory planning. Due to the introduction of error compensation from the previous round, the initial attitude deviation of the new trajectory is reduced from 0.5 degrees to 0.05 degrees, significantly improving the smoothness of the transition.

[0180] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A cleaning control method, characterized in that, Applied to detachable photovoltaic module cleaning robots or intelligent transport devices, the process includes the following steps: S1: Obtain the timing data of the vertical distance between the bottom surface of the cleaning head and the current photovoltaic panel surface, determine whether the vertical distance continues to increase and exceeds the preset detachment threshold, and generate a detachment trigger signal if so. S2: Based on the detachment trigger signal, freeze and record the instantaneous spatial pose vector of the cleaning head, the shoulder joint pitch angle, the elbow joint pitch angle, and the instantaneous angular velocity and angular acceleration of the shoulder and elbow joints to construct detachment memory modeling data containing the offset trend direction and rate. S3: Retrieve the installation tilt angle, edge position, and local curvature feature parameters of the next target photovoltaic panel from the pre-stored 3D geographic information database of the photovoltaic array, and generate the geometric prior constraint set of the target panel surface by combining the motion inertia parameters in the memory-free modeling data; S4: Using a kinematic simulation engine, the memory-free modeling data and the prior geometric constraints of the target plate are input into the kinematic solution model. Under the condition of satisfying the constraints of the joint motion range and acceleration continuity of the robotic arm, a non-jump transition trajectory sequence jointly driven by the shoulder and elbow joints is generated. S5: Extract the fine-tuning action segment of the non-jump transition trajectory sequence as the suspension pre-adjustment execution command. When the washing head is in the suspension transition period, inject the command into the servo driver of the shoulder joint pitch motor and the elbow joint pitch motor to complete the initial posture alignment and coarse height positioning. S6: Real-time monitoring of the contact signal fed back by the distance sensor. When the bottom surface of the cleaning head is detected to be in contact with the target photovoltaic panel surface again, switch to closed-loop pose adjustment logic and execute the remaining fine-tuning segment of the non-jump transition trajectory sequence to achieve bonding. S7: Dynamically optimize the data update strategy for detached modeling during subsequent panel switching process based on the calculation results of real-time distance deviation and tilt angle deviation after contact.

2. The cleaning control method according to claim 1, characterized in that, It also includes the following steps: S2.1: Obtain the distance sensor timing data and tilt angle sensor reading at the moment the release trigger signal takes effect, and use the coordinate transformation matrix to map the vertical distance of the bottom surface of the cleaning head and the tilt angle relative to the horizontal plane to the coordinate system of the mobile vehicle base, and generate the instantaneous spatial pose state vector to establish the absolute position reference and attitude reference of the cleaning head in three-dimensional space. S2.2: Based on the instantaneous spatial pose state vector, read the encoder feedback pulses of the shoulder joint pitch motor and the elbow joint pitch motor, read the real-time rotation angle of the shoulder joint and elbow joint through the servo driver, and generate the shoulder joint pitch angle and the elbow joint pitch angle to quantify the specific configuration parameters of the upper arm and forearm of the robotic arm in the vertical plane. S2.3: Perform time series difference operation on the shoulder joint pitch angle and elbow joint pitch angle, calculate the angular displacement change rate in combination with the sampling interval of the servo control cycle, and further differentiate the angular displacement change rate to obtain the angular acceleration value, and generate the instantaneous angular velocity of the shoulder joint, the instantaneous angular velocity of the elbow joint, the instantaneous angular acceleration of the shoulder joint and the instantaneous angular acceleration of the elbow joint to characterize the motion inertia characteristics of the robotic arm end effector at the moment of disengagement; S2.4: The least squares method is used to fit the evolution trend of the instantaneous spatial pose state vector within a preset time window before separation, and the displacement direction component and velocity component of the cleaning head relative to the current plate surface normal are extracted to generate the offset trend direction and offset trend rate to describe the dynamic separation trajectory characteristics when the cleaning head separates from the contact. S2.5: The instantaneous spatial pose state vector, shoulder joint pitch angle, elbow joint pitch angle, shoulder joint instantaneous angular velocity, elbow joint instantaneous angular velocity, shoulder joint instantaneous angular acceleration, elbow joint instantaneous angular acceleration, offset trend direction and offset trend rate are structured and encapsulated, and written into the controller's non-volatile storage unit to construct memory-free modeling data, so as to form a complete initial state snapshot for driving subsequent trajectory pre-simulation.

3. The cleaning control method according to claim 1, characterized in that, It also includes the following steps: S3.1: Based on the data reading instruction activated by the detachment trigger signal, perform an index query operation on the pre-stored photovoltaic array three-dimensional geographic information database to extract the installation tilt angle scalar, edge position coordinate vector and local curvature feature matrix of the next target photovoltaic panel, and generate a set of static geometric parameters of the target panel surface; S3.2: Based on the static geometric parameter set of the target plate surface, the edge position coordinate vector is mapped to the coordinate system of the mobile vehicle base using a coordinate transformation algorithm, and the target plate surface normal unit vector is calculated by combining the installation tilt angle scalar to generate the target plate surface spatial pose reference vector; S3.3: Based on the offset trend direction and offset trend rate in the memory-free modeling data, and combined with the instantaneous angular velocity of the shoulder joint, the instantaneous angular velocity of the elbow joint, the instantaneous angular acceleration of the shoulder joint, and the instantaneous angular acceleration of the elbow joint, a motion inertial parameter vector characterizing the dynamic characteristics of the robotic arm end effector at the moment of separation is constructed. S3.4: Based on the target plate surface spatial pose reference vector and the motion inertial parameter vector, perform multi-dimensional parameter fusion operation, and superimpose the local curvature feature matrix as a safety margin factor onto the motion envelope boundary to generate a target plate surface geometric prior constraint set containing position constraints, attitude constraints and dynamic continuity constraints. S3.5: Based on the prior geometric constraints set of the target plate, perform data structure serialization and encapsulation processing to convert it into a standard input format that can be recognized by the kinematic simulation engine, and generate a standardized constraint input stream for driving the generation of abrupt transition trajectory sequences.

4. The cleaning control method according to claim 1, characterized in that, It also includes the following steps: S4.1: Based on the instantaneous spatial pose state vector and motion inertial parameters in the memory-free modeling data, combined with the installation tilt angle and edge position information in the geometric prior constraint set of the target plate, a multi-dimensional state space mapping model including the initial boundary conditions and the final boundary conditions is constructed to establish the initial solution interval of the shoulder joint pitch angle and the elbow joint pitch angle in the time domain. S4.2: Perform joint motion range verification processing on the initial solution range output by the multidimensional state space mapping model, eliminate angle combinations that exceed the physical limits of the upper arm and forearm, generate a set of feasible posture candidates that conform to the safety boundary of the mechanical structure envelope, and provide a reasonable search space for subsequent trajectory planning. S4.3: Use the kinematic simulation engine to perform acceleration continuity constraint mapping operation on the feasible posture candidate set, calculate the rate of change of angular acceleration of the shoulder joint pitch motor and elbow joint pitch motor at each discrete time step, screen out smooth motion path segments with angular acceleration change rate less than the preset abrupt change threshold, and form a preliminary non-jump trajectory prototype. S4.4: Based on the preliminary non-jump trajectory prototype, perform multi-segment polynomial interpolation optimization processing, perform high-order smooth fitting on the shoulder joint pitch angle sequence and elbow joint pitch angle sequence, eliminate velocity reversal and jerk mutation at path nodes, and generate a shoulder-elbow joint joint driven original transition trajectory sequence with strict acceleration continuity. S4.5: Perform end effector motion envelope safety verification on the original transition trajectory sequence driven by the shoulder and elbow joints, simulate the spatial scanning volume of the cleaning head during the suspended transition period and perform collision detection with the local curvature features of the target plate surface, and output the final safety-verified non-jump transition trajectory sequence as the direct data source for the suspended pre-adjustment execution command.

5. The cleaning control method according to claim 1, characterized in that, It also includes the following steps: S5.1: Based on the timestamp index in the non-jump transition trajectory sequence, the shoulder joint pitch angle time series data and the elbow joint pitch angle time series data are segmented and truncated to obtain the suspended pre-adjustment action segment covering the interval from the beginning stage of the suspended transition period to 60% of the total duration. S5.2: The kinematic inverse algorithm is used to map and transform the end effector pose vector in the suspended pre-tuned motion segment to generate a discretized joint angular displacement command set corresponding to the shoulder joint pitch motor and the elbow joint pitch motor. S5.3: Based on the discretized joint angular displacement instruction set, the trapezoidal velocity curve planning algorithm is used to smoothly reconstruct the joint angular velocity parameters to output the shoulder joint pitch motor control quantity sequence and elbow joint pitch motor control quantity sequence that satisfy the acceleration continuity constraint. S5.4: The shoulder joint pitch motor control sequence and the elbow joint pitch motor control sequence are injected in real time into the position loop input terminals of the shoulder joint pitch motor servo driver and the elbow joint pitch motor servo driver through the fieldbus communication protocol, so as to drive the upper arm and forearm to perform the initial posture alignment action. S5.5: Based on the real-time vertical distance data fed back by the distance sensor, the amplitude gain of the currently executing shoulder joint pitch motor control sequence and elbow joint pitch motor control sequence is dynamically adjusted to complete the coarse positioning of the cleaning head relative to the target photovoltaic panel surface and maintain the stability of the movement in the suspended state.

6. The cleaning control method according to claim 1, characterized in that, It also includes the following steps: S6.1: Perform high-frequency sampling and sliding window filtering on the vertical spacing time-series data collected by the distance sensor to extract the vertical spacing change rate feature and generate a contact judgment flag, which serves as the input condition for triggering the control mode switching. S6.2: Based on the state transition signal of the contact determination flag, execute the dynamic transfer logic of the servo driver control authority, smoothly transfer the control authority of the shoulder joint pitch motor and elbow joint pitch motor from the trajectory tracking module to the deviation feedback adjustment module, so as to generate a mode switching completion status signal. S6.3: Activate the breakpoint continuation mechanism of the non-jump transition trajectory sequence using the mode switching completion status signal, extract the angle increment instruction set in the remaining fine-tuning segment, and construct the residual trajectory execution queue as the driving basis for precise positioning. S6.4: Based on the residual trajectory, execute the angle increment instruction set in the queue, and combine it with the distance deviation and tilt angle deviation calculated in real time to execute the multi-axis collaborative impedance compensation algorithm to generate the final motor torque instruction containing the stiffness adaptive adjustment. S6.5: Based on the final motor torque command, drive the shoulder joint pitch motor and elbow joint pitch motor to perform micro-displacement correction actions to eliminate residual pose error at the end of the robotic arm.

7. The cleaning control method according to claim 1, characterized in that, It also includes the following steps: S7.1: Obtain the distance deviation residual sequence and tilt angle deviation residual sequence at the moment of contact, and perform weighted moving average filtering on the distance deviation residual sequence and tilt angle deviation residual sequence to extract the steady-state error feature vector characterizing the current plate switching accuracy; S7.2: Based on the steady-state error feature vector, the preset memory-free correction coefficient matrix is ​​identified and updated online using the recursive least squares algorithm to generate a dynamic correction gain matrix containing historical error compensation factors; S7.3: Read the target panel geometric prior constraint set from the pre-stored photovoltaic array 3D geographic information database, and apply the dynamic correction gain matrix to the installation tilt angle parameter and edge position parameter of the target panel geometric prior constraint set to reconstruct an updated version of the target panel geometric prior constraint set with error self-compensation capability; S7.4: Call the offset trend direction and rate data in the memory-free modeling data generated in the previous loop, and combine it with the updated target plate surface geometric prior constraint set. Then, recalculate the expected values ​​of the shoulder joint pitch angle and elbow joint pitch angle through the inverse kinematics mapping function to generate the optimized memory-free modeling data update package. S7.5: Write the optimized memory-free modeling data update package into the controller's non-volatile storage unit to replace the original memory-free modeling data.

8. A detachable photovoltaic module cleaning robot, used to execute the cleaning control method according to any one of claims 1-7, characterized in that, include: Multi-axis robotic arm, including upper arm and forearm; The upper arm is driven by a shoulder joint pitch motor, and the proximal end of the forearm is hinged to the distal end of the upper arm and driven by an elbow joint pitch motor; both the shoulder joint pitch motor and the elbow joint pitch motor have corresponding servo drivers and encoders. A cleaning head is installed at the end of the forearm, and a cleaning brush is provided on the bottom surface of the cleaning head; the cleaning head is also provided with a distance sensor and a tilt angle sensor; The controller is electrically connected to the distance sensor and the tilt angle sensor respectively, and is electrically connected to the servo drivers corresponding to the shoulder joint pitch motor and the elbow joint pitch motor through a communication interface.

9. The detachable photovoltaic module cleaning robot according to claim 8, characterized in that, The cleaning head includes a housing, and the cleaning brush is mounted on the side of the housing facing the photovoltaic module. A cleaning drive is provided inside the housing, and the output end of the cleaning drive is connected to the rotation shaft of the cleaning brush through a transmission component to drive the cleaning brush to rotate around its axis.

10. An intelligent transport device, characterized in that, The invention includes a detachable photovoltaic module cleaning robot as described in claim 8, and a mobile lift, wherein the lift has a positioning and heading angle sensing module and is electrically connected to the controller; the mobile lift has a lifting platform, and a turntable structure is provided on the lifting platform; the base end of the multi-axis robotic arm is mounted on the turntable structure and rotates relative to the lifting platform through the turntable structure.

Images