An autonomous planning path system and method based on a photovoltaic cleaning robot

Abstract

The application relates to the technical field of photovoltaic device cleaning, and discloses an autonomous path planning system and method based on a photovoltaic cleaning robot. The system comprises an environment perception module, which collects three-dimensional point cloud data of a photovoltaic panel surface and reduces noise; a path planning module, which constructs a topological map and optimizes a global path; an obstacle detection module, which identifies obstacles and extracts their physical properties; a dynamic adjustment module, which analyzes path conflicts and dynamically replans; and an execution control module, which generates robot motion instructions. The method realizes efficient path planning and cleaning operation of the photovoltaic cleaning robot in a complex environment through cooperative work of the above modules. Compared with a traditional cleaning method, the application improves cleaning efficiency and quality, reduces cost, enhances the adaptability of the equipment to a complex environment, and provides strong technical support for the development of the photovoltaic industry.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic equipment cleaning technology, specifically to an autonomous path planning system and method based on a photovoltaic cleaning robot. Background Technology

[0002] Against the backdrop of the global energy transition, photovoltaic (PV) power generation, as a sustainable and clean energy source, is rapidly expanding in scale. As the core component of a PV power generation system, the cleanliness of the PV panel's surface directly affects its power generation efficiency. However, the cleaning of PV panels currently faces numerous challenges.

[0003] Traditional manual cleaning methods have many drawbacks. Firstly, manual cleaning is costly. As the scale of photovoltaic power plants continues to increase, the number of workers required also increases accordingly. This involves not only labor costs but also personnel training, safety, and a series of other expenses. For example, a large photovoltaic power plant may require a cleaning team of dozens or even hundreds of people, with annual labor costs reaching several million yuan. Secondly, manual cleaning is inefficient. Manual cleaning mainly relies on handheld cleaning tools for wiping, which is slow and difficult to meet the cleaning needs of large-scale photovoltaic power plants. Moreover, manual cleaning is greatly affected by weather and terrain factors. In severe weather conditions, such as high temperatures, heavy rain, and strong winds, cleaning operations cannot be carried out, leading to delays in the cleaning work.

[0004] While mechanical cleaning equipment has improved cleaning efficiency to some extent, it also has significant shortcomings. Most existing mechanical cleaning equipment lacks adaptability to complex environments. Photovoltaic power stations often have diverse and complex terrains, including mountains, hills, and deserts, making it difficult for some mechanical cleaning equipment to move and operate smoothly on these surfaces. In mountainous photovoltaic power stations, equipment may slip or tip over due to steep slopes. Furthermore, these devices have limited ability to identify and avoid obstacles. The area around photovoltaic panels may contain various obstacles such as weeds, rocks, and bird droppings. If mechanical cleaning equipment cannot detect and avoid these obstacles in time, it will not only affect the cleaning effect but may also damage the equipment itself, increasing maintenance costs.

[0005] In terms of path planning, existing cleaning equipment uses relatively simple methods. Some equipment uses fixed path planning, which cannot be adjusted according to the actual situation of the photovoltaic panels, resulting in incomplete cleaning and blind spots. Even some equipment with path planning capabilities often fails to effectively integrate multi-source data, resulting in insufficient environmental perception. For example, relying solely on information from a single sensor makes it difficult to accurately determine the surface condition of the photovoltaic panels, the specific location and properties of obstacles, etc., thus failing to plan the optimal cleaning path and reducing cleaning efficiency and energy utilization.

[0006] Furthermore, current cleaning equipment has a low level of intelligence, lacking the ability to make autonomous decisions and adapt. During the cleaning process, in the event of unforeseen circumstances such as sensor malfunctions or equipment component damage, the equipment often cannot handle the situation independently, requiring manual intervention. This not only affects the continuity of cleaning work but also increases maintenance costs. With the continuous development of the photovoltaic industry, higher demands are being placed on the intelligence, efficiency, and adaptability of photovoltaic cleaning technologies. Therefore, developing an autonomous path planning system and method based on photovoltaic cleaning robots that can solve the above problems has significant practical implications. Summary of the Invention

[0007] The purpose of this invention is to provide an autonomous path planning system and method based on a photovoltaic cleaning robot to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an autonomous path planning system based on a photovoltaic cleaning robot, the system comprising:

[0009] The environmental perception module is used to collect and process three-dimensional point cloud data of the photovoltaic panel surface using multi-source heterogeneous sensors to obtain photovoltaic surface point cloud data; and to perform noise removal processing on the photovoltaic surface point cloud data based on an adaptive filtering algorithm to obtain photovoltaic surface noise-reduced data.

[0010] The path planning module is used to perform dynamic map construction processing on the photovoltaic surface noise reduction data using an improved SLAM algorithm to obtain a photovoltaic surface topology map; and to perform global path optimization processing on the photovoltaic surface topology map based on a hierarchical reinforcement learning algorithm to obtain the initial optimal path.

[0011] The obstacle detection module is used to perform obstacle recognition processing on real-time acquired photovoltaic panel images using a spatiotemporal fusion convolutional neural network to obtain obstacle distribution information; and combines infrared thermal imaging technology to extract features of the physical properties of obstacles to obtain obstacle attribute data.

[0012] The dynamic adjustment module is used to perform conflict analysis on obstacle attribute data and the initial optimal path through a fuzzy logic controller to obtain the path conflict coefficient; and to perform dynamic path replanning on the path conflict coefficient based on a quantum genetic algorithm to obtain the adjusted global path.

[0013] The execution control module is used to perform joint trajectory planning on the adjusted global path using the inverse kinematics algorithm to obtain the robot's motion trajectory command; and to perform drive signal conversion processing on the robot's motion trajectory command through the multi-axis cooperative control program to obtain the photovoltaic cleaning robot's execution command.

[0014] Preferably, the environment sensing module includes:

[0015] The point cloud acquisition unit is used to acquire three-dimensional point clouds on the surface of photovoltaic panels using the fusion technology of lidar and depth camera.

[0016] An adaptive filtering unit is used to perform noise separation processing on point cloud data based on a hybrid algorithm of wavelet transform and Kalman filtering.

[0017] The data fusion unit is used to spatiotemporally align the noise-reduced point cloud data with the photovoltaic panel tilt angle sensor data to generate photovoltaic surface noise-reduced data.

[0018] Preferably, the function formula of the improved SLAM algorithm in the path planning module is as follows:

[0019]

[0020] In the formula, M t Let z be the dynamic map constructed at time t, M be the set of candidate maps, and z be the dynamic map constructed at time t. t For sensor observation data, x t For the robot's pose, α k Let m be the confidence weight of the k-th submap. k Let β be the spatial feature description of the k-th sub-map, and β be the map smoothing constraint coefficient. Let K be the gradient of the submap, and K be the total number of submaps.

[0021] Preferably, the structure of the spatiotemporal fusion convolutional neural network in the obstacle detection module includes:

[0022] A spatial feature extraction layer is used to extract the spatial dimension features of photovoltaic images through 3D convolutional kernels;

[0023] The time series association layer is used to model the temporal changes of consecutive frame images using a long short-term memory network;

[0024] The attribute classification layer is used to combine infrared thermal imaging data with image features for multimodal fusion classification and output obstacle attribute data.

[0025] Preferably, the optimization objective function of the quantum genetic algorithm in the dynamic adjustment module is:

[0026]

[0027] In the formula, P is the set of paths, and d i Let E be the length of the i-th path segment. i Let w1 and w2 be the path energy consumption, w1 and w2 be the weight coefficients, the optimization objective be to maximize F(P), and N be the total number of path segments.

[0028] Preferably, the multi-axis collaborative control program in the execution control module includes:

[0029] The trajectory interpolation unit is used to smoothly interpolate joint trajectories based on fifth-order polynomial curves.

[0030] The drive conversion unit is used to convert trajectory commands into pulse signals for each joint motor using an inverse kinematics model;

[0031] The feedback correction unit is used to compensate for the phase error of the pulse signal in real time by combining encoder feedback data.

[0032] Preferably, the improved SLAM algorithm further includes:

[0033] The dynamic object culling unit is used to detect moving obstacles using optical flow and remove their corresponding point clouds from the map.

[0034] The map update unit is used to perform incremental map updates based on static environmental features using a Bayesian probability model.

[0035] Preferably, the spatiotemporal fusion convolutional neural network further includes:

[0036] The attention mechanism unit is used to weightedly fuse salient regions of infrared thermal imaging data through the channel attention module;

[0037] The anti-occlusion unit is used to complete and reconstruct partially occluded obstacle images using a generative adversarial network.

[0038] Preferably, the quantum genetic algorithm employs the following encoding method:

[0039] Quantum chromosome encoding encodes the coordinates and turning angles of path nodes into a superposition state of qubits;

[0040] Adaptive rotating gate update dynamically adjusts the rotation angle of the qubits based on the path conflict coefficient to accelerate convergence.

[0041] Preferably, the present invention also includes an autonomous path planning method based on a photovoltaic cleaning robot, the method comprising the following steps:

[0042] S1: Use multi-source heterogeneous sensors to collect and process three-dimensional point cloud data of photovoltaic panel surface to obtain photovoltaic surface point cloud data. Then, use an adaptive filtering algorithm to remove noise from the photovoltaic surface point cloud data to obtain photovoltaic surface noise reduction data.

[0043] S2: The photovoltaic surface noise reduction data is dynamically mapped using an improved SLAM algorithm to obtain a photovoltaic surface topology map. Then, the photovoltaic surface topology map is globally optimized using a hierarchical reinforcement learning algorithm to obtain the initial optimal path.

[0044] S3: Use a spatiotemporal fusion convolutional neural network to perform obstacle recognition processing on real-time acquired photovoltaic panel images to obtain obstacle distribution information. Combine this with infrared thermal imaging technology to extract features of the physical properties of the obstacles and obtain obstacle attribute data.

[0045] S4: The obstacle attribute data and the initial optimal path are analyzed and processed by the fuzzy logic controller to obtain the path conflict coefficient. The path conflict coefficient is then dynamically reprogrammed based on the quantum genetic algorithm to obtain the adjusted global path.

[0046] S5: The inverse kinematics algorithm is used to perform joint trajectory planning on the adjusted global path to obtain the robot's motion trajectory command. The multi-axis cooperative control program is then used to convert the robot's motion trajectory command into a drive signal to obtain the execution command of the photovoltaic cleaning robot.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The environmental perception module utilizes multi-source heterogeneous sensors to acquire 3D point clouds and combines them with adaptive filtering algorithms for noise reduction. The fusion technology of LiDAR and depth cameras enables comprehensive acquisition of photovoltaic panel surface information from different angles, accurately acquiring data even under complex lighting conditions. The adaptive filtering algorithm, through a hybrid approach of wavelet transform and Kalman filtering, effectively removes noise, providing high-precision data for subsequent path planning. This makes the robot's perception of the photovoltaic panel surface condition more accurate, significantly improving the reliability of cleaning operations and reducing cleaning omissions or equipment damage caused by misjudgments. The path planning module constructs a dynamic map using an improved SLAM algorithm and optimizes the global path with a hierarchical reinforcement learning algorithm. The improved SLAM algorithm's formula comprehensively considers multiple factors to construct the map; dynamic object removal units and map update units ensure real-time map accuracy, enabling the robot to quickly locate its own position and surrounding environment. The hierarchical reinforcement learning algorithm plans paths at both macroscopic and microscopic levels, finding the initial optimal path. Compared to traditional fixed paths or simple planning methods, this significantly improves cleaning efficiency, reduces cleaning time, and enhances the overall power generation efficiency of the photovoltaic power station.

[0049] The obstacle detection module utilizes a spatiotemporal fusion convolutional neural network and infrared thermal imaging technology to not only identify obstacle distribution but also extract physical attributes. A spatial feature extraction layer, a time-series correlation layer, and an attribute classification layer work together to accurately identify obstacles. An attention mechanism unit and an anti-occlusion unit handle occlusion issues, improving detection accuracy. Combining infrared thermal imaging technology to distinguish obstacles of different materials provides detailed information for path planning, enabling the robot to avoid obstacles in advance, ensuring equipment safety, and reducing maintenance costs. The dynamic adjustment module analyzes path conflicts through a fuzzy logic controller and replans the path using a quantum genetic algorithm. The fuzzy logic controller derives the conflict coefficient based on obstacle attributes and path relationships. The quantum genetic algorithm, guided by an optimization objective function, employs quantum chromosome encoding and adaptive rotating door updates to rapidly adjust the path in complex environments, ensuring the robot successfully avoids obstacles and continues to complete the cleaning task, improving the robot's adaptability and operational capabilities in complex environments.

[0050] The execution control module transforms the planned path into precise motion commands for the robot through inverse kinematics algorithms and multi-axis collaborative control programs. The trajectory interpolation unit uses fifth-order polynomial curve smoothing interpolation to ensure smooth robot movement. The drive conversion unit accurately converts signals to drive the motors, and the feedback correction unit combines encoder data to compensate for errors in real time, ensuring the robot moves precisely along the planned path, guaranteeing uniform cleaning quality, and preventing uneven cleaning from affecting the photovoltaic panel's power generation efficiency.

[0051] This invention automates and intelligently cleans photovoltaic panels, reducing reliance on manual labor and lowering labor costs. It improves cleaning efficiency and equipment safety, minimizing equipment damage and maintenance costs. Precise cleaning and path planning enhance the power generation efficiency of photovoltaic panels, increasing electricity output and creating greater economic benefits for the photovoltaic industry, thus promoting its sustainable development. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the autonomous path planning system based on a photovoltaic cleaning robot described in this invention.

[0053] Figure 2 Flowchart for map updating in the improved SLAM algorithm;

[0054] Figure 3 A flowchart for obstacle handling using a spatiotemporal fusion convolutional neural network;

[0055] Figure 4 A flowchart for encoding and updating the quantum genetic algorithm. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1-4 This invention relates to an autonomous path planning system based on a photovoltaic cleaning robot, and its specific implementation scheme is described in detail below.

[0058] This system mainly includes an environmental perception module, a path planning module, an obstacle detection module, a dynamic adjustment module, and an execution control module.

[0059] The environmental perception module is responsible for data acquisition and processing of the photovoltaic panel surface. Utilizing multi-source heterogeneous sensors, it acquires 3D point cloud data of the photovoltaic panel surface through a fusion technology of LiDAR and depth cameras. Then, based on an adaptive filtering algorithm—a hybrid algorithm employing wavelet transform and Kalman filtering—noise is removed from the acquired photovoltaic surface point cloud data to eliminate interference. Finally, the denoised point cloud data is spatiotemporally aligned with the photovoltaic panel tilt angle sensor data to obtain the final denoised photovoltaic surface data.

[0060] The path planning module operates based on the photovoltaic surface noise reduction data obtained from the environmental perception module. An improved SLAM algorithm is used to dynamically construct a map from this data, resulting in a photovoltaic surface topology map. Then, a hierarchical reinforcement learning algorithm is used to perform global path optimization on the generated photovoltaic surface topology map, generating an initial optimal path and planning a preliminary cleaning route for the robot.

[0061] The obstacle detection module utilizes a spatiotemporal fusion convolutional neural network to process real-time acquired images of photovoltaic panels. This network identifies obstacles and obtains obstacle distribution information through spatial feature extraction, temporal series correlation, and attribute classification layers. Simultaneously, it combines infrared thermal imaging technology to extract features of the obstacles' physical properties, acquiring obstacle attribute data to subsequently assess their impact on path planning.

[0062] The dynamic adjustment module receives obstacle attribute data from the obstacle detection module and the initial optimal path generated by the path planning module. A fuzzy logic controller performs conflict analysis on both data to obtain a path conflict coefficient. Based on a quantum genetic algorithm, and with maximizing the objective function as the optimization direction, dynamic path replanning is performed on the path conflict coefficient to obtain an adjusted global path, ensuring that the robot can adjust its path promptly when encountering obstacles.

[0063] The execution control module uses an inverse kinematics algorithm to perform joint trajectory planning on the adjusted global path, obtaining the robot's motion trajectory command. Through a multi-axis cooperative control program, which includes a trajectory interpolation unit, a drive conversion unit, and a feedback correction unit, the robot's motion trajectory command is converted into drive signals, ultimately yielding the photovoltaic cleaning robot's execution command, controlling the robot to perform cleaning operations according to the planned path.

[0064] The specific embodiments of the present invention are described in detail below.

[0065] Example 1:

[0066] This embodiment describes the environmental perception module. In practical applications, photovoltaic panels are typically located in complex outdoor environments, potentially subject to interference factors such as dust and rain. The fusion technology of LiDAR and depth camera plays a crucial role. LiDAR acquires distance information by emitting a laser beam and measuring the time it takes for the reflected light to pass through, while the depth camera directly captures the depth data of objects. Combined, these technologies can obtain three-dimensional point cloud data of the photovoltaic panel surface from different angles, comprehensively and accurately reflecting the shape and characteristics of the panel surface.

[0067] An adaptive filtering unit based on a hybrid algorithm of wavelet transform and Kalman filtering is employed. Wavelet transform enables multi-scale decomposition of the signal, separating noise from the useful signal in point cloud data. Its multi-resolution analysis characteristics can effectively capture noise in different frequency bands. Kalman filtering, on the other hand, recursively estimates the optimal state based on the previous time-to-time estimate and the current time-to-time observation, continuously correcting the point cloud data and removing noise interference in the presence of noise.

[0068] The data fusion unit performs spatiotemporal alignment between the denoised point cloud data and the photovoltaic panel tilt sensor data. Assuming the tilt angle measured by the photovoltaic panel tilt sensor is θ, at time t, the timestamp of the denoised point cloud data is matched with the timestamp of the tilt sensor data. Based on the tilt angle θ, the spatial position of the point cloud data at different locations is adjusted so that the point cloud data more accurately reflects the actual state of the photovoltaic panel, generating denoised photovoltaic surface data.

[0069] Example 2: The path planning module is one of the core components of this system. The improved SLAM algorithm uses the following function formula when constructing a dynamic map:

[0070]

[0071] Among them, M tThe dynamic map built at time t is based on current sensor observation data and robot pose information. M is the candidate map set, containing possible map building schemes. The system needs to select the optimal scheme from these candidates to build the dynamic map for the current time. t This refers to sensor observation data, including data collected by sensors such as LiDAR and depth cameras. This data forms the basis for map construction. t This represents the robot's pose, i.e., its position and orientation information at time t. Knowing the robot's position on the map is essential for accurate map construction. α k represents the confidence weight of the k-th submap. Different submaps have different levels of importance in the overall map construction, and the confidence weight is used to measure this difference. k is the spatial feature description of the k-th sub-map, used to characterize the shape, location, and other spatial features of the sub-map. β is the map smoothing constraint coefficient, used to control the smoothness of the map and avoid excessive fluctuations or discontinuities. The gradient of the submap reflects the changes in its spatial features. K represents the total number of submaps. Dividing the entire map into multiple submaps for processing helps improve the efficiency and accuracy of map construction.

[0072] During map construction, the system continuously calculates the formula based on collected sensor data to find the optimal dynamic map construction scheme. Based on the obtained photovoltaic surface topology map, the hierarchical reinforcement learning algorithm further optimizes the global path, generating an initial optimal path by considering factors such as the shape and layout of the photovoltaic panels and the requirements of the cleaning task. For example, in a large photovoltaic power plant, where photovoltaic panels are distributed in an array, the hierarchical reinforcement learning algorithm will plan an initial optimal path that can efficiently clean the panels while avoiding frequent robot turning and collisions, based on the panel distribution in different areas.

[0073] Example 3:

[0074] In the obstacle detection module, the spatiotemporal fusion convolutional neural network is a key technology for obstacle recognition. The spatial feature extraction layer extracts the spatial dimension features of the photovoltaic image through a 3D convolutional kernel. The 3D convolutional kernel performs convolution operations on the length, width, and time dimensions of the image. Assuming the size of the 3D convolutional kernel is n×m×p, where n and m represent the kernel size on the image plane, and p represents the kernel size in the time dimension. Through this convolution operation, features such as the spatial position and shape of objects in the image can be extracted, such as the outline and size of obstacles.

[0075] The time-series correlation layer utilizes a Long Short-Term Memory (LSTM) network to model the temporal changes of consecutive frame images. LSTM can effectively handle long-term dependencies in time-series data. In photovoltaic panel images, by analyzing multiple consecutive frames, it can capture the motion trajectory and changing trends of obstacles. For example, when a bird or other moving object passes by the photovoltaic panel, LSTM can identify its motion state and determine whether it is an obstacle.

[0076] The attribute classification layer combines infrared thermal imaging data with image features for multimodal fusion classification, outputting obstacle attribute data. Infrared thermal imaging technology can detect temperature differences in objects; different objects have different temperature characteristics. By fusing infrared thermal imaging data with image features, classification algorithms can determine the physical properties of obstacles, such as material and electrical charge. Let the infrared thermal imaging data be represented as I... ir The image features are represented as F img The classification algorithm can be represented as C(I ir ,F img This algorithm obtains obstacle attribute data, providing more detailed information for subsequent path planning.

[0077] Example 4:

[0078] The dynamic adjustment module plays a crucial role in the system, responding to unexpected situations and ensuring the robot's smooth operation. The optimization objective function of the quantum genetic algorithm is:

[0079]

[0080] In this formula, P represents the path set, which contains all possible path options the robot can choose during the cleaning process. i E represents the length of the i-th path segment. Path length is an important indicator of path efficiency; a shorter path can reduce the robot's running time and energy consumption. i The path energy consumption varies depending on factors such as terrain and robot movement. w1 and w2 are weighting coefficients used to adjust the importance of path length and path energy consumption in the optimization objective. These weighting coefficients can be flexibly set according to the actual application scenario and requirements. N is the total number of path segments. In actual path planning, dividing the entire path into multiple path segments for analysis and optimization helps to plan the path more accurately.

[0081] The fuzzy logic controller performs conflict analysis based on obstacle attribute data and the initial optimal path to obtain the path conflict coefficient. Assume the path conflict coefficient is C. conflictThe path conflict coefficient reflects the degree of interference of obstacles with the initial optimal path. The quantum genetic algorithm dynamically adjusts the path based on the path conflict coefficient, continuously iterating and optimizing the objective function F(P) to find the path scheme that maximizes F(P), i.e., the adjusted global path. In practical applications, when encountering large obstacles, the path conflict coefficient increases, and the quantum genetic algorithm will be more inclined to choose paths with lower energy consumption that bypass the obstacles, ensuring that the robot can complete the cleaning task safely and efficiently.

[0082] Example 5:

[0083] The execution control module is responsible for translating the planned path into the robot's actual motion. The trajectory interpolation unit performs smooth interpolation of the joint trajectories based on a quintic polynomial curve. The expression for the quintic polynomial curve is q(t) = a0 + a1t + a2t. 2 +a3t 3 +a4t 4 +a5t 5 Where q(t) represents the joint position at time t, and a0 to a5 are polynomial coefficients. By giving boundary conditions such as the starting and ending positions, velocities, and accelerations, these coefficients can be determined, thereby achieving a smooth transition of the joint trajectory and avoiding jitter and shocks during robot movement.

[0084] The drive conversion unit converts trajectory commands into pulse signals for each joint motor using an inverse kinematics model. Assuming the robot's kinematic model is f(q), where q is the joint angle vector, the inverse kinematics model is f(q). -1 (x), where x represents the position and orientation of the robot's end effector. Using the inverse kinematics model, the required rotation angle of each joint can be calculated based on the desired position and orientation of the robot's end effector, and then converted into pulse signals for the motors of each joint to control the rotation of the motors.

[0085] The feedback correction unit combines encoder feedback data to compensate for the phase error of the pulse signal in real time. The encoder can measure the actual rotation angle of the joint in real time; assuming the angle measured by the encoder is θ. enc The desired joint angle is θ des The phase error is Δθ = θ enc -θ des Based on the phase error, the feedback correction unit adjusts the pulse signal to compensate for the motor's rotation, ensuring that the robot's actual movement matches the planned path and improving the accuracy and stability of the robot's movement.

[0086] Example 6:

[0087] In addition to the core formula for map construction, the improved SLAM algorithm also includes a dynamic object culling unit and a map update unit. The dynamic object culling unit detects moving obstacles and removes their corresponding point clouds from the map using optical flow. Optical flow calculates object motion information based on the brightness changes of pixels in an image. Assuming that in two adjacent frames I1(x,y) and I2(x,y), the optical flow vector of pixel (x,y) is (u,v), according to the optical flow constraint equation I... x u+I y v+I t =0, where I x I y These are the gradients of the image in the x and y directions, respectively, I t This represents the rate of change of the image over time. By solving this equation, the motion velocity of each pixel can be obtained, thus detecting moving obstacles. The point cloud corresponding to the detected moving obstacles is then removed from the map to prevent it from interfering with map construction and path planning.

[0088] The map update unit performs incremental map updates based on static environmental features using a Bayesian probability model. The basic formula of the Bayesian probability model is... In map updating, A represents the map's state (such as the characteristics of a certain area on the map), and B represents new sensor observation data. Based on the new observation data, the probability distribution of the map state is updated using Bayes' theorem, achieving incremental map updates. For example, when the robot moves to a new area and acquires new sensor data, the map updating unit updates the map based on the Bayesian probability model and previous map information, making the map more accurately reflect environmental changes.

[0089] Example 7:

[0090] In addition to its basic structure, the spatiotemporal fusion convolutional neural network includes an attention mechanism unit and an anti-occlusion unit. The attention mechanism unit uses a channel attention module to weightedly fuse salient regions of the infrared thermal imaging data. Assuming the number of channels in the infrared thermal imaging data is C, the channel attention module first performs global average pooling and global max pooling on the data from each channel, obtaining two feature vectors F. avg and F max Then, the two feature vectors are processed by a multilayer perceptron (MLP) to obtain the channel attention weight vector W. Finally, the infrared thermal imaging data and the weight vector W are weighted and fused to highlight the features of the salient region and enhance the ability to identify obstacles.

[0091] The anti-occlusion unit utilizes a Generative Adversarial Network (GAN) to complete and reconstruct partially occluded obstacle images. The GAN consists of a generator G and a discriminator D. The generator aims to generate a completed image similar to the real image, while the discriminator determines whether the generated image is realistic. During training, the generator and discriminator continuously optimize against each other. It is assumed that the generator's input is a partially occluded image I. occluded The generated completed image is I generated The output of the discriminator is D(I) generated The generator minimizes 1-D(I) generated To optimize itself, the discriminator maximizes D(I) by maximizing D(I). generated This allows the anti-occlusion unit to optimize itself. In this way, the anti-occlusion unit can complete and reconstruct partially occluded obstacle images, improving the accuracy of obstacle detection.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An autonomous path planning system based on a photovoltaic cleaning robot, characterized in that, include: The environmental perception module is used to collect and process three-dimensional point cloud data of the photovoltaic panel surface using multi-source heterogeneous sensors to obtain photovoltaic surface point cloud data. The photovoltaic surface point cloud data is then processed to remove noise based on an adaptive filtering algorithm, resulting in denoised photovoltaic surface data. The path planning module is used to perform dynamic map construction processing on the photovoltaic surface noise reduction data using an improved SLAM algorithm to obtain a photovoltaic surface topology map; and to perform global path optimization processing on the photovoltaic surface topology map based on a hierarchical reinforcement learning algorithm to obtain the initial optimal path. The obstacle detection module is used to perform obstacle recognition processing on real-time acquired photovoltaic panel images using a spatiotemporal fusion convolutional neural network to obtain obstacle distribution information; and combines infrared thermal imaging technology to extract features of the physical properties of obstacles to obtain obstacle attribute data. The dynamic adjustment module is used to perform conflict analysis on obstacle attribute data and the initial optimal path through a fuzzy logic controller to obtain the path conflict coefficient. The path conflict coefficient is dynamically replanned based on the quantum genetic algorithm to obtain the adjusted global path. The execution control module is used to perform joint trajectory planning on the adjusted global path using the inverse kinematics algorithm to obtain the robot's motion trajectory command; and to perform drive signal conversion processing on the robot's motion trajectory command through the multi-axis cooperative control program to obtain the photovoltaic cleaning robot's execution command.

2. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 1, characterized in that, The environment sensing module includes: The point cloud acquisition unit is used to acquire three-dimensional point clouds on the surface of photovoltaic panels using the fusion technology of lidar and depth camera. An adaptive filtering unit is used to perform noise separation processing on point cloud data based on a hybrid algorithm of wavelet transform and Kalman filtering. The data fusion unit is used to spatiotemporally align the noise-reduced point cloud data with the photovoltaic panel tilt angle sensor data to generate photovoltaic surface noise-reduced data.

3. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 1, characterized in that, The function formula for the improved SLAM algorithm in the path planning module is as follows: In the formula, M t Let z be the dynamic map constructed at time t, M be the set of candidate maps, and z be the dynamic map constructed at time t. t For sensor observation data, x t For the robot's pose, α k Let m be the confidence weight of the k-th submap. k Let β be the spatial feature description of the k-th sub-map, and β be the map smoothing constraint coefficient. Let K be the gradient of the submap, and K be the total number of submaps.

4. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 1, characterized in that, The structure of the spatiotemporal fusion convolutional neural network in the obstacle detection module includes: A spatial feature extraction layer is used to extract the spatial dimension features of photovoltaic images through 3D convolutional kernels; The time series association layer is used to model the temporal changes of consecutive frame images using a long short-term memory network; The attribute classification layer is used to combine infrared thermal imaging data with image features for multimodal fusion classification and output obstacle attribute data.

5. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 1, characterized in that, The optimization objective function of the quantum genetic algorithm in the dynamic adjustment module is: In the formula, P is the set of paths, and d i Let E be the length of the i-th path segment. i Let w1 and w2 be the path energy consumption, w1 and w2 be the weight coefficients, the optimization objective be to maximize F(P), and N be the total number of path segments.

6. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 1, characterized in that, The multi-axis collaborative control program in the execution control module includes: The trajectory interpolation unit is used to smoothly interpolate joint trajectories based on fifth-order polynomial curves. The drive conversion unit is used to convert trajectory commands into pulse signals for each joint motor using an inverse kinematics model; The feedback correction unit is used to compensate for the phase error of the pulse signal in real time by combining encoder feedback data.

7. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 3, characterized in that, The improved SLAM algorithm also includes: The dynamic object culling unit is used to detect moving obstacles using optical flow and remove their corresponding point clouds from the map. The map update unit is used to perform incremental map updates based on static environmental features using a Bayesian probability model.

8. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 4, characterized in that, The spatiotemporal fusion convolutional neural network also includes: The attention mechanism unit is used to weightedly fuse salient regions of infrared thermal imaging data through the channel attention module; The anti-occlusion unit is used to complete and reconstruct partially occluded obstacle images using a generative adversarial network.

9. The autonomous path planning system based on a photovoltaic cleaning robot according to claim 5, characterized in that, The quantum genetic algorithm uses the following encoding method: Quantum chromosome encoding encodes the coordinates and turning angles of path nodes into a superposition state of qubits; Adaptive rotating gate update dynamically adjusts the rotation angle of the qubits based on the path conflict coefficient to accelerate convergence.

10. A method for autonomous path planning based on a photovoltaic cleaning robot, characterized in that, Includes the following steps: S1: Use multi-source heterogeneous sensors to collect and process three-dimensional point cloud data of photovoltaic panel surface to obtain photovoltaic surface point cloud data. Then, use an adaptive filtering algorithm to remove noise from the photovoltaic surface point cloud data to obtain photovoltaic surface noise reduction data. S2: The photovoltaic surface noise reduction data is dynamically mapped using an improved SLAM algorithm to obtain a photovoltaic surface topology map. Then, the photovoltaic surface topology map is globally optimized using a hierarchical reinforcement learning algorithm to obtain the initial optimal path. S3: Use a spatiotemporal fusion convolutional neural network to perform obstacle recognition processing on real-time acquired photovoltaic panel images to obtain obstacle distribution information. Combine this with infrared thermal imaging technology to extract features of the physical properties of the obstacles and obtain obstacle attribute data. S4: The obstacle attribute data and the initial optimal path are analyzed and processed by the fuzzy logic controller to obtain the path conflict coefficient. The path conflict coefficient is then dynamically reprogrammed based on the quantum genetic algorithm to obtain the adjusted global path. S5: The inverse kinematics algorithm is used to perform joint trajectory planning on the adjusted global path to obtain the robot's motion trajectory command. The multi-axis cooperative control program is then used to convert the robot's motion trajectory command into a drive signal to obtain the execution command of the photovoltaic cleaning robot.

Images