Photovoltaic cleaning robot based on automatic path planning and operation method thereof

Abstract

The invention discloses a photovoltaic cleaning robot based on automatic path planning and an operation method thereof, relates to the technical field of robots, and solves the problem of cleaning blind spots in an existing cleaning robot and the endless loop risk of a path planning algorithm. According to the invention, the surface information of the photovoltaic panel is accurately captured through the multispectral acquisition module, a high-resolution and high-definition comprehensive image is generated through the data fusion processing module, then dirt is accurately identified on different scales by using a multi-scale pyramid structure and an adaptive feature fusion mechanism, and an attention mechanism and environmental perception data are combined, so that the real-time performance of the system is improved. The recognition threshold is dynamically adjusted, and misrecognition and missing detection are reduced; and finally, enabling the robot to autonomously learn in a complex environment and optimize an obstacle avoidance strategy through a dynamic obstacle avoidance algorithm based on deep reinforcement learning. The intelligent level and the cleaning effect of an existing photovoltaic cleaning robot are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of robot technology, and more specifically to a photovoltaic cleaning robot based on automatic path planning and an operation method thereof. Background Art

[0002] With the continuous growth of global energy demand and the enhancement of environmental protection awareness, solar energy has been widely used as a clean and renewable energy source. As the main form of solar power generation, the power generation efficiency of photovoltaic power stations directly affects the economy and environmental protection of energy utilization. However, dust and dirt on the surface of photovoltaic panels will significantly reduce the photoelectric conversion efficiency. Therefore, regular cleaning of photovoltaic panels has become a key measure to improve the efficiency of photovoltaic power generation. Traditional manual cleaning methods are not only labor-intensive and costly, but also difficult to achieve efficient and uniform cleaning in large-area photovoltaic power stations. Photovoltaic cleaning robots based on automatic path planning have emerged. They can automatically plan cleaning paths without human intervention, clean photovoltaic panels efficiently and evenly, and significantly improve the operation and maintenance efficiency and economic benefits of photovoltaic power stations.

[0003] Photovoltaic cleaning robots based on automatic path planning are an efficient cleaning tool developed in recent years. The core of the robot is to realize automatic cleaning of photovoltaic panels through advanced machine vision, path planning algorithms and walking systems. These robots are usually equipped with high-precision sensors and cameras, which can identify dirt and obstacles on photovoltaic panels in real time, and generate the optimal cleaning path through path planning algorithms. For example, the existing technology is a photovoltaic cleaning robot based on machine vision and path planning. The robot collects images of photovoltaic panels in real time through cameras, uses image processing technology to identify dirt areas, and generates the optimal cleaning path through the A* algorithm. In addition, there are path planning methods based on laser radar and SLAM (simultaneous localization and mapping) technology, which further improve the navigation accuracy and autonomy of the robot. The development of these technologies enables photovoltaic cleaning robots to complete cleaning tasks efficiently and accurately in complex environments.

[0004] However, although photovoltaic cleaning robots based on automatic path planning have made significant progress in improving cleaning efficiency and reducing labor costs, there are still some potential shortcomings and challenges. First, under certain lighting conditions, such as strong reflections or interlaced shadows, the machine vision system may not be able to accurately identify or distinguish dirt from normal light and shadow changes, resulting in misjudgment and missed scans. The robot may mistakenly believe that the photovoltaic panel is clean and skip the area that needs to be cleaned, thereby affecting the power generation efficiency. To overcome the blind spot problem, the robot needs to perform additional scanning or confirmation steps, which will increase energy consumption and cleaning time. Second, in complex or dynamically changing environments, such as irregular layout of photovoltaic panels or temporary obstacles, the path planning algorithm may fall into a "dead loop", that is, repeatedly trying to access the same path or failing to find a valid path. This will cause a significant drop in cleaning efficiency, and the robot will stagnate for a long time and fail to effectively complete the cleaning task. At the same time, continuous attempts to find a path will increase the robot's energy consumption and shorten its working time.

[0005] Therefore, the present invention discloses a photovoltaic cleaning robot based on automatic path planning and an operating method thereof to solve the blind spot problem of machine vision and the dead loop risk of the path planning algorithm existing in the existing photovoltaic cleaning robots. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention discloses a photovoltaic cleaning robot based on automatic path planning and an operating method thereof, which improves the cleaning ability and path planning ability of the photovoltaic cleaning robot.

[0007] The present invention adopts the following technical solutions:

[0008] A photovoltaic cleaning robot based on automatic path planning, comprising a multi-spectral acquisition module, a data fusion processing module, a dirt dynamic recognition module, an environment perception module and a path planning module;

[0009] The multi-spectral acquisition module is used to acquire the surface image of the photovoltaic panel under the visible light, near infrared and ultraviolet spectrum, perform geometric correction and radiation correction, and store the corrected image data;

[0010] The data fusion processing module is used to read multispectral image data, perform image registration, multi-scale fusion and preprocessing, and generate a comprehensive image that integrates image information under different spectra;

[0011] The dirt dynamic identification module is used to extract multi-scale features based on the adaptive deep dirt analysis framework, identify abnormal areas including dirt, scratches and abnormal light and shadow changes, dynamically adjust the dirt identification threshold, and classify the identified abnormal areas to generate classification labels and dirt distribution maps; the adaptive deep dirt analysis framework automatically learns the feature representation and classification boundaries of dirt through the attention enhanced convolutional neural network algorithm based on the comprehensive image output by the data fusion processing module; the adaptive deep dirt analysis framework includes an input layer, a multi-scale feature extraction layer, a feature aggregation layer, a threshold adjustment layer, a multi-category classification layer and a dirt distribution map generation layer;

[0012] The environmental perception module is used to collect environmental data in the photovoltaic power station in real time, perform preprocessing, build a dynamic environmental matrix model, and store the preprocessed environmental data, wherein the environmental data includes light intensity, temperature, humidity, shadow movement, and obstacle position;

[0013] The path planning module is used to read the dynamic environment matrix model and the dirt distribution map, generate an initial cleaning path through an adaptive illumination compensation path planning algorithm, and monitor the position of obstacles in real time during the cleaning process, and dynamically adjust the cleaning path through a dynamic obstacle avoidance algorithm based on deep reinforcement learning.

[0014] As a further technical solution of the present invention, the multi-spectral acquisition module first divides the incident light into visible light, near-infrared and ultraviolet spectral bands through spectral separation, and each spectral band ensures the purity and accuracy of the spectrum through an interference filter; then, the multi-spectral acquisition module converts the separated spectral information into digital image data through a charge coupled device, and performs geometric correction and radiation correction on the collected image through an image correction unit; the image correction unit eliminates image distortion through affine transformation, and eliminates the radiation intensity difference of spectral data at different wavelengths through a radiation correction model; the radiation correction model corrects the radiation value of each pixel through polynomial fitting.

[0015] As a further technical solution of the present invention, the data fusion processing module aligns images under different spectra through an image registration algorithm; the image registration algorithm extracts feature points in the image through a feature point matching-based method SIFT, and matches the transformation matrix between images under different spectra through a random sample consistency algorithm; the random sample consistency algorithm selects a subset of feature points randomly for multiple times, fits the transformation model and verifies the model consistency to select the best transformation matrix to ensure that the pixels of different spectral images correspond consistently; then, the data fusion processing module fuses the high-frequency and low-frequency information of different spectral images through a multi-scale pyramid decomposition method, and performs denoising, contrast enhancement and edge detection operations on the fused image through an image preprocessing method; the working steps of the multi-scale pyramid decomposition method are as follows:

[0016] S301, decomposing each spectral image into image layers of multiple scales through a Gaussian pyramid;

[0017] S302, performing weighted averaging on the same scale layer of different spectral images to generate a fused image layer;

[0018] S303, recombining the fused image layers into a comprehensive image through inverse pyramid transformation;

[0019] Based on the fused comprehensive image, the image preprocessing method dynamically adjusts the filtering parameters according to the local characteristics of the image through the non-local mean, and enhances the contrast through the histogram equalization and Canny edge detection algorithm, thereby improving the visual effect of the image and extracting the edge information in the image.

[0020] As a further technical solution of the present invention, the working method of the adaptive deep fouling analysis framework comprises the following steps:

[0021] Step 401: input the comprehensive image obtained by the data fusion processing module into the adaptive deep dirt analysis framework through the input layer.

[0022] Step 402: extracting multi-scale features from the input through the multi-scale feature extraction layer; the multi-scale feature extraction layer decomposes the comprehensive image into feature maps of multiple scales through a multi-scale pyramid structure, and extracts local and global features using convolution and pooling operations;

[0023] Step 403: The feature aggregation layer aggregates the feature maps generated by the multi-scale extraction layer to form a high-dimensional feature representation; the feature aggregation layer calculates the weight of each feature map through a self-attention mechanism, dynamically adjusts the importance of each feature map, and aggregates feature maps of different scales through feature aggregation;

[0024] Step 404: The light intensity, temperature and humidity data provided by the environment perception module are integrated with the feature map through the threshold adjustment layer, and the threshold of dirt recognition is dynamically adjusted; the threshold adjustment layer integrates the environment data with the feature map through a multimodal fusion method to generate a comprehensive feature representation, and predicts the optimal threshold under the current environmental conditions through a deep learning model;

[0025] Step 405: classify the multi-scale features extracted in step 402 through the multi-class classification layer to identify abnormal areas including dirt, scratches, and abnormal light and shadow changes. The multi-class classification layer generates the probability that each pixel belongs to a different category through a softmax classifier;

[0026] Step 406: Generate a dirt distribution map through the dirt distribution map generation layer according to the classification results of the multi-category classification layer; the dirt distribution map generation layer generates a high-definition dirt distribution map containing dirt location and degree information through morphological operations and connected region analysis.

[0027] As a further technical solution of the present invention, the method for dynamically adjusting the threshold value of dirt recognition by the threshold adjustment layer includes:

[0028] Step 501: Real-time detection of the temperature, humidity and light intensity conditions of the current environment through the temperature, humidity and light intensity data provided by the environment sensing module;

[0029] Step 502: Define the reference threshold T0 and the illumination influence coefficient K I , Temperature influence coefficient K T , Humidity influence coefficient K H , where the reference threshold limit T0 represents the threshold determined under standard light, temperature and humidity conditions, including strong light, normal light and weak light levels; the light influence coefficient K is used to quantify the influence of light intensity on the threshold; the temperature influence coefficient K T Used to quantify the impact of temperature on the threshold; the humidity impact coefficient K H Used to quantify the impact of humidity on the threshold; let the current light intensity be I and the reference light intensity be I0. When the light intensity is higher than the strong light level, the reflectivity of the dirt increases, the image brightness changes greatly, and the dynamic threshold T is lowered to avoid misjudgment; when the light intensity is lower than the weak light level, the reflectivity of the dirt decreases, the image brightness changes little, and the dynamic threshold T is increased to avoid missed detection; the calculation formula of the dynamic threshold T is:

[0030] T=T0+K I (I-I0)+K T (T env -T0)+K H (H env -T0) (1)

[0031] In formula (1), T env Indicates the current ambient temperature; H env Indicates the current ambient humidity;

[0032] Step 503: dynamically adjust the threshold of dirt recognition through the deep learning model, use the light, temperature and humidity data as additional input features, fuse them with the image features, and predict the optimal threshold T under the current environmental conditions through the output layer of the model;

[0033] Step 504: Apply the calculated dynamic threshold T to the dirt identification process to adjust the generation standard of the binary dirt mask. The specific steps include:

[0034] Step r1, processing the comprehensive image through the dirt analysis model to generate a probability map of each pixel belonging to dirt;

[0035] Step r2: Compare each pixel point. If the probability that the pixel point belongs to dirt is greater than the dynamic threshold T, the pixel point is marked as a dirt area; otherwise, it is marked as a non-dirt area.

[0036] Step r3: Generate a binary dirt mask, where the dirt area is marked as 1 and the non-dirt area is marked as 0.

[0037] As a further technical solution of the present invention, the working steps of the attention-enhanced convolutional neural network algorithm include:

[0038] Step 601: The dynamically generated attention map is fused with the multi-scale feature map through the attention enhancement feature fusion formula, so that the model is concentrated in the dirt-dense area; the formula expression of the attention enhancement feature fusion formula is:

[0039]

[0040] In formula (2), AF ij represents the feature map after attention enhancement at position (i, j); σ represents the ReLU nonlinear activation function; AM ij Represents the attention map at position (i, j), which is used to guide the fusion process of feature maps. * represents the element-by-element multiplication operation, which is used to convert the attention weight AM ij Applied to the fusion results of multi-scale feature maps; a k Represents the weight of the kth feature map, which is used to balance the contribution of feature maps of different scales to the final feature representation; represents the feature map extracted by the kth convolutional layer at position (i, j), which is used to capture information in the image at multiple scales; N is the total number of convolutional layers, indicating the number of feature maps involved in feature fusion;

[0041] Step 602: Optimize the sample recognition accuracy near the classification boundary by using a boundary-sensitive loss function; the boundary-sensitive loss function constructs a loss term by using the intersection-over-union ratio between the prediction result and the true label and combining the distance from the prediction result to the nearest classification boundary, thereby amplifying the loss near the boundary area; the formula expression of the boundary-sensitive loss function is:

[0042]

[0043] In formula (3), L dary represents the boundary sensitive loss, which is used to guide the training process of the model; M is the total number of training samples, which indicates the number of samples considered when calculating the loss function; IoU (H (m) ,hm ) represents the prediction result H of the mth sample (m) With the true label h m The intersection-over-union ratio between Indicates the distance between the prediction result of the mth sample and the nearest classification boundary; w s represents a hyperparameter, which is used to control the influence of the distance term on the loss function; exp represents an exponential function, which is used to amplify the loss near the boundary area;

[0044] Step 603: The threshold of dirt recognition is automatically adjusted according to the prediction results of the model and preset conditions during the training process through the dynamic threshold adjustment mechanism. The calculation formula is:

[0045]

[0046] In formula (4), TAU t+1 is the threshold at the next moment, used for dirt identification; TAU t represents the current threshold, which serves as the basis for adjustment; CL(s) is the confidence score of the sth sample, which indicates the confidence of the model in the classification result of the sample; N(l) represents the preset confidence threshold, which is used to judge whether the confidence of the sample is high enough; α represents the learning rate, which is used to control the speed and amplitude of threshold adjustment; sign represents the sign function, which is used to judge the increase or decrease of the threshold; S represents the total size of the threshold adjustment sample set.

[0047] As a further technical solution of the present invention, the environmental perception module divides the physical space of the photovoltaic power station into multiple grid units through a gridding method, and each grid unit contains environmental data in the unit area, and the environmental data includes light intensity, temperature, humidity, shadow movement and obstacle position. The gridding method organizes the environmental data into a structured matrix form through space division and data mapping. Then, the environmental perception module maps the environmental data to the corresponding grid units through interpolation and data filling methods to generate a dynamic environmental matrix model; the interpolation includes nearest neighbor interpolation, bilinear interpolation and spline interpolation; the dynamic environmental matrix model dynamically updates the environmental data through a real-time data stream processing method; the real-time data stream processing method receives and processes the environmental data in real time through a message queue and an event-driven mechanism.

[0048] As a further technical solution of the present invention, the adaptive light compensation path planning algorithm first receives the dynamic environment matrix model constructed in real time by the sensor network, reads the dirt distribution map, analyzes the light intensity data in the environment matrix through the light impact assessment model, and evaluates the impact of light in different areas on the power generation efficiency of the photovoltaic panel; then, the light impact assessment model converts the light intensity value into the corresponding power generation efficiency loss value through the mapping relationship between light intensity and power generation efficiency, and quantifies the impact of uneven light on the performance of the photovoltaic panel; based on the evaluation result of the light impact assessment model, the adaptive light compensation path planning algorithm adopts a priority allocation mechanism to comprehensively calculate the priority of each cleaning area according to the degree of dirt and the degree of light impact; the priority allocation mechanism weightedly sums the dirt density in the dirt distribution map and the power generation efficiency loss value output by the light impact assessment model through a weighted summation method to obtain a comprehensive priority value for each cleaning area; finally, the adaptive light compensation path planning algorithm, based on the calculated priority value, preferentially traverses the areas with high priority while maintaining the overall efficiency of the cleaning path through a heuristic search algorithm to generate an initial cleaning path.

[0049] As a further technical solution of the present invention, the dynamic obstacle avoidance algorithm based on deep reinforcement learning includes a perception layer, a feature extraction layer, a decision layer, a dynamic adjustment layer and an online learning layer; the working method of the dynamic obstacle avoidance algorithm based on deep reinforcement learning includes:

[0050] Step 901: obtaining dynamic information of the working environment of the photovoltaic cleaning robot through the perception layer;

[0051] Step 902: The feature extraction layer performs deep processing on the data acquired by the perception layer; the feature extraction layer extracts high-level abstract features from the original data through multi-layer convolution and pooling operations;

[0052] Step 903: Make a decision through the decision layer. The decision layer uses the deep Q network to continuously iteratively update the parameters of the policy network through the gradient descent method according to the current environment state vector, combined with the past experience in the experience replay pool and the current policy network, so as to maximize the cumulative reward signal. During the decision-making process, the decision layer also balances the conflicts among the cleaning speed, obstacle avoidance success rate and collision penalty items through a multi-objective value function.

[0053] Step 904: When the dynamic adjustment layer detects environmental changes or obstacles, the obstacle avoidance process is triggered; the dynamic adjustment layer predicts the future movement trajectory of the obstacle through LSTM-based sequence prediction, and calculates the optimal obstacle avoidance path in combination with the action suggestions output by the current policy network; at the same time, the dynamic adjustment layer also evaluates environmental uncertainty through an uncertainty quantification method, and adjusts the balance between exploration and utilization based on environmental uncertainty to ensure that robust decisions can still be made in an uncertain environment;

[0054] Step 905: Learn environmental data and obstacle avoidance experience during the cleaning process through the online learning layer, and update the strategy network in real time through the incremental learning method.

[0055] As a further technical solution of the present invention, an operation method of a photovoltaic cleaning robot based on automatic path planning comprises the following steps:

[0056] Step 1: Use the environmental perception module to comprehensively scan the surrounding environment and identify the location, shape, spacing, light intensity, temperature, humidity, shadow movement and obstacle location of the photovoltaic panels; after detecting the photovoltaic panels, use the multi-spectral acquisition module to scan the surface of the photovoltaic panels using spectral information of different bands to collect data on dirt distribution, type and degree;

[0057] Step 2: Generate a comprehensive image by fusing image information under different spectra through the data fusion processing module.

[0058] Step 3: Based on the fused data, the dirt dynamic identification module is used to analyze the dirt situation on the surface of the photovoltaic panel in real time, including the type of dirt, distribution area and degree of pollution;

[0059] Step 4: According to the output result of the dynamic dirt recognition module, a cleaning path is planned by a path planning module; the dynamic dirt recognition module plans an initial path by an adaptive light compensation path planning algorithm taking into account the layout of the photovoltaic panels, the distribution of dirt and the performance limitations of the robot; the performance limitations of the robot include battery life and cleaning efficiency;

[0060] Step 5: During the cleaning process, the robot dynamically adjusts the cleaning path through a dynamic obstacle avoidance algorithm based on deep reinforcement learning;

[0061] Step 6: During the cleaning process, the robot adjusts the cleaning force and speed according to the output result of the dirt dynamic recognition module;

[0062] Step 7: During the cleaning process, the working status and cleaning progress of the robot are monitored in real time through the cloud platform, and when an abnormal situation occurs, early warning information is fed back through the email communication protocol; the abnormal information includes insufficient power and poor cleaning effect;

[0063] Step 8. After the cleaning task is completed, the robot automatically returns to the designated parking location or charges, and uploads the cleaning process data to the cloud platform for analysis and storage; the cleaning process data includes the cleaning area, cleaning efficiency, cleaning quality, environmental perception data and equipment status data.

[0064] The positive beneficial effects of the present invention are:

[0065] 1. The present invention can capture information under different spectra by collecting images of the photovoltaic panel surface under visible light, near infrared and ultraviolet spectra, effectively making up for the limitations of single spectrum images. This helps to identify dirt, scratches and abnormal light and shadow changes that are difficult to find under a single spectrum. By performing image registration, multi-scale fusion and preprocessing on multi-spectral images, a comprehensive image that integrates image information under different spectra is generated. This not only improves the resolution and clarity of the image, but also enhances the adaptability to complex environments and reduces the "blind spots" of machine vision.

[0066] 2. The present invention extracts feature maps of different scales through a multi-scale pyramid structure to ensure that dirt can be effectively identified at different scales. The adaptive feature fusion mechanism dynamically adjusts the weights of feature maps of different scales, further enhancing the ability to identify complex dirt. The efficiency and robustness of feature aggregation are improved through the attention mechanism. The attention mechanism can dynamically adjust the importance of feature maps to ensure that key features receive full attention. In addition, combined with the light intensity, temperature and humidity data provided by the environmental perception module, the dirt recognition threshold is dynamically adjusted to ensure that dirt can be accurately identified under different environmental conditions. This greatly reduces the problems of misidentification and missed detection caused by environmental changes.

[0067] 3. The present invention uses an adaptive light compensation path planning algorithm to model the light intensity, which can more accurately estimate the surface conditions of the photovoltaic panels under actual lighting conditions. This provides reliable lighting information for path planning and reduces path planning errors caused by changes in lighting. Combined with the information provided by the light compensation model, the path cost is dynamically adjusted so that the cleaning robot can give priority to paths with lighting conditions that are more conducive to cleaning efficiency. The dynamic adjustment of the path cost ensures the flexibility and adaptability of path planning and reduces the risk of "dead loops". By real-time monitoring of changes in light intensity and dynamically adjusting the light compensation coefficient, it is ensured that the path planning strategy can be adjusted according to actual lighting conditions. This further improves the accuracy and robustness of path planning.

[0068] 4. The present invention adopts a dynamic obstacle avoidance algorithm based on deep reinforcement learning. By learning obstacle avoidance strategies through a deep Q network, the cleaning path can be dynamically adjusted according to real-time changes in the environment. The deep Q network continuously optimizes its decision-making strategy through interaction with the environment, ensuring that the robot can complete the cleaning task efficiently and safely in a complex and changing environment. And by storing and reusing past interaction experience, the correlation between samples is reduced, and the stability and convergence speed of training are improved. This enables the robot to learn effective obstacle avoidance strategies more quickly and reduce the risk of "dead loops" caused by path planning errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a schematic diagram of the overall architecture of a photovoltaic cleaning robot based on automatic path planning according to the present invention;

[0070] Figure 2 This is a working principle framework diagram of the dynamic obstacle avoidance algorithm based on deep reinforcement learning of the present invention;

[0071] Figure 3 It is a schematic diagram of the principle framework of the data fusion processing module of the present invention;

[0072] Figure 4 This is a working principle diagram of the adaptive illumination compensation path planning algorithm of the present invention;

[0073] Figure 5 A method step diagram of dynamically adjusting the threshold value of dirt recognition by the threshold value adjustment layer of the present invention;

[0074] Figure 6 This is a step diagram of the operating method of the photovoltaic cleaning robot based on automatic path planning of the present invention. DETAILED DESCRIPTION

[0075] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] In the above embodiment, if Figure 1-Figure 6As shown: In a photovoltaic cleaning robot based on automatic path planning, the multi-spectral acquisition module is used to acquire the surface image of the photovoltaic panel under the visible light, near-infrared and ultraviolet spectra, perform geometric correction and radiation correction, and store the corrected image data; the multi-spectral acquisition module first divides the incident light into visible light, near-infrared and ultraviolet spectral bands through spectral separation, and each spectral band ensures the purity and accuracy of the spectrum through an interference filter; then, the multi-spectral acquisition module converts the separated spectral information into digital image data through a charge coupler, and performs geometric correction and radiation correction on the collected image through an image correction unit; the image correction unit eliminates image distortion through affine transformation, and eliminates the radiation intensity difference of spectral data at different wavelengths through a radiation correction model; the radiation correction model corrects the radiation value of each pixel through polynomial fitting.

[0077] In the specific implementation, the multi-spectral acquisition module is initialized when it is started, and the parameters of the interference filter and the working mode of the CCD are set to ensure that all components work normally. After the initialization is completed, the module enters the standby state and waits for the control system of the photovoltaic cleaning robot to send the acquisition instruction. When the control system sends the acquisition instruction, the multi-spectral acquisition module starts working. The interference filter performs spectral separation on the incident light and divides the light into visible light, near-infrared and ultraviolet spectral bands. The light of each band passes through the corresponding interference filter to ensure the purity and accuracy of the spectrum. The separated spectral information is photoelectrically converted by the charge coupled device (CCD) to generate digital image data. The CCD scans the charge amount of each pixel line by line to generate an analog electrical signal, which is converted into a digital signal by an analog-to-digital converter (ADC). The digital image data is transmitted to the image processing unit through a high-speed data interface for further processing. The image processing unit performs real-time processing on the collected digital image data, including geometric correction and radiation correction. Geometric correction eliminates distortion in the image through affine transformation to ensure the geometric accuracy of the image. Radiation correction corrects the radiation value of each pixel through polynomial fitting to eliminate the difference in radiation intensity of spectral data at different wavelengths. The corrected image data is stored in the local memory for use by the subsequent dirt identification and path planning modules. During the entire working process, the multispectral acquisition module can update the image data in real time according to the requirements of the control system to ensure that the photovoltaic cleaning robot can clean according to the latest environmental information. Among them, the interference filter selectively transmits light of a specific wavelength through the interference effect of the multi-layer dielectric film, thereby ensuring the purity and accuracy of each spectral band. CCD is a highly sensitive photoelectric conversion device that can convert optical signals into electrical signals, and then convert the electrical signals into digital signals through an analog-to-digital converter (ADC). CCD reads the charge amount of each pixel by scanning line by line to generate digital image data. Affine transformation is a linear transformation that can correct distortions such as translation, rotation, scaling and shearing in the image. Radiation correction eliminates the difference in radiation intensity of spectral data at different wavelengths through the radiation correction model to ensure the radiation consistency of the image. In specific implementation, the multispectral acquisition module can generate high-resolution and high-definition image data through spectral separation and CCD technology to ensure the purity and accuracy of the image. This provides a reliable foundation for subsequent dirt identification and path planning, and improves the environmental adaptability and robustness of the entire system. Secondly, geometric correction and radiation correction eliminate distortion and radiation intensity differences in the image, ensure the geometric accuracy and radiation consistency of the image, and improve the quality and reliability of the image. This not only helps to improve the accuracy of dirt identification, but also provides accurate environmental information for path planning, allowing the cleaning robot to complete cleaning tasks efficiently and safely in complex and changing environments.In addition, the multi-spectral acquisition module is able to work under different lighting conditions, and through spectral separation and radiation correction, it ensures that high-quality image data can be obtained under different lighting conditions. This enhances the adaptability of the photovoltaic cleaning robot in complex environments, enabling it to work normally under various weather and lighting conditions.

[0078] The data fusion processing module is used to read multispectral image data, perform image registration, multi-scale fusion and preprocessing, and generate a comprehensive image that integrates image information under different spectra; Figure 3 As shown: the data fusion processing module aligns images under different spectra through an image registration algorithm; the image registration algorithm extracts feature points in the image through a feature point matching method SIFT, and matches the transformation matrix between images under different spectra through a random sample consistency algorithm; the random sample consistency algorithm selects feature point subsets randomly for multiple times, fits the transformation model and verifies the model consistency to select the best transformation matrix to ensure that the pixels of different spectral images correspond consistently; then, the data fusion processing module fuses the high-frequency and low-frequency information of different spectral images through a multi-scale pyramid decomposition method, and performs denoising, contrast enhancement and edge detection operations on the fused image through an image preprocessing method; the working steps of the multi-scale pyramid decomposition method are as follows:

[0079] S301, decomposing each spectral image into image layers of multiple scales through a Gaussian pyramid;

[0080] S302, performing weighted averaging on the same scale layer of different spectral images to generate a fused image layer;

[0081] S303, recombining the fused image layers into a comprehensive image through inverse pyramid transformation;

[0082] Based on the fused composite image, the image preprocessing method dynamically adjusts the filtering parameters according to the local characteristics of the image through the non-local mean, and enhances the contrast through the histogram equalization and Canny edge detection algorithm, thereby improving the visual effect of the image and extracting the edge information in the image. In the specific implementation, the core of the image registration technology is to solve the image misalignment problem caused by factors such as shooting angle, distance, and time, and ensure that images acquired from different perspectives or different time points can be accurately aligned in space. The SIFT (Scale-Invariant Feature Transform) algorithm is the key to achieving this goal. It achieves scale and rotation invariance by detecting key points in the image and calculating its descriptor, and can maintain a high matching accuracy even when the lighting conditions change. The RANSAC (RANdom SAmple Consensus) algorithm further improves the robustness and accuracy of the registration. It estimates the model parameters by randomly selecting sample point pairs, and continuously optimizes the model through iteration, and finally selects the best matching point pairs, constructs the geometric transformation relationship between the two images, and thus achieves accurate image registration.

[0083] The main purpose of multi-scale pyramid decomposition is to process images at different scales in order to better capture the details of the image. The technology first decomposes the original image into multiple image layers of different resolutions, each layer representing a specific scale of the original image. Then, the images are weighted averaged and fused at the same scale, which can effectively retain the detailed features of the image while smoothing out the noise. Finally, the processed image layers are recombined through the inverse pyramid transform to obtain a high-quality image that integrates multispectral information. This method can not only effectively improve the clarity of the image, but also enhance the contrast of the image, providing a better basis for subsequent dirt detection.

[0084] Non-local mean denoising is a denoising method based on image similarity. It assumes that there are a lot of repetitive structures in natural images, and similar image blocks can be found even in different positions. Therefore, this method replaces the original pixel value by calculating the weighted average of each pixel in the image and its surrounding similar pixels. The weight is determined based on the similarity between pixels. This method can better preserve the edges and other important features of the image while removing noise, avoiding the edge blurring problem that may be caused by traditional filters.

[0085] Histogram equalization is a commonly used image enhancement technology. Its basic idea is to adjust the grayscale distribution of the image to make the grayscale value of the image more uniform, thereby achieving the effect of enhancing the image contrast. Specifically, histogram equalization redistributes the probability of each grayscale level in the image so that the new grayscale distribution is as close to a uniform distribution as possible, thereby highlighting the details of the image and improving the visual effect of the image. This technology is particularly suitable for processing images with low contrast and unclear details, and can significantly improve the quality of the image.

[0086] The Canny edge detection algorithm is a gradient-based edge detection method that aims to accurately identify the boundaries of objects from an image. The algorithm first performs a Gaussian filter on the image to reduce the impact of noise, and then calculates the gradient strength and direction of the image to determine the location of the edge. Next, the edge is refined by non-maximum suppression, retaining only those pixels that may be true edges. Finally, the final edge is determined by double threshold processing and edge connection to ensure that the detected edge is both continuous and accurate. The Canny algorithm is widely used in various image processing scenarios due to its good noise resistance and edge detection accuracy, and plays an important role in improving the photovoltaic cleaning robot's perception of the surface state of photovoltaic panels. Through precise image registration and multi-scale fusion technology, the module can effectively integrate image information from different spectra to generate a high-quality comprehensive image, which is crucial for accurately identifying the type and distribution of dirt on the surface of photovoltaic panels. At the same time, the application of image preprocessing technology not only improves the visual quality of the image, but also enhances the contrast and edge features of the image, making dirt detection more sensitive and accurate. This means that the robot can find areas that need to be cleaned earlier and take timely action to avoid the problem of reduced power generation efficiency of photovoltaic panels due to long-term accumulation of dirt. In addition, high-quality images also provide a reliable basis for intelligent planning of cleaning paths, helping the robot avoid obstacles and choose the shortest and safest cleaning route, thereby improving cleaning efficiency and reducing unnecessary energy consumption. Overall, the application of the data fusion processing module not only significantly improves the intelligence level and operating ability of photovoltaic cleaning robots, but also brings more scientific and efficient solutions to the operation and maintenance management of photovoltaic power stations.

[0087] The dirt dynamic recognition module is used to extract multi-scale features based on the adaptive deep dirt analysis framework, identify abnormal areas including dirt, scratches and abnormal light and shadow changes, dynamically adjust the dirt recognition threshold, and classify the identified abnormal areas to generate classification labels and dirt distribution maps; the adaptive deep dirt analysis framework automatically learns the feature representation and classification boundaries of dirt based on the comprehensive image output by the data fusion processing module through the attention enhanced convolutional neural network algorithm; the adaptive deep dirt analysis framework includes an input layer, a multi-scale feature extraction layer, a feature aggregation layer, a threshold adjustment layer, a multi-category classification layer and a dirt distribution map generation layer; the working method of the adaptive deep dirt analysis framework includes the following steps: step 401, inputting the comprehensive image obtained by the data fusion processing module into the adaptive deep dirt analysis framework through the input layer, step 402, extracting multi-scale features from the input through the multi-scale feature extraction layer; the multi-scale feature extraction layer decomposes the comprehensive image into feature maps of multiple scales through a multi-scale pyramid structure, and extracts local and global features using convolution and pooling operations; step 403, The layer aggregates the feature maps generated by the multi-scale extraction layer to form a high-dimensional feature representation; the feature aggregation layer calculates the weight of each feature map through the self-attention mechanism, dynamically adjusts the importance of each feature map, and aggregates feature maps of different scales through feature aggregation; step 404, the light intensity, temperature and humidity data provided by the environment perception module are fused with the feature map through the threshold adjustment layer, and the threshold of dirt recognition is dynamically adjusted; the threshold adjustment layer fuses the environmental data with the feature map through a multimodal fusion method to generate a comprehensive feature representation, and predicts the optimal threshold under the current environmental conditions through a deep learning model; step 405, the multi-scale features extracted in step 402 are classified through the multi-class classification layer to identify abnormal areas including dirt, scratches and abnormal light and shadow changes; the multi-class classification layer generates the probability that each pixel belongs to a different category through a softmax classifier; step 406, the dirt distribution map generation layer generates a dirt distribution map according to the classification results of the multi-class classification layer; the dirt distribution map generation layer generates a high-definition dirt distribution map containing dirt location and degree information through morphological operations and connected area analysis. The method for dynamically adjusting the threshold value of dirt recognition by the threshold adjustment layer includes: step 501, detecting the temperature, humidity and light intensity data of the current environment in real time through the temperature, humidity and light intensity data provided by the environment perception module; step 502, defining the reference threshold limit T0 and the light influence coefficient K I , Temperature influence coefficient K T , Humidity influence coefficient K H, where the reference threshold limit T0 represents the threshold determined under standard light, temperature and humidity conditions, including strong light, normal light and weak light levels; the light influence coefficient K is used to quantify the influence of light intensity on the threshold; the temperature influence coefficient K T Used to quantify the impact of temperature on the threshold; the humidity impact coefficient K H Used to quantify the impact of humidity on the threshold; let the current light intensity be I and the reference light intensity be I0. When the light intensity is higher than the strong light level, the reflectivity of the dirt increases, the image brightness changes greatly, and the dynamic threshold T is lowered to avoid misjudgment; when the light intensity is lower than the weak light level, the reflectivity of the dirt decreases, the image brightness changes little, and the dynamic threshold T is increased to avoid missed detection; the calculation formula of the dynamic threshold T is:

[0088] T=T0+K I (I-I0)+K T (T env -T0)+K H (H env -T0) (1)

[0089] In formula (1), T env Indicates the current ambient temperature; H env Indicates the current ambient humidity;

[0090] Step 503: dynamically adjust the threshold of dirt recognition through the deep learning model, use the light, temperature and humidity data as additional input features, fuse them with the image features, and predict the optimal threshold T under the current environmental conditions through the output layer of the model;

[0091] Step 504: Apply the calculated dynamic threshold T to the dirt identification process and adjust the generation standard of the binary dirt mask. The specific steps include step r1: Process the comprehensive image through the dirt analysis model to generate a probability map of each pixel belonging to dirt; step r2: Compare each pixel point. If the probability of the pixel point belonging to dirt is greater than the dynamic threshold T, mark the pixel point as a dirt area; otherwise, mark it as a non-dirt area; step r3: Generate a binary dirt mask, in which the dirt area is marked as 1 and the non-dirt area is marked as 0. The working steps of the attention enhanced convolutional neural network algorithm include:

[0092] Step 601: The dynamically generated attention map is fused with the multi-scale feature map through the attention enhancement feature fusion formula, so that the model is concentrated in the dirt-dense area; the formula expression of the attention enhancement feature fusion formula is:

[0093]

[0094] In formula (2), AF ijrepresents the feature map after attention enhancement at position (i, j); σ represents the ReLU nonlinear activation function; AM ij Represents the attention map at position (i, j), which is used to guide the fusion process of feature maps. * represents the element-by-element multiplication operation, which is used to convert the attention weight AM ij Applied to the fusion results of multi-scale feature maps; a k Represents the weight of the kth feature map, which is used to balance the contribution of feature maps of different scales to the final feature representation represents the feature map extracted by the kth convolutional layer at position (i, j), which is used to capture information in the image at multiple scales; N is the total number of convolutional layers, indicating the number of feature maps involved in feature fusion;

[0095] Step 602: Optimize the sample recognition accuracy near the classification boundary by using a boundary-sensitive loss function; the boundary-sensitive loss function constructs a loss term by using the intersection-over-union ratio between the prediction result and the true label and combining the distance from the prediction result to the nearest classification boundary, thereby amplifying the loss near the boundary area; the formula expression of the boundary-sensitive loss function is:

[0096]

[0097] In formula (3), L dary represents the boundary sensitive loss, which is used to guide the training process of the model; M is the total number of training samples, which indicates the number of samples considered when calculating the loss function; IoU (H (m) ,h m ) represents the prediction result H of the mth sample (m) With the true label h m The intersection-over-union ratio between Indicates the distance between the prediction result of the mth sample and the nearest classification boundary; w s represents a hyperparameter, which is used to control the influence of the distance term on the loss function; exp represents an exponential function, which is used to amplify the loss near the boundary area;

[0098] Step 603: The threshold of dirt recognition is automatically adjusted according to the prediction results of the model and preset conditions during the training process through the dynamic threshold adjustment mechanism. The calculation formula is:

[0099]

[0100] In formula (4), TAU t+1 is the threshold at the next moment, used for dirt identification; TAU tRepresents the threshold at the current moment, which serves as the basis for adjustment. CL(s) is the confidence score of the sth sample, which indicates the confidence of the model in the classification result of the sample. N(l) represents the preset confidence threshold, which is used to determine whether the confidence of the sample is high enough. α represents the learning rate, which is used to control the speed and amplitude of threshold adjustment. Sign represents the sign function, which is used to judge the increase or decrease of the threshold. S represents the total size of the threshold adjustment sample set.

[0101] In the specific implementation, multi-scale features are extracted through multi-layer convolution and pooling operations, ensuring that the algorithm can capture dirt features of different scales and improve the accuracy and robustness of dirt recognition. Through global pooling operations or attention mechanisms, the information in the multi-scale feature map is aggregated to form a global feature representation, further improving the accuracy and robustness of dirt recognition. The threshold of dirt recognition is dynamically adjusted through adaptive methods, ensuring that the algorithm can adapt to different environmental conditions and improving the accuracy and robustness of dirt recognition. The identified abnormal areas are classified through classification algorithms and classification labels are generated, ensuring that the algorithm can accurately identify and classify different abnormal areas, such as dirt, scratches, and abnormal light and shadow changes. Through the data fusion processing module, multimodal image data is fused into a comprehensive image, ensuring the richness and diversity of the input data and improving the accuracy and robustness of dirt recognition. By generating a dirt distribution map, reliable dirt distribution information is provided to the path planning module, ensuring the rationality and effectiveness of the cleaning path. Based on deep learning technology, especially the attention-enhanced convolutional neural network algorithm, the feature representation and classification boundaries of dirt are automatically learned, ensuring that the algorithm can efficiently and accurately identify dirt. Through the online learning mechanism, the dirt recognition model is continuously optimized to ensure that the algorithm can continue to learn and improve in the long run, improving the accuracy and robustness of dirt recognition.

[0102] In the specific implementation, the hardware working environment of the attention-enhanced convolutional neural network algorithm and the adaptive deep dirt analysis framework includes: multi-core CPU (such as Intel Xeon Gold series) and high-speed GPU (such as NVIDIA Tesla V100 or higher) to accelerate the neural network training and reasoning process. SSD or NVMe SSD is used to store a large amount of image data collected by the photovoltaic cleaning robot, as well as temporary files and results generated during the model training process. And high-speed network interface to ensure the efficiency of data transmission, support real-time reception of image data from the photovoltaic cleaning robot, and quickly transmit the processing results back. Sufficiently large RAM (such as 256GB or more) to support large-scale data processing and memory requirements for model training. The operating system and software environment is a Linux system, equipped with GPU acceleration libraries such as CUDA and cuDNN, and a deep learning framework (such as TensorFlow or PyTorch) to implement and deploy the attention-enhanced convolutional neural network algorithm.

[0103] Based on the above hardware environment, Experiment 1 was designed to verify the positive and beneficial effects of the attention-enhanced convolutional neural network algorithm in the photovoltaic panel dirt recognition task. The experiment was divided into Group A and Group B. Group A used the attention-enhanced convolutional neural network algorithm, and Group B used the traditional convolutional neural network algorithm (such as VGG or ResNet as the basic structure, but without the attention mechanism). Both groups of experiments were carried out in the same hardware working environment mentioned above to ensure the consistency of experimental conditions. The experimental data comes from the image set containing dirt, scratches and abnormal light and shadow changes actually collected by the photovoltaic cleaning robot. The experimental data record is shown in Table 1:

[0104] Table 1 Experimental record data table

[0105]

[0106]

[0107] Through comparative experiments, it can be seen that Group A using the attention-enhanced convolutional neural network algorithm is significantly better than Group B using the traditional algorithm in terms of recognition accuracy, precision, recall and F1 score. Specifically, the average recognition accuracy of Group A increased by about 4.6 percentage points, and the precision and recall rates also increased, and the F1 score was about 4.5 percentage points higher, indicating that the attention enhancement mechanism effectively improves the algorithm's ability to recognize dirt, scratches and abnormal light and shadow changes. In addition, although the processing time of Group A is slightly higher than that of Group B (a difference of about 10ms), this slight increase in time is acceptable considering its significant performance improvement. At the same time, the model size of Group A is slightly larger than that of Group B, but considering the storage capacity of modern computing devices, this difference does not have a significant impact on the deployment and transmission of the model.

[0108] The environmental perception module is used to collect environmental data in the photovoltaic power station in real time, perform preprocessing, build a dynamic environmental matrix model, and store the preprocessed environmental data, wherein the environmental data includes light intensity, temperature, humidity, shadow movement, and obstacle position; the environmental perception module divides the physical space of the photovoltaic power station into multiple grid cells by a gridding method, each grid cell contains environmental data in the unit area, and the environmental data includes light intensity, temperature, humidity, shadow movement, and obstacle position; the gridding method organizes the environmental data into a structured matrix form by space division and data mapping; then, the environmental perception module maps the environmental data to the corresponding grid cells by an interpolation and data filling method to generate a dynamic environmental matrix model; the interpolation includes nearest neighbor interpolation, bilinear interpolation, and spline interpolation; the dynamic environmental matrix model dynamically updates the environmental data by a real-time data stream processing method; the real-time data stream processing method receives and processes the environmental data in real time by a message queue and an event-driven mechanism.

[0109] In the specific implementation, the gridding method divides the physical space of the photovoltaic power station into multiple grid cells of equal size through spatial division. The size of each grid cell can be adjusted according to actual needs, and the appropriate grid size is usually selected to balance the computational complexity and data accuracy. Each grid cell stores environmental data such as the light intensity, temperature, humidity, shadow movement, and obstacle position of the area. The gridding method divides the complex physical space into structured grid cells, which facilitates data management and processing and improves the organization and accessibility of the data. The interpolation methods include nearest neighbor interpolation, bilinear interpolation, and spline interpolation. The nearest neighbor interpolation is suitable for simple scenes by selecting the nearest data point for filling; the bilinear interpolation is suitable for scenes with smooth changes by linear interpolation through the four nearest data points; the spline interpolation is suitable for scenes with complex changes by interpolation through polynomial fitting. The data filling method ensures that each grid cell has complete environmental data by predicting and supplementing missing data points. The interpolation and data filling methods ensure that each grid cell has complete environmental data, improve the integrity and accuracy of the dynamic environment matrix model, and provide reliable data support for path planning and obstacle avoidance. The working mode of the environmental perception module in the photovoltaic cleaning robot based on automatic path planning is a highly automated and real-time process. First, the module collects environmental data in real time through the sensor network, including light intensity, temperature, humidity, shadow movement and obstacle position. These data are transmitted to the environmental perception module through wired or wireless communication. Next, the module divides the physical space of the photovoltaic power station into multiple grid cells through a gridding method, and each grid cell contains the environmental data in the unit area. Specifically, the application of the environmental perception module in the photovoltaic cleaning robot based on automatic path planning has significant advantages and characteristics. Compared with the same type of hardware, the environmental perception module collects environmental data in real time through a variety of sensors to ensure the real-time and accuracy of the data. Compared with the traditional hardware that collects data regularly, the environmental perception module can obtain the latest environmental information more timely, which improves the response speed and real-time performance of the system. Through the message queue and event-driven mechanism, the environmental perception module can receive and process the data collected by the sensor in real time and generate the latest dynamic environment matrix model. Compared with the traditional batch processing method, the real-time data stream processing method can update the environmental data more quickly, ensuring that the photovoltaic cleaning robot can perform dynamic path planning and obstacle avoidance according to the latest environmental information. The environmental perception module divides the physical space of the photovoltaic power station into multiple grid cells through a gridding method, and each grid cell contains the environmental data in the unit area. The gridding method divides the complex physical space into structured grid cells, which facilitates data management and processing, and improves the organization and accessibility of data. Compared with the traditional point-to-point data management method, the gridding method can manage and process a large amount of environmental data more efficiently.Through the interpolation and data filling methods, the environment perception module can map the environmental data to the corresponding grid cells and generate a dynamic environment matrix model. The interpolation method ensures that each grid cell has complete environmental data, which improves the integrity and accuracy of the dynamic environment matrix model. Compared with the traditional data filling method, the interpolation method can more accurately reflect the environmental conditions in the photovoltaic power station and improve the robustness and reliability of the system. The environment perception module generates a dynamic environment matrix model through real-time data acquisition, gridding, interpolation and data filling technologies. The dynamic environment matrix model can reflect the environmental conditions in the photovoltaic power station in real time, providing reliable data support for the path planning and obstacle avoidance of the photovoltaic cleaning robot. Compared with the traditional static environment model, the dynamic environment matrix model can more accurately reflect environmental changes and improve the intelligence level of path planning and obstacle avoidance. Through the message queue and event-driven mechanism, the environment perception module can receive and process the data collected by the sensor in real time and generate the latest dynamic environment matrix model. The real-time data stream processing method ensures that the environment perception module can perform dynamic path planning and obstacle avoidance according to the latest environmental information, avoiding path planning errors and obstacle avoidance failures caused by data delays. Compared with traditional batch processing, real-time data stream processing methods can update environmental data more quickly and improve the intelligence level of the system.

[0110] The path planning module is used to read the dynamic environment matrix model and the dirt distribution map, generate an initial cleaning path through an adaptive illumination compensation path planning algorithm, and monitor the position of obstacles in real time during the cleaning process, and dynamically adjust the cleaning path through a dynamic obstacle avoidance algorithm based on deep reinforcement learning. The adaptive illumination compensation path planning algorithm first receives the dynamic environment matrix model constructed in real time by the sensor network, reads the dirt distribution map, analyzes the light intensity data in the environment matrix through the illumination impact assessment model, and evaluates the impact of illumination in different areas on the power generation efficiency of the photovoltaic panel; then, the illumination impact assessment model converts the illumination intensity value into the corresponding power generation efficiency loss value through the mapping relationship between illumination intensity and power generation efficiency, and quantifies the impact of uneven illumination on the performance of the photovoltaic panel; based on the evaluation result of the illumination impact assessment model, the adaptive illumination compensation path planning algorithm adopts a priority allocation mechanism to comprehensively calculate the priority of each cleaning area according to the degree of dirt and the degree of illumination impact; the priority allocation mechanism weightedly sums the dirt density in the dirt distribution map and the power generation efficiency loss value output by the illumination impact assessment model through the weighted summation method to obtain the comprehensive priority value of each cleaning area; finally, the adaptive illumination compensation path planning algorithm, based on the calculated priority value, preferentially traverses the areas with high priority while maintaining the overall efficiency of the cleaning path through a heuristic search algorithm to generate an initial cleaning path.

[0111] In specific implementation, Figure 4As shown in the figure, the adaptive light compensation path planning algorithm evaluates the impact of light in different areas on the power generation efficiency of photovoltaic panels through the light impact assessment model, and quantifies the impact of uneven light on the performance of photovoltaic panels. Compared with the traditional path planning method, the light impact assessment model can determine the cleaning priority more scientifically and improve the cleaning efficiency and effect. Through the priority allocation mechanism, the adaptive light compensation path planning algorithm comprehensively calculates the priority of each cleaning area according to the degree of dirt and the degree of light impact, ensuring the rationality and effectiveness of the cleaning path. Compared with the traditional fixed priority path planning method, the priority allocation mechanism can adapt to environmental changes more flexibly and improve the cleaning efficiency and effect. The adaptive light compensation path planning algorithm generates the initial cleaning path through the heuristic search algorithm to ensure that the path planning module can traverse the areas with high priority first. Compared with the traditional shortest path algorithm, the heuristic search algorithm can generate the optimal cleaning path more efficiently, ensuring that the photovoltaic cleaning robot can complete the efficient cleaning task within a limited time. During the cleaning process, the adaptive light compensation path planning algorithm monitors the obstacle position in real time through the dynamic obstacle avoidance algorithm based on deep reinforcement learning, and dynamically adjusts the cleaning path. Compared with the traditional static path planning method, the dynamic obstacle avoidance algorithm can complete the cleaning task more safely and efficiently, avoiding path planning errors and obstacle avoidance failures caused by obstacles. Through the real-time data stream processing method, the adaptive illumination compensation path planning algorithm can update the environmental data in real time to ensure the real-time and accuracy of the data. Compared with the traditional batch processing method, the real-time data stream processing method can update the environmental data more quickly, improving the intelligence level of the system.

[0112] In the specific implementation, the hardware working environment of the adaptive illumination compensation path planning algorithm includes: sensor network: including illumination sensors, temperature sensors, humidity sensors, infrared sensors, etc., which are used to collect environmental data in the photovoltaic power station in real time. Multispectral camera: used to collect multispectral images of the surface of the photovoltaic panel and generate a dirt distribution map. Embedded computer: used to process the data collected by the sensor network and the multispectral camera, and generate a dynamic environment matrix model and a dirt distribution map. Storage device: used to store environmental data and dirt distribution map to ensure the security and integrity of the data. Central processing unit (CPU): used to run the adaptive illumination compensation path planning algorithm and generate an initial cleaning path. Graphics processing unit (GPU): used to accelerate the calculation tasks in the path planning algorithm and improve the operation efficiency of the algorithm. Drive motor: used to control the movement of the photovoltaic cleaning robot and clean according to the generated cleaning path. Obstacle avoidance sensor: including ultrasonic sensor, laser radar, etc., used to monitor the position of obstacles in real time and dynamically adjust the cleaning path. Wireless communication module: used to realize data transmission between the path planning module and the execution module to ensure real-time update and adjustment of the cleaning path.

[0113] Based on the above hardware environment, a group of photovoltaic panel cleaning robots were selected, and the adaptive light compensation path planning algorithm (Group A) and the shortest path planning algorithm (Group B) were used for comparative experiments. Both groups of experiments were carried out under the same photovoltaic panel array and dirt distribution conditions. The cleaning time, cleaning coverage and photovoltaic panel power generation efficiency improvement data of Group A and Group B after adaptive light compensation were compared. Five experiments were conducted in each group. The experimental results are shown in Table 2:

[0114] Table 2 Experimental results record

[0115]

[0116] It can be seen from the data table that the adaptive light compensation path planning algorithm can significantly reduce the cleaning time and improve the cleaning efficiency. It can also cover the surface of the photovoltaic panel more comprehensively, improve the cleaning effect, and significantly improve the power generation efficiency of the photovoltaic panel.

[0117] Furthermore, the dynamic obstacle avoidance algorithm based on deep reinforcement learning includes a perception layer, a feature extraction layer, a decision layer, a dynamic adjustment layer and an online learning layer; the working method of the dynamic obstacle avoidance algorithm based on deep reinforcement learning includes:

[0118] Step 901: obtaining dynamic information of the working environment of the photovoltaic cleaning robot through the perception layer;

[0119] Step 902: The feature extraction layer performs deep processing on the data acquired by the perception layer; the feature extraction layer extracts high-level abstract features from the original data through multi-layer convolution and pooling operations;

[0120] Step 903: Make a decision through the decision layer. The decision layer uses the deep Q network to continuously iteratively update the parameters of the policy network through the gradient descent method according to the current environment state vector, combined with the past experience in the experience replay pool and the current policy network, so as to maximize the cumulative reward signal. During the decision-making process, the decision layer also balances the conflicts among the cleaning speed, obstacle avoidance success rate and collision penalty items through a multi-objective value function.

[0121] Step 904: When the dynamic adjustment layer detects environmental changes or obstacles, the obstacle avoidance process is triggered; the dynamic adjustment layer predicts the future movement trajectory of the obstacle through LSTM-based sequence prediction, and calculates the optimal obstacle avoidance path in combination with the action suggestions output by the current policy network; at the same time, the dynamic adjustment layer also evaluates environmental uncertainty through an uncertainty quantification method, and adjusts the balance between exploration and utilization based on environmental uncertainty to ensure that robust decisions can still be made in an uncertain environment;

[0122] Step 905: Learn environmental data and obstacle avoidance experience during the cleaning process through the online learning layer, and update the strategy network in real time through the incremental learning method.

[0123] In the photovoltaic cleaning robot based on automatic path planning, the dynamic obstacle avoidance algorithm based on deep reinforcement learning deeply integrates the advantages of deep learning and reinforcement learning, and simulates the decision-making process of humans in the face of complex environments by building a multi-layer neural network. Specifically, the algorithm collects environmental information through the perception layer, such as the distribution of dirt on the surface of the photovoltaic panel, the location of obstacles and the intensity of light, etc., and this information is then sent to the feature extraction layer for deep processing. The feature extraction layer uses the convolution and pooling operations of the convolutional neural network (CNN) to automatically learn and extract high-dimensional features that are useful for obstacle avoidance decisions from the original data. This process effectively reduces data redundancy and improves decision-making efficiency. In terms of working method, the algorithm follows the basic framework of reinforcement learning, namely the "perception-decision-action-feedback" cycle. First, the perception layer continuously monitors the environment and obtains real-time data; then, the feature extraction layer processes this data to form an environmental state vector that can be used for decision-making. The decision layer is based on the deep Q network (DQN), combined with experience replay and gradient descent method, to continuously optimize the parameters of the policy network to maximize the long-term cumulative reward. Here, the design of the multi-objective value function is particularly critical. It not only considers the maximization of the cleaning speed, but also weighs the obstacle avoidance success rate and collision penalty to ensure that the robot can safely avoid obstacles while cleaning efficiently.

[0124] In the specific implementation, the dynamic adjustment layer uses the long short-term memory network (LSTM) to predict the future movement trajectory of obstacles, and combines the output of the current strategy network to calculate the optimal obstacle avoidance path in real time. In addition, through the uncertainty quantification method, the dynamic adjustment layer can also evaluate the uncertainty of the environment and adjust the balance between exploration and utilization accordingly, enhancing the robot's adaptability in uncertain environments. This dynamic adjustment mechanism enables the robot to make rapid and robust decisions when faced with emergencies, ensuring the continuity and safety of cleaning operations.

[0125] Compared with traditional photovoltaic cleaning robots and their obstacle avoidance algorithms, the dynamic obstacle avoidance algorithm based on deep reinforcement learning has significant advantages and characteristics. First, its powerful environmental perception ability enables the robot to more accurately identify obstacles and dirt distribution, providing strong support for precise cleaning. Secondly, through the combination of deep learning and reinforcement learning, the robot can continuously learn and optimize obstacle avoidance strategies, improving its adaptability to complex environments and decision-making efficiency. In addition, the introduction of the dynamic adjustment layer enables the robot to make rapid and robust decisions when faced with emergencies, ensuring the continuity and safety of cleaning operations. Finally, the incremental update mechanism based on online learning enables the robot to continuously accumulate obstacle avoidance experience and optimize the strategy network, achieving continuous performance improvement.

[0126] Specifically, the application effect of the dynamic obstacle avoidance algorithm based on deep reinforcement learning in the photovoltaic cleaning robot was verified through experiments; this experiment selected a group of identical hardware platforms, equipped with a dynamic obstacle avoidance algorithm based on deep reinforcement learning (Group A) and a traditional obstacle avoidance algorithm (Group B), and conducted comparative tests in a simulated photovoltaic panel cleaning environment. The experimental environment was set to contain scenes with different lighting conditions and randomly distributed obstacles to ensure the consistency of each experimental condition. During the experiment, the cleaning efficiency, obstacle avoidance success rate, path smoothness, and energy efficiency data were recorded. The experimental data records are shown in Table 3:

[0127] Table 3 Algorithm experiment comparison table

[0128]

[0129] From the data table 3, we can see that the average cleaning efficiency of group A algorithm reaches 1.28m 2 / min, which is about 16% higher than that of Group B, indicating that the algorithm of Group A can clean a larger area of ​​photovoltaic panels in the same time, thereby improving work efficiency. In addition, the obstacle avoidance success rate of the algorithm of Group A reached 99.0% on average, which is much higher than the 88.4% of Group B. This proves that the algorithm of Group A has higher stability and reliability in complex environments, and can more effectively avoid collisions with obstacles. At the same time, the cleaning path generated by the algorithm of Group A is smoother, with a standard deviation of only 0.07m, which is about 50% less than the 0.14m of Group B. This means that the algorithm of Group A is more sophisticated in planning the path, reducing unnecessary turns and pauses, thereby improving cleaning efficiency and energy utilization. Secondly, the energy consumption efficiency of the algorithm of Group A is 0.0026kWh / m 2 , compared with 0.0033kWh / m 2 This shows that while maintaining efficient cleaning, the A-group algorithm can also significantly reduce energy consumption, which is in line with the green and energy-saving environmental protection concept.

[0130] Furthermore, an operation method of a photovoltaic cleaning robot based on automatic path planning includes the following steps: Step 1, a comprehensive scan of the surrounding environment is performed through an environmental perception module to identify the position, shape, spacing, light intensity, temperature, humidity, shadow movement and obstacle position of the photovoltaic panel; after the photovoltaic panel is detected, the multi-spectral acquisition module is used to scan the surface of the photovoltaic panel using spectral information of different bands to collect dirt distribution, type and degree data;

[0131] Step 2: Generate a comprehensive image by fusing image information under different spectra through the data fusion processing module.

[0132] Step 3: Based on the fused data, the dirt dynamic identification module is used to analyze the dirt situation on the surface of the photovoltaic panel in real time, including the type of dirt, distribution area and degree of pollution;

[0133] Step 4: According to the output result of the dynamic dirt recognition module, a cleaning path is planned by a path planning module; the dynamic dirt recognition module plans an initial path by an adaptive light compensation path planning algorithm taking into account the layout of the photovoltaic panels, the distribution of dirt and the performance limitations of the robot; the performance limitations of the robot include battery life and cleaning efficiency;

[0134] Step 5: During the cleaning process, the robot dynamically adjusts the cleaning path through a dynamic obstacle avoidance algorithm based on deep reinforcement learning;

[0135] Step 6: During the cleaning process, the robot adjusts the cleaning force and speed according to the output result of the dirt dynamic recognition module;

[0136] Step 7: During the cleaning process, the working status and cleaning progress of the robot are monitored in real time through the cloud platform, and when an abnormal situation occurs, early warning information is fed back through the email communication protocol; the abnormal information includes insufficient power and poor cleaning effect;

[0137] Step 8. After the cleaning task is completed, the robot automatically returns to the designated parking location or charges, and uploads the cleaning process data to the cloud platform for analysis and storage; the cleaning process data includes the cleaning area, cleaning efficiency, cleaning quality, environmental perception data and equipment status data.

[0138] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A photovoltaic cleaning robot based on automatic path planning, characterized in that: It includes multi-spectral acquisition module, data fusion processing module, dirt dynamic identification module, environment perception module and path planning module; The multi-spectral acquisition module is used to acquire the surface image of the photovoltaic panel under the visible light, near infrared and ultraviolet spectrum, perform geometric correction and radiation correction, and store the corrected image data; The data fusion processing module is used to read multispectral image data, perform image registration, multi-scale fusion and preprocessing, and generate a comprehensive image that integrates image information under different spectra; The dirt dynamic identification module is used to extract multi-scale features based on the adaptive deep dirt analysis framework, identify abnormal areas including dirt, scratches and abnormal light and shadow changes, dynamically adjust the dirt identification threshold, and classify the identified abnormal areas to generate classification labels and dirt distribution maps; the adaptive deep dirt analysis framework automatically learns the feature representation and classification boundaries of dirt through the attention enhanced convolutional neural network algorithm based on the comprehensive image output by the data fusion processing module; the adaptive deep dirt analysis framework includes an input layer, a multi-scale feature extraction layer, a feature aggregation layer, a threshold adjustment layer, a multi-category classification layer and a dirt distribution map generation layer; The environmental perception module is used to collect environmental data in the photovoltaic power station in real time, perform preprocessing, build a dynamic environmental matrix model, and store the preprocessed environmental data, wherein the environmental data includes light intensity, temperature, humidity, shadow movement, and obstacle position; The path planning module is used to read the dynamic environment matrix model and the dirt distribution map, generate an initial cleaning path through an adaptive illumination compensation path planning algorithm, and monitor the position of obstacles in real time during the cleaning process, and dynamically adjust the cleaning path through a dynamic obstacle avoidance algorithm based on deep reinforcement learning.

2. A photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The multi-spectral acquisition module first divides the incident light into visible light, near-infrared and ultraviolet spectral bands through spectral separation, and each spectral band is ensured to have a pure and accurate spectrum through an interference filter; then, the multi-spectral acquisition module converts the separated spectral information into digital image data through a charge coupled device, and performs geometric correction and radiation correction on the collected image through an image correction unit; the image correction unit eliminates image distortion through affine transformation, and eliminates the difference in radiation intensity of spectral data at different wavelengths through a radiation correction model; the radiation correction model corrects the radiation value of each pixel through polynomial fitting.

3. A photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The data fusion processing module aligns images under different spectra through an image registration algorithm; the image registration algorithm extracts feature points in the image through a feature point matching-based method SIFT, and matches the transformation matrix between images under different spectra through a random sample consistency algorithm; the random sample consistency algorithm selects a subset of feature points randomly for multiple times, fits the transformation model and verifies the model consistency to select the best transformation matrix to ensure that the pixels of different spectral images correspond consistently; then, the data fusion processing module fuses the high-frequency and low-frequency information of different spectral images through a multi-scale pyramid decomposition method, and performs denoising, contrast enhancement and edge detection operations on the fused image through an image preprocessing method; the working steps of the multi-scale pyramid decomposition method are as follows: S301, decomposing each spectral image into image layers of multiple scales by using a Gaussian pyramid; S302, performing weighted averaging on the same scale layer of different spectral images to generate a fused image layer; S303, recombining the fused image layers into a comprehensive image through inverse pyramid transformation; Based on the fused comprehensive image, the image preprocessing method dynamically adjusts the filtering parameters according to the local characteristics of the image through the non-local mean, and enhances the contrast through the histogram equalization and Canny edge detection algorithm, thereby improving the visual effect of the image and extracting the edge information in the image.

4. A photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The working method of the adaptive deep fouling analysis framework comprises the following steps: Step 401: input the comprehensive image obtained by the data fusion processing module into the adaptive deep dirt analysis framework through the input layer. Step 402: extracting multi-scale features from the input through the multi-scale feature extraction layer; the multi-scale feature extraction layer decomposes the comprehensive image into feature maps of multiple scales through a multi-scale pyramid structure, and extracts local and global features using convolution and pooling operations; Step 403: The feature aggregation layer aggregates the feature maps generated by the multi-scale extraction layer to form a high-dimensional feature representation; the feature aggregation layer calculates the weight of each feature map through a self-attention mechanism, dynamically adjusts the importance of each feature map, and aggregates feature maps of different scales through feature aggregation; Step 404: The light intensity, temperature and humidity data provided by the environment perception module are integrated with the feature map through the threshold adjustment layer, and the threshold of dirt recognition is dynamically adjusted; the threshold adjustment layer integrates the environment data with the feature map through a multimodal fusion method to generate a comprehensive feature representation, and predicts the optimal threshold under the current environmental conditions through a deep learning model; Step 405: classify the multi-scale features extracted in step 402 through the multi-class classification layer to identify abnormal areas including dirt, scratches, and abnormal light and shadow changes; the multi-class classification layer generates the probability that each pixel belongs to a different category through a softmax classifier; Step 406: Generate a dirt distribution map through the dirt distribution map generation layer according to the classification results of the multi-category classification layer; the dirt distribution map generation layer generates a high-definition dirt distribution map containing dirt location and degree information through morphological operations and connected region analysis.

5. A photovoltaic cleaning robot based on automatic path planning according to claim 4, characterized in that: The method for dynamically adjusting the threshold value of dirt recognition by the threshold adjustment layer includes: Step 501: Real-time detection of the temperature, humidity and light intensity conditions of the current environment through the temperature, humidity and light intensity data provided by the environment sensing module; Step 502: Define the reference threshold T0 and the illumination influence coefficient K I , Temperature influence coefficient K T , Humidity influence coefficient K H , where the reference threshold limit T0 represents the threshold determined under standard lighting, temperature and humidity conditions, including strong light, normal light and weak light levels; the lighting influence coefficient K is used to quantify the influence of lighting intensity on the threshold; the temperature influence coefficient K T Used to quantify the impact of temperature on the threshold; the humidity impact coefficient K H Used to quantify the impact of humidity on the threshold; assuming the current light intensity is I, and the reference light intensity is I0, when the light intensity I is higher than the strong light level, the reflectivity of the dirt increases, the image brightness changes greatly, and the dynamic threshold T is reduced to avoid misjudgment; when the light intensity I is lower than the weak light level, the reflectivity of the dirt decreases, the image brightness changes little, and the dynamic threshold T is increased to avoid missed detection; the calculation formula of the dynamic threshold T is: T=T0+K I (I-I0)+K T (T env -T0)+K H (H env -T0) (1) In formula (1), T env Indicates the current ambient temperature; H env Indicates the current ambient humidity; Step 503: dynamically adjust the threshold of dirt recognition through the deep learning model, use the light, temperature and humidity data as additional input features, fuse them with the image features, and predict the optimal threshold T under the current environmental conditions through the output layer of the model; Step 504: Apply the calculated dynamic threshold T to the dirt identification process to adjust the generation standard of the binary dirt mask. The specific steps include: Step r1, processing the comprehensive image through the dirt analysis model to generate a probability map of each pixel belonging to dirt; Step r2: Compare each pixel point. If the probability that the pixel point belongs to dirt is greater than the dynamic threshold T, the pixel point is marked as a dirt area; otherwise, it is marked as a non-dirt area. Step r3: Generate a binary dirt mask, where the dirt area is marked as 1 and the non-dirt area is marked as 0.

6. A photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The working steps of the attention-enhanced convolutional neural network algorithm include: Step 601: The dynamically generated attention map is fused with the multi-scale feature map through the attention enhancement feature fusion formula, so that the model is concentrated in the dirt-dense area; the formula expression of the attention enhancement feature fusion formula is: In formula (2), AF ij represents the feature map after attention enhancement at position (i, j); σ represents the ReLU nonlinear activation function; AM ij Represents the attention map at position (i, j), which is used to guide the fusion process of feature maps; * represents the element-by-element multiplication operation, which is used to convert the attention weight AM ij Applied to the fusion results of multi-scale feature maps; α k Represents the weight of the kth feature map, which is used to balance the contribution of feature maps of different scales to the final feature representation; represents the feature map extracted by the kth convolutional layer at position (i, j), which is used to capture information in the image at multiple scales; N is the total number of convolutional layers, indicating the number of feature maps involved in feature fusion; Step 602: Optimize the sample recognition accuracy near the classification boundary by using a boundary-sensitive loss function; the boundary-sensitive loss function constructs a loss term by using the intersection-over-union ratio between the prediction result and the true label and combining the distance from the prediction result to the nearest classification boundary, thereby amplifying the loss near the boundary area; the formula expression of the boundary-sensitive loss function is: In formula (3), L dary represents the boundary sensitive loss, which is used to guide the training process of the model; M is the total number of training samples, which indicates the number of samples considered when calculating the loss function; IoU (H (m) ,h m ) represents the prediction result H of the mth sample (m) With the true label h m The intersection-over-union ratio between them is used to measure the accuracy of classification; Indicates the distance between the prediction result of the mth sample and the nearest classification boundary; w s represents a hyperparameter, which is used to control the influence of the distance term on the loss function; exp represents an exponential function, which is used to amplify the loss near the boundary area; Step 603: The threshold of dirt recognition is automatically adjusted according to the prediction results of the model and preset conditions during the training process through a dynamic threshold adjustment mechanism. The calculation formula is: In formula (4), TAU t+1 is the threshold at the next moment, used for dirt identification; TAU t represents the current threshold, which serves as the basis for adjustment; CL(s) is the confidence score of the sth sample, which indicates the confidence of the model in the classification result of the sample; N(l) represents the preset confidence threshold, which is used to judge whether the confidence of the sample is high enough; α represents the learning rate, which is used to control the speed and amplitude of threshold adjustment; sign represents the sign function, which is used to judge the increase or decrease of the threshold; S represents the total size of the threshold adjustment sample set.

7. A photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The environmental perception module divides the physical space of the photovoltaic power station into multiple grid units through a gridding method, and each grid unit contains environmental data in the unit area, and the environmental data includes light intensity, temperature, humidity, shadow movement and obstacle position; the gridding method organizes the environmental data into a structured matrix form through space division and data mapping; then, the environmental perception module maps the environmental data to the corresponding grid units through interpolation and data filling methods to generate a dynamic environmental matrix model; the interpolation includes nearest neighbor interpolation, bilinear interpolation and spline interpolation; the dynamic environmental matrix model dynamically updates the environmental data through a real-time data stream processing method; the real-time data stream processing method receives and processes the environmental data in real time through a message queue and an event-driven mechanism.

8. The photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The adaptive illumination compensation path planning algorithm first receives the dynamic environment matrix model constructed in real time by the sensor network, reads the dirt distribution map, analyzes the light intensity data in the environment matrix through the illumination impact assessment model, and evaluates the impact of illumination in different areas on the power generation efficiency of the photovoltaic panel; then, the illumination impact assessment model converts the illumination intensity value into the corresponding power generation efficiency loss value through the mapping relationship between illumination intensity and power generation efficiency, and quantifies the impact of uneven illumination on the performance of the photovoltaic panel; based on the evaluation result of the illumination impact assessment model, the adaptive illumination compensation path planning algorithm adopts a priority allocation mechanism to comprehensively calculate the priority of each cleaning area according to the degree of dirt and the degree of illumination impact; the priority allocation mechanism weightedly sums the dirt density in the dirt distribution map and the power generation efficiency loss value output by the illumination impact assessment model through the weighted summation method to obtain the comprehensive priority value of each cleaning area; finally, the adaptive illumination compensation path planning algorithm, based on the calculated priority value, preferentially traverses the areas with high priority while maintaining the overall efficiency of the cleaning path through a heuristic search algorithm to generate an initial cleaning path.

9. The photovoltaic cleaning robot based on automatic path planning according to claim 1, characterized in that: The dynamic obstacle avoidance algorithm based on deep reinforcement learning includes a perception layer, a feature extraction layer, a decision layer, a dynamic adjustment layer and an online learning layer; the working method of the dynamic obstacle avoidance algorithm based on deep reinforcement learning includes: Step 901: obtaining dynamic information of the working environment of the photovoltaic cleaning robot through the perception layer; Step 902: The feature extraction layer performs deep processing on the data acquired by the perception layer; the feature extraction layer extracts high-level abstract features from the original data through multi-layer convolution and pooling operations; Step 903: Make a decision through the decision layer. The decision layer uses the deep Q network to continuously iteratively update the parameters of the policy network through the gradient descent method according to the current environment state vector, combined with the past experience in the experience replay pool and the current policy network, so as to maximize the cumulative reward signal. During the decision-making process, the decision layer also balances the conflicts among the cleaning speed, obstacle avoidance success rate and collision penalty items through a multi-objective value function. Step 904: When the dynamic adjustment layer detects environmental changes or obstacles, the obstacle avoidance process is triggered; the dynamic adjustment layer predicts the future movement trajectory of the obstacle through LSTM-based sequence prediction, and calculates the optimal obstacle avoidance path in combination with the action suggestions output by the current policy network; at the same time, the dynamic adjustment layer also evaluates environmental uncertainty through an uncertainty quantification method, and adjusts the balance between exploration and utilization based on environmental uncertainty to ensure that robust decisions can still be made in an uncertain environment; Step 905: Learn environmental data and obstacle avoidance experience during the cleaning process through the online learning layer, and update the strategy network in real time through the incremental learning method.

10. An operation method of a photovoltaic cleaning robot based on automatic path planning, characterized in that: A photovoltaic cleaning robot based on automatic path planning as described in any one of claims 1 to 9 comprises the following steps: Step 1: Use the environmental perception module to comprehensively scan the surrounding environment and identify the location, shape, spacing, light intensity, temperature, humidity, shadow movement and obstacle location of the photovoltaic panels; after detecting the photovoltaic panels, use the multi-spectral acquisition module to scan the surface of the photovoltaic panels using spectral information of different bands to collect data on dirt distribution, type and degree; Step 2: Generate a comprehensive image by fusing image information under different spectra through the data fusion processing module. Step 3: Based on the fused data, the dirt dynamic identification module is used to analyze the dirt situation on the surface of the photovoltaic panel in real time, including the type of dirt, distribution area and degree of pollution; Step 4: According to the output result of the dynamic dirt recognition module, a cleaning path is planned by a path planning module; the dynamic dirt recognition module plans an initial path by an adaptive light compensation path planning algorithm taking into account the layout of the photovoltaic panels, the distribution of dirt and the performance limitations of the robot; the performance limitations of the robot include battery life and cleaning efficiency; Step 5: During the cleaning process, the robot dynamically adjusts the cleaning path through a dynamic obstacle avoidance algorithm based on deep reinforcement learning; Step 6: During the cleaning process, the robot adjusts the cleaning force and speed according to the output result of the dirt dynamic recognition module; Step 7: During the cleaning process, the working status and cleaning progress of the robot are monitored in real time through the cloud platform, and when an abnormal situation occurs, early warning information is fed back through the email communication protocol; the abnormal information includes insufficient power and poor cleaning effect; Step 8. After the cleaning task is completed, the robot automatically returns to the designated parking location or charges, and uploads the cleaning process data to the cloud platform for analysis and storage; the cleaning process data includes the cleaning area, cleaning efficiency, cleaning quality, environmental perception data and equipment status data.

Images