A visual feedback-based photovoltaic panel frame dust cleaning system

Abstract

The application discloses a visual feedback-based photovoltaic panel frame dust cleaning system, which comprises a frame dust target recognition module and a fixed-point dustless cleaning module; the frame dust target recognition module is used for recognizing a dust position based on a dust detection model; and the fixed-point dustless cleaning module is used for cleaning dust based on the dust position. The application solves the problem of photovoltaic panel frame dust by using a machine vision method and a waterless cleaning mode, and has higher dust recognition precision and efficient waterless cleaning capacity, compared with a traditional photovoltaic cleaning robot which has low efficiency, high power consumption and incomplete local cleaning due to no cleaning feedback, thereby providing important guiding significance for intelligent operation and maintenance development of power stations.

Description

Technical Field

[0001] This application belongs to the field of solar photovoltaic power generation technology, specifically relating to a photovoltaic panel frame dust cleaning system based on visual feedback. Background Technology

[0002] With the long-term operation of photovoltaic (PV) systems, the problem of dust accumulation and cleaning has become increasingly prominent. Dust from the atmosphere accumulates on the surface of PV panels, hindering heat dissipation and severely reducing the power generation efficiency and system safety of the PV system. In particular, dust accumulation on the frame of the PV panels can cause the maximum power point (MPP) of the PV module to be lower than that of other modules. Due to the "weakest link" effect of the series connection, this can reduce the power generation of the entire string of modules, even by more than 10%. Currently, PV cleaning robots are undoubtedly a new and efficient technology for solving the problem of dust accumulation on PV panels, and have achieved good practical application results in PV fields. However, traditional PV cleaning robots still have problems such as low work efficiency, high power consumption, and incomplete cleaning. Operational experience shows an urgent need for a system that can precisely target and efficiently clean dust accumulation on the frame of PV panels, providing on-site maintenance personnel with a simple and efficient maintenance tool. Summary of the Invention

[0003] This application aims to address the shortcomings of existing technologies by proposing a photovoltaic panel frame dust cleaning system based on visual feedback. By utilizing an embedded dust accumulation visual detection model and a waterless cleaning method, it can achieve accurate identification and efficient cleaning of dust accumulation on the photovoltaic panel frame. This overcomes the problems of low efficiency, high power consumption, and incomplete cleaning due to lack of feedback in traditional photovoltaic cleaning robots, thereby improving dust removal efficiency and saving energy. This is of great significance for realizing refined operation and maintenance of photovoltaic systems.

[0004] To achieve the above objectives, this application provides the following solution:

[0005] A photovoltaic panel frame dust cleaning system based on visual feedback includes: a frame dust accumulation target recognition module and a fixed-point dust-free cleaning module;

[0006] The edge dust accumulation target recognition module is used to identify the location of dust accumulation based on the dust accumulation detection model;

[0007] The fixed-point dust-free cleaning module is used to clean dust based on the location of the accumulated dust.

[0008] Preferably, the border dust accumulation target recognition module includes: an image acquisition device and a dust accumulation recognition device;

[0009] The image acquisition device is used to acquire images of photovoltaic panels;

[0010] The dust accumulation identification device is used to detect the photovoltaic panel image based on the dust accumulation detection model to obtain the dust accumulation location.

[0011] Preferably, the dust accumulation detection model includes: an image preprocessing network, a backbone network, a neck network, and a head network;

[0012] The image preprocessing network is used to scale the photovoltaic panel image, normalize it, and sort its channels to obtain the processed image.

[0013] The backbone network is used to perform convolutional pooling operations on the processed image to obtain feature images at different levels.

[0014] The neck network is used to fuse the feature images at different levels to obtain a fused image;

[0015] The head network is used to predict the target based on the fused image to obtain the location of the ash accumulation.

[0016] Preferably, the backbone network includes: a first convolutional unit and a second convolutional unit;

[0017] The first convolutional unit includes one layer of FOUCS units;

[0018] The second convolutional unit includes four sets of CBS units and DSConv units.

[0019] Preferably, the fixed-point dust-free cleaning module includes: a moving unit, a dust-generating unit, and a dust removal unit;

[0020] The moving unit is used to control the cleaning system to move to the dust accumulation location;

[0021] The dust-generating unit is used to complete the dust-generating process and generate a dust-air mixture.

[0022] The dust removal unit is used to extract the dust-air mixture, settle large dust particles, and filter small dust particles.

[0023] Preferably, the mobile unit includes: a Beidou navigation device, a negative pressure mobile carrier, and drive wheels;

[0024] The workflow of the mobile unit includes: the Beidou navigation module provides position information on the photovoltaic panel surface based on the location of the accumulated dust, and the drive wheel drives the negative pressure mobile carrier to control the cleaning system to move on the photovoltaic panel surface.

[0025] Preferably, the dust-generating unit includes: a lightweight nylon roller brush, a DC motor, and a coupling;

[0026] The working process of the dust-raising unit includes: connecting the DC motor to the nylon lightweight roller brush via the coupling; the DC motor drives the nylon lightweight roller brush to rotate; the bristles of the nylon lightweight roller brush sweep across the surface of the photovoltaic panel, destroying the dust deposition structure and lifting it up, thus completing the dust-raising process and generating the dust-air mixture gas.

[0027] Preferably, the dust removal unit includes: a centrifugal fan, a filter screen, and a dust collection box;

[0028] The working process of the dust removal unit includes: the centrifugal fan guides the dust-air mixture into the dust collection box through the air intake channel; larger dust particles are deposited at the bottom of the dust collection box under the action of gravity; smaller dust particles are filtered through the filter screen and separated from the air; the clean air after dust removal is drawn away by the centrifugal fan and discharged from the cleaning system through the vent.

[0029] Preferably, the fixed-point dust-free cleaning module further includes: a sealing strip and a dust suction chamber;

[0030] The sealed area formed by the sealing strip and the dust suction chamber confines suspended dust particles, preventing dust from falling onto the cleaned photovoltaic modules and causing secondary pollution.

[0031] Compared with the prior art, the beneficial effects of this application are as follows:

[0032] This application utilizes machine vision-based methods and waterless cleaning to solve the problem of dust accumulation on photovoltaic panel frames. Compared to traditional photovoltaic cleaning robots, which suffer from low efficiency, high power consumption, and incomplete cleaning due to lack of feedback, this application offers higher accuracy in dust accumulation identification and efficient waterless cleaning capabilities, providing significant guidance for the intelligent operation and maintenance of power plants. This application allows for adjusting the cleaning system's dwell time based on the accumulation and distribution of dust on the frames to achieve targeted and intelligent cleaning. Compared to traditional photovoltaic cleaning systems, it enhances frame dust accumulation identification capabilities, resulting in higher cleaning efficiency and lower power consumption. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of this application, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the system structure according to an embodiment of this application;

[0035] Figure 2 This is an overall schematic diagram of an embodiment of this application;

[0036] Figure 3 This is a schematic diagram illustrating the working principle of the dust-generating unit in an embodiment of this application;

[0037] Figure 4 This is a schematic diagram illustrating the working principle of the dust removal unit in an embodiment of this application;

[0038] Figure 5 This is a schematic diagram illustrating the working principle of the dust collection box in an embodiment of this application. Detailed Implementation

[0039] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] In this embodiment, as Figure 1 , 2 As shown, a photovoltaic panel frame dust cleaning system based on visual feedback includes: a frame dust target recognition module and a fixed-point dust-free cleaning module.

[0042] The border-based dust accumulation target recognition module is used to identify the location of dust accumulation based on a dust accumulation detection model. The border-based dust accumulation target recognition module includes an image acquisition device and a dust accumulation recognition device.

[0043] An image acquisition device is used to acquire images of photovoltaic panels; a dust accumulation recognition device is used to detect the photovoltaic panel images based on a dust accumulation detection model to obtain the location of the dust accumulation. In this embodiment, the image acquisition device includes a high-definition visible light camera, a 360° control gimbal, and a gimbal controller. The gimbal controller is used to control the gimbal to rotate the high-definition visible light camera in a 360° manner to acquire images of the photovoltaic panels within the field of view. The dust accumulation recognition device includes an embedded microprocessor, which uses a dust accumulation detection model to identify the dust accumulation gathered on the edge of the photovoltaic panel in the image, records the BeiDou navigation coordinates of the system and the gimbal attitude data at the time of detection, and calculates the location information of the dust accumulation to be cleaned.

[0044] The dust accumulation detection model includes an image preprocessing network, a backbone network, a neck network, and a head network. The image preprocessing network is used to scale, normalize, and sort the channels of the photovoltaic panel image to obtain the processed image. The backbone network is used to perform convolutional pooling operations on the processed image to obtain feature images at different levels. The neck network is used to fuse the feature images at different levels to obtain the fused image. The head network is used to predict the target based on the fused image to obtain the dust accumulation location.

[0045] In this embodiment, the dust accumulation detection model comprises four parts: an image preprocessing network, an input network, a backbone network, a neck network, and a head network. The main recognition process includes: the image preprocessing network scales the acquired photovoltaic panel image to 640*640 pixels and performs normalization (BN) and channel sorting operations to obtain a processed image; the processed image is then input into the backbone network via the input network for feature extraction. Through convolutional layers, pooling layers, and other operations, a series of feature maps of different sizes and levels are gradually obtained. Lower-level feature maps focus more on image texture and edge information, while higher-level feature maps focus more on semantic information; to obtain more comprehensive linguistic information, the neck network fuses the feature maps of different levels generated by the backbone network to obtain a fused image, thus acquiring richer and more diverse feature information; after feature fusion, the YOLOv5 in the head network treats each grid cell of the fused image as a candidate bounding box and performs target prediction on each bounding box to obtain the dust accumulation location.

[0046] The backbone network includes: a first convolutional unit and a second convolutional unit; the first convolutional unit includes one layer of FOUCS units; the second convolutional unit includes four groups of CBS units and DSConv units.

[0047] In this embodiment, the first convolutional unit is configured with one FOUCS layer, which has a 3×3 convolutional kernel and 32 output channels, transforming the input 3×640×640 image into a 32×320×320 feature map. The second convolutional unit is configured with four CBS units and DSConv units. The first group of the network has 64 output channels, transforming the input 32×320×320 image into a 64×160×160 feature map; the second group of the network has 128 output channels, transforming the input 64×160×160 image into a 128×80×80 feature map; the third group of the network has 256 output channels, transforming the input 128×80×80 image into a 256×40×40 feature map; and the fourth group of the network has 512 output channels, transforming the input 256×40×40 image into a 512×20×20 feature map. The output is fed into the Neck network.

[0048] The DSConv unit described above consists of Conv, Batch Normalization (BN), an attention mechanism module, and an activation function. To reduce the number of parameters and computational cost of the model, a new convolutional layer is used instead of the standard convolution. This new convolutional layer is obtained by linearly transforming the BN layer and the standard convolutional layer, as detailed below:

[0049]

[0050]

[0051]

[0052] W = W BN ·W conv (4)

[0053] b = W BN ·b conv +b BN (5)

[0054] Equation (1) is the standard form of the BN layer, x i For input data, The data is after BN transformation, γ and β are the parameters to be learned, μ is the mean of the input data, and σ 2 Let ∈ be the variance of the input data, and let ∈ be a very small number to prevent the denominator from being 0.

[0055] In order to be able to merge with convolution, equation (1) is rewritten in the form of convolution, thus deriving the format of equation (2), where For input, For output, This is the weight matrix. Let be the bias matrix. Equation (3) represents the result of combining convolution and BN, where For the output of the convolutional layer, f i，j Convolutional layer input, W conv b represents the weights of the convolutional layer. conv For the bias of the convolutional layer, W BN b represents the weights of the BN layer. BN This is the bias for the BN layer. W is the new convolutional kernel weight, and b is the new convolutional kernel bias.

[0056] In the DSConv unit described above, a CA attention mechanism module is added between the two convolution operations to capture contextual information in the input data. The CA module sets two independent pooling kernels (H, 1) and (1, W) to perform average pooling on each channel of the feature map in the horizontal and vertical directions. The calculation process is as follows:

[0057]

[0058]

[0059] Where, x c For the c-th input channel, This is the output of the c-th channel with height h. This is the output of the c-th channel with width w. To fully utilize the spatial and positional information obtained above, the CA module concatenates a pair of feature maps and uses convolution to compress the channels, reducing the number of channels from c to [the desired value]. c / r, the calculation formula is as follows:

[0060] f=δ(Conv([z h , z w ])) (8)

[0061] Where [·, ·] is the concatenation operation, Conv is the convolution operation, δ is the non-linear activation function, and f represents the intermediate feature map with fused height and width information.

[0062] Furthermore, a 1×1 convolution transformation is used to decompose f into two separate vectors fi. h f w Transform into F h and F w This yields a vector with the same number of channels as the input x. The calculation process is shown below:

[0063] g h =σ(F h (f h (9)

[0064] g w =σ(F w (f w (10)

[0065] Finally, the formula for calculating the output E of the CA attention module is as follows:

[0066]

[0067] The CA attention mechanism module utilizes attention maps in both horizontal and vertical directions to simulate the complexities of irregular shapes in height and width, assigning different weights to each feature point through attention vectors. Furthermore, after the final convolution operation, the H-Swish activation function is introduced to enhance the algorithm's ability to extract features from accumulated dust, improving the network's ability to focus on both local and global image information. The H-Swish activation function is shown below:

[0068]

[0069] Here, ReLU6 represents a rectified linear unit function with a maximum output value of 6, and x is the input to the activation function. H-Swish uses a piecewise function mechanism to process the input value in steps, which can greatly reduce the number of network memory accesses, thereby reducing latency costs.

[0070] The fixed-point dust-free cleaning module is used for dust cleaning based on the location of dust accumulation. The module includes a moving unit, a dust-raising unit, and a dust-removing unit. It also includes sealing strips and a suction chamber. The enclosed area formed by the sealing strips and suction chamber confines suspended dust particles, preventing dust from falling onto the cleaned photovoltaic modules and causing secondary pollution.

[0071] The moving unit is used to control the cleaning system to move to the dust accumulation location. The moving unit includes: a Beidou navigation device, a negative pressure moving carrier, and drive wheels. The working process of the moving unit includes: the Beidou navigation module provides position information on the photovoltaic panel surface based on the dust accumulation location, and then the drive wheels drive the negative pressure moving carrier to control the movement of the cleaning system on the photovoltaic panel surface.

[0072] The dust collection unit is used to complete the dust collection process, generating a dust-air mixture. The dust collection unit includes: a lightweight nylon roller brush, a DC motor, and a coupling. The working principle of the dust collection unit is as follows: Figure 3 As shown, the process includes: connecting a DC motor to a lightweight nylon roller brush via a coupling; the DC motor drives the lightweight nylon roller brush to rotate; the bristles of the lightweight nylon roller brush sweep across the surface of the photovoltaic panel, disrupting the dust deposition structure and lifting it up, thus completing the dust agitation process and generating a dust-air mixture.

[0073] The dust collection unit is used to extract dust-air mixtures, settle large dust particles, and filter out small dust particles. The dust collection unit includes a centrifugal fan, a filter screen, and a dust collection box. The working principle of the dust collection unit is as follows: Figure 4As shown, the process includes: a centrifugal fan guiding a dust-air mixture through an air intake duct into a dust collection box; larger dust particles settle at the bottom of the box under gravity, while smaller dust particles are filtered through a screen and separated from the air; the clean air after dust removal is then drawn away by the centrifugal fan and discharged from the cleaning system through a vent. The working principle of the dust collection box is as follows. Figure 5 As shown.

[0074] In summary, this embodiment first uses a high-definition visible light camera to acquire images of dust accumulation on the photovoltaic panel frame. Then, a dust accumulation detection model identifies the degree of dust accumulation and determines its distance. Based on the identified distance, the system's fixed-point movement range is controlled. The rotational speeds of the roller brush motor, centrifugal fan, and cleaning system drive wheels are controlled by the dust accumulation detection model's results. Finally, the centrifugal fan collects the dust from the photovoltaic panel surface, which has been dispersed by the roller brush, into a dust collection box. This invention allows for adjusting the cleaning system's dwell time based on the degree and distribution of dust accumulation on the frame to achieve targeted and intelligent cleaning. Compared to traditional photovoltaic cleaning systems, it enhances frame dust accumulation identification capabilities, resulting in higher cleaning efficiency and lower power consumption.

[0075] The embodiments described above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Any modifications and improvements made to the technical solutions of this application by those skilled in the art without departing from the spirit of this application shall fall within the protection scope defined by the claims of this application.

Claims

1. A visual feedback based cleaning system for dust accumulation on the frame of a photovoltaic panel, characterized in that, The application relates to a dust cleaning system for photovoltaic panels. The dust cleaning system comprises a frame dusting target recognition module and a fixed-point dustless cleaning module. The frame dusting target recognition module is used for recognizing a dusting position based on a dusting detection model. The fixed-point dustless cleaning module is used for dust cleaning based on the dusting position. The frame dusting target recognition module comprises an image acquisition device and a dusting recognition device. The image acquisition device is used for acquiring a photovoltaic panel image. The dusting recognition device is used for detecting the photovoltaic panel image based on the dusting detection model to obtain the dusting position. The dusting detection model comprises an image preprocessing network, a backbone network, a neck network and a head network. The image preprocessing network is used for performing scaling processing on the photovoltaic panel image and performing normalization and channel sorting processing to obtain a processed image. The backbone network is used for performing convolution and pooling operations on the processed image to obtain different-level feature images. The neck network is used for performing feature fusion on the different-level feature images to obtain a fused image. The head network is used for performing target prediction based on the fused image to obtain the dusting position. The backbone network comprises a first convolution unit and a second convolution unit. The first convolution unit comprises one layer of FOCUS units. The second convolution unit comprises four groups of CBS units and DSConv units. The first convolution unit is configured with one layer of FOCUS units, a 3*3 convolution kernel and an output channel of 32, and the input image of 3*640*640 is changed into a feature image of 32*320*320. The second convolution unit is configured with four groups of CBS units and DSConv units. Formula (1) is the standard form of BN layer, is the input data, is the data after BN conversion, , is the parameter to be learned, is the mean of the input data, is the variance of the input data, is a very small number to prevent the denominator from being 0; rewrite formula (1) into the form of convolution, thus deriving formula (2), wherein is the input, is the output, is the weight matrix, is the bias matrix; formula (3) represents the result after convolution and BN are combined, wherein, is the output of the convolution layer, is the input of the convolution layer, is the weight of the convolution layer, is the bias of the convolution layer, is the weight of the BN layer, is the bias of the BN layer; is the new convolution kernel weight, is the new convolution kernel bias; in the above DSConv unit, the CA attention mechanism module is added between the two convolution operations to capture the context information in the input data; two independent pooling kernels (H, 1) and (1, W) are set in the CA module to perform average pooling on each channel in the horizontal direction and the vertical direction of the feature map: in, For the input of the first c aisle, For height is h The c Channel output, For width is w The c The CA module concatenates a pair of feature maps and uses convolutional transformations to compress the channels, reducing the number of channels from [previous number]. c Become c / r : wherein, is a concatenation operation, The first group of networks is configured with an output channel of 64, and the input image of 32*320*320 is changed into a feature image of 64*160*160. is a convolution operation, is a nonlinear activation function, f denotes an intermediate feature map with fused height and width information; using a 1 x 1 convolutional transformation to f two separate vectors , transformed into and , thus obtaining a vector with the same number of channels as the input x , the calculation process is shown as follows: Output of the CA attention module E The calculation formula is: The second group of networks is configured with an output channel of 128, and the input image of 64*160*160 is changed into a feature image of 128*80*80. wherein, The third group of networks is configured with an output channel of 256, and the input image of 128*80*80 is changed into a feature image of 256*40*40. represents a rectified linear unit function with an output maximum of 6, x is an input of the activation function; the H-Swish adopts a piecewise function mechanism to perform step number on the input value; The fourth group of networks is configured with an output channel of 512, and the input image of 256*40*40 is changed into a feature image of 512*20*20, and the output enters the Neck network. The overall structure of the DSConv unit is composed of a Conv, a BN, an attention mechanism module and an activation function. The new convolution layer is obtained by linear transformation of the BN layer and the standard convolution layer. The CA attention mechanism module uses attention maps in horizontal and vertical directions to simulate the complex situation of irregular patterns in height and width, and different weights are assigned to each feature point through the attention vector. The H-Swish activation function is introduced after the last convolution operation to enhance the feature extraction ability of the algorithm and improve the attention ability of the network to local and global information. The ReLU6 is an activation function. The fixed-point dustless cleaning module comprises a moving unit, a dust raising unit and a dust removing unit. The moving unit is used to control the cleaning system to move to the dust accumulation position; The dust raising unit is used to complete the dust raising process and generate dust-air mixture gas; The dust removing unit is used to extract the dust-air mixture gas, settle large dust particles and filter small dust particles; The moving unit comprises a Beidou navigation device, a negative pressure moving carrier and a driving wheel; The working process of the moving unit comprises the following steps: providing position information of the dust accumulation position on the surface of the photovoltaic panel by the Beidou navigation device, and driving the negative pressure moving carrier to move on the surface of the photovoltaic panel by the driving wheel; The dust raising unit comprises a nylon light weight rolling brush, a direct current motor and a coupling; The working process of the dust raising unit comprises the following steps: connecting the direct current motor and the nylon light weight rolling brush through the coupling, rotating the nylon light weight rolling brush by the direct current motor, sweeping the brush wire of the nylon light weight rolling brush over the surface of the photovoltaic panel to destroy the dust deposition structure and raise the dust, completing the dust raising process and generating the dust-air mixture gas; The dust removing unit comprises a centrifugal fan, a filter screen and a dust collecting box; The working process of the dust removing unit comprises the following steps: guiding the dust-air mixture gas to enter the dust collecting box through the air guide channel by the centrifugal fan, depositing large dust particles at the bottom of the dust collecting box under the action of gravity, filtering small dust particles through the filter screen and separating the small dust particles from the air, extracting clean air after dust removal by the centrifugal fan and discharging the clean air out of the cleaning system through the ventilation opening; The fixed-point dust-free cleaning module further comprises sealing wool and a dust suction cavity; The closed area formed by the sealing wool and the dust suction cavity limits the suspended dust particles therein and prevents the dust from falling on the cleaned photovoltaic module to cause secondary pollution.

Images