A mechanical arm-based intelligent recycling and identification system and method for decommissioned photovoltaic modules
By employing multi-view image acquisition, environmental adaptive preprocessing, and an improved YOLOX algorithm, combined with a binocular vision system, the problems of robustness in identification and positioning accuracy in solar panel recycling were solved, achieving efficient and accurate identification of recycled photovoltaic modules.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2026-02-05
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies, the target identification robustness during the solar panel recycling process is poor. It is affected by changes in light intensity, weather interference, and surface stains, resulting in a limited identification range and making it difficult to achieve efficient and automated recycling.
By employing a multi-view image acquisition unit, an environment-adaptive preprocessing unit, a fusion recognition unit, and a 3D positioning unit, combined with an improved YOLOX algorithm and a binocular vision system, highly robust recognition and accurate positioning of photovoltaic modules can be achieved.
It achieves high-precision identification of solar panels in complex outdoor environments, with an identification accuracy of over 95%, a damage level identification error of less than level 1, and a three-dimensional positioning error controlled within ±2mm, reducing system complexity and improving recycling efficiency.
Smart Images
Figure CN122157230A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of retired photovoltaic module recycling and machine vision, specifically relating to an intelligent recycling and identification system and method for retired photovoltaic modules based on a robotic arm. Background Technology
[0002] With the rapid development of the solar photovoltaic industry, the amount of scrapped solar panels is increasing year by year, highlighting the urgency and importance of solar panel recycling. Mobile solar panel recycling systems have become an important form of equipment for solar panel recycling due to their advantages of flexibility and on-site recycling capabilities. The target identification module, as the "eyes" of the recycling system, directly determines the level of automation and reliability of the recycling operation through its identification accuracy and efficiency.
[0003] In existing technologies, target recognition during solar panel recycling often employs traditional machine vision algorithms, such as template matching or simple feature extraction. These methods have the following drawbacks: First, they have poor robustness. In outdoor recycling scenarios, they are easily affected by factors such as changes in light intensity, weather interference (such as rain and dust), and stains or damage on the surface of the solar panels, leading to recognition errors or missed recognition. Second, they have a limited recognition range. Traditional single-camera systems have a fixed field of view, making it difficult to cover solar panel targets in different locations and with different orientations around the recycling system. This results in the system needing to frequently adjust its position to find targets, reducing recycling efficiency.
[0004] To address the aforementioned issues, there is an urgent need for a target recognition module that is adaptable to mobile outdoor recycling scenarios, possesses high robustness, a wide field of view, and integrated recognition and positioning functions, in order to improve the automation capabilities of solar panel recycling systems. Summary of the Invention
[0005] To address the aforementioned problems, the present invention aims to provide an intelligent recycling and identification system and method for retired photovoltaic modules based on a robotic arm.
[0006] The specific technical solution for achieving the objective of this invention is as follows:
[0007] A robotic arm-based intelligent recycling and identification system for decommissioned photovoltaic modules includes a multi-view image acquisition unit, an environmental adaptive preprocessing unit, a fusion identification unit, a three-dimensional positioning unit, and a data synchronization unit.
[0008] The multi-view image acquisition unit is used to acquire close-range target images from multiple perspectives within the working range of the robotic arm;
[0009] The environment adaptive preprocessing unit is used to preprocess the target image acquired by the multi-view image acquisition unit according to the environmental data;
[0010] The fusion recognition unit is used to extract and recognize features from the preprocessed image and output the target classification result.
[0011] The three-dimensional positioning unit performs three-dimensional positioning based on data from the multi-view image acquisition unit and the fusion recognition unit.
[0012] The data synchronization unit integrates the recognition results output by the fusion recognition unit with the coordinate information output by the three-dimensional positioning unit, generates a standardized data frame, and then transmits it to the control end of the robotic arm.
[0013] Furthermore, the multi-view image acquisition unit includes a main camera and a follow-up camera set at multiple angles;
[0014] Used for acquiring panoramic and long-distance target images from multiple angles;
[0015] Each of the cameras is equipped with an adaptive fill light, which automatically adjusts the fill light power according to the ambient light intensity.
[0016] Furthermore, the environment adaptive preprocessing unit uses a light intensity sensor to collect ambient light intensity data at certain time intervals, and automatically switches the corresponding preprocessing parameter template according to the data to adapt to low light environment, normal light environment and strong light environment, respectively, and preprocesses the collected target image.
[0017] The preprocessing parameter template includes Retinex algorithm scale parameters, filter window size, and edge enhancement threshold.
[0018] Furthermore, the preprocessing procedure for the acquired target image by the environment adaptive preprocessing unit includes:
[0019] First, the original RGB image is converted to HSV color space through illumination equalization processing based on the Retinex algorithm. The multi-scale Retinex algorithm is used only for the luminance channel to separate the illumination component and the reflection component and make adaptive adjustments. Finally, the grayscale stretching is fused with the original H and S channels to solve the problem of local dark or bright images caused by uneven outdoor lighting.
[0020] Next, a denoising process based on a combination of median filtering and Gaussian filtering is performed, using a composite strategy of "median first, then Gaussian" to achieve smooth denoising while preserving image edge details.
[0021] Finally, contour extraction is performed based on edge enhancement algorithms. Combining the Sobel and Laplacian operators, the gradient magnitude map is first calculated in the horizontal and vertical directions. Then, the Laplacian operator enhances the gradient magnitude map a second time. The edge region is separated from the background by adaptive threshold segmentation, generating a clear contour mask map, which provides clear target contour information for subsequent feature extraction.
[0022] Furthermore, the fusion recognition unit includes a feature extraction subunit and a classification recognition subunit;
[0023] The feature extraction subunit is constructed based on the improved YOLOX algorithm. The improved YOLOX algorithm embeds the CBAM attention mechanism module in the C3 module of the YOLO backbone network. This module is located after each residual block and achieves dual weighting through the channel attention submodule and the spatial attention submodule in sequence.
[0024] The channel attention submodule adopts a parallel structure of "global average pooling + global max pooling", and learns channel weights by combining two fully connected layers to enhance the response of key feature channels; the spatial attention submodule performs pooling and 1×1 convolution on the channel-weighted feature map to learn spatial weights, highlight the features of the target region, and effectively extract the shape features, texture features and damage features of the solar panel in the preprocessed image.
[0025] The classification and recognition subunit is trained on a dataset constructed using images of solar panels. It uses image features extracted by the feature extraction subunit to classify the data and obtain classification results for "solar panels" and "non-solar panels" and their corresponding damage levels.
[0026] Furthermore, the fusion recognition unit sequentially includes:
[0027] Input layer: In the photovoltaic module inspection scenario, the preprocessed images are first standardized, including normalizing pixel values and adjusting image size, so that the data meets the requirements of model training and inference. To adapt to the characteristics of diverse photovoltaic module models and different installation angles on the production line, the input layer adopts a multi-scale training strategy to dynamically adjust the image resolution, allowing the model to learn target features at different scales.
[0028] Backbone layer: The backbone layer adopts the CSPDarknet structure, which divides the feature extraction process into multiple stages through the idea of cross-stage local networks. Each stage consists of residual blocks and cross-stage connections.
[0029] In photovoltaic module testing, the backbone network extracts underlying features such as the frame, cell arrangement, and junction box of the photovoltaic module step by step through multi-layer convolution operations.
[0030] Feature fusion layer: The feature fusion layer achieves the fusion of features at different scales through a path aggregation network;
[0031] The feature fusion layer is based on a PAN structure and adopts a bidirectional feature fusion strategy of bottom-up and top-down to enhance the semantic expression of low-level features and make up for the insufficient positioning accuracy of high-level features.
[0032] Prediction Layer: In photovoltaic module inspection, the prediction layer analyzes each preset anchor box through convolution operations to predict the probability that the target belongs to different photovoltaic module categories, and outputs the position coordinates of the target bounding box.
[0033] To adapt to the differences in photovoltaic module shapes, the prediction layer adopts a dynamic anchor frame mechanism, which automatically adjusts the anchor frame size and proportion based on training data, so that the model can accurately locate photovoltaic module targets of various shapes such as photovoltaic panels and solar cells of different specifications.
[0034] Finally, the prediction layer removes duplicate prediction boxes using a non-maximum suppression algorithm, outputting the optimal detection result.
[0035] Furthermore, the three-dimensional positioning unit performs preprocessing and stereo matching on the left and right eye images acquired by the binocular vision system in the multi-view image acquisition unit to obtain the preliminary three-dimensional coordinates of the target;
[0036] The depth information collected by the lidar is used to correct the target's initial three-dimensional coordinates;
[0037] Perform coordinate transformation to change the target's coordinates from the image coordinate system to the robot arm coordinate system.
[0038] This invention also provides a method for intelligent recycling and identification of decommissioned photovoltaic modules based on a robotic arm, comprising the following steps:
[0039] Acquire multi-view images of close-range targets within the robotic arm's operating range;
[0040] Preprocess the target images acquired by the multi-view image acquisition unit based on environmental data;
[0041] The preprocessed image undergoes feature extraction and recognition, and the target classification result is output.
[0042] 3D localization based on the identified target;
[0043] The fusion recognition results are associated with the coordinate information of the 3D positioning, and a standardized data frame is generated and transmitted to the control end of the robotic arm.
[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0045] (1) Robust and adaptable to complex outdoor scenes: Wide coverage is achieved through multi-view image acquisition unit, combined with light equalization and composite noise reduction processing of environment adaptive preprocessing unit, effectively resisting interference from light changes, dust, stains, etc.; Improved YOLOX algorithm adds CBAM attention mechanism to enhance target feature extraction, and the recognition accuracy can reach more than 95%, and the damage level recognition error is less than level 1.
[0046] (2) Integrated identification and positioning to reduce system complexity: The three-dimensional positioning unit integrates binocular vision and lidar positioning, and works with the coordinate transformation module to realize real-time conversion from the image coordinate system to the robot arm base coordinate system. No additional independent positioning equipment is required, and the coordinate transformation error is controlled within ±2mm, providing direct data support for the robot arm to grasp accurately.
[0047] (3) Reliable data transmission and good coordination: The data synchronization unit adopts timestamp synchronization and CRC-16 verification mechanism to ensure accurate correlation and transmission integrity between the identification results and coordinate information. The retransmission mechanism further improves reliability and avoids robot arm operation errors caused by data asynchrony or loss.
[0048] The present invention will be further described below with reference to specific embodiments. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the architecture of the intelligent recycling and identification system and method for retired photovoltaic modules based on a robotic arm, as described in this invention.
[0050] Figure 2 This is a schematic diagram of the processing flow of the improved YOLOX algorithm of the present invention.
[0051] Figure 3 This is a schematic diagram of the improved YOLOX algorithm of the present invention. Detailed Implementation
[0052] Example
[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0054] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.
[0055] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.
[0056] Combination Figure 1 A robotic arm-based intelligent recycling and identification system for decommissioned photovoltaic modules includes a multi-view image acquisition unit, an environmental adaptive preprocessing unit, a fusion identification unit, a three-dimensional positioning unit, and a data synchronization unit.
[0057] The multi-view image acquisition unit is used to acquire close-range target images from multiple perspectives within the working range of the robotic arm;
[0058] The environment adaptive preprocessing unit is used to preprocess the target image acquired by the multi-view image acquisition unit according to the environmental data;
[0059] The fusion recognition unit is used to extract and recognize features from the preprocessed image and output the target classification result.
[0060] The three-dimensional positioning unit performs three-dimensional positioning based on data from the multi-view image acquisition unit and the fusion recognition unit.
[0061] The data synchronization unit integrates the recognition results output by the fusion recognition unit with the coordinate information output by the three-dimensional positioning unit, generates a standardized data frame, and then transmits it to the control end of the robotic arm.
[0062] The multi-view image acquisition unit includes a main camera and a follow-up camera set at multiple angles;
[0063] Used for acquiring panoramic and long-distance target images from multiple angles;
[0064] Each of the aforementioned cameras is equipped with an adaptive fill light, which is used to automatically adjust the fill light power according to the ambient light intensity;
[0065] The multi-view image acquisition unit in this embodiment includes a main camera fixed to the front of the mobile chassis, side cameras symmetrically arranged on both sides of the mobile chassis, and a follow-up camera mounted on the end of the robotic arm. The main camera uses a wide-angle lens to acquire panoramic images within a 120° range in front. The side cameras use telephoto lenses to acquire images of distant targets within a 50° range on both sides. The follow-up camera is linked to the movement trajectory of the robotic arm and adjusts the shooting angle in real time to acquire detailed images of close-range targets within the working range of the robotic arm. Each camera is equipped with an adaptive fill light module, which automatically adjusts the fill light power according to the ambient light intensity.
[0066] The environment adaptive preprocessing unit uses a light intensity sensor to collect ambient light intensity data at certain time intervals, and automatically switches the corresponding preprocessing parameter template according to the data to adapt to low light environment, normal light environment and strong light environment, and preprocesses the collected target image.
[0067] The preprocessing parameter template includes Retinex algorithm scale parameters, filter window size, and edge enhancement threshold;
[0068] In this embodiment, the environment adaptive preprocessing unit is electrically connected to the multi-view image acquisition unit, which is a key link connecting image acquisition and recognition. The built-in BH1750 light intensity sensor collects ambient light intensity data at 100ms intervals and automatically switches the corresponding preprocessing parameter template according to the data to adapt to low light environment, normal lighting environment and strong light environment respectively. The template includes key parameters such as Retinex algorithm scale parameters, filter window size, edge enhancement threshold, etc., to ensure the optimal processing effect in different environments.
[0069] In addition, the preprocessing procedure for the acquired target image by the environment adaptive preprocessing unit includes:
[0070] First, the original RGB image is converted to HSV color space through illumination equalization processing based on the Retinex algorithm. The multi-scale Retinex algorithm (setting three Gaussian kernels with σ=15, 80, 250) is used only for the luminance channel to separate the illumination component and the reflection component and make adaptive adjustments. Finally, the grayscale stretching is used to fuse with the original H and S channels to solve the problem of local dark or bright images caused by uneven outdoor lighting.
[0071] Next, a denoising process based on a combination of median filtering and Gaussian filtering is performed. A composite strategy of "median first, then Gaussian" is adopted. The median filter with a 3×3 window removes salt-and-pepper noise, and the Gaussian filter with a 5×5 window and a standard deviation σ=1.2 suppresses Gaussian noise, achieving smooth denoising while preserving image edge details.
[0072] Finally, contour extraction is performed based on edge enhancement algorithms. Combining the Sobel and Laplacian operators, the gradient magnitude map is first calculated in the horizontal and vertical directions. Then, the Laplacian operator enhances the gradient magnitude map a second time. The edge region is separated from the background by adaptive threshold segmentation, generating a clear contour mask map, which provides clear target contour information for subsequent feature extraction.
[0073] The fusion recognition unit, which is electrically connected to the environment adaptive preprocessing unit, is the core of the target recognition module and includes a feature extraction subunit and a classification recognition subunit.
[0074] The feature extraction subunit is constructed based on the improved YOLOX algorithm. The improved YOLOX algorithm embeds the CBAM attention mechanism module in the C3 module of the YOLO backbone network. This module is located after each residual block and achieves dual weighting through the channel attention submodule and the spatial attention submodule in sequence.
[0075] The channel attention submodule adopts a parallel structure of "global average pooling + global max pooling", and combines two fully connected layers to learn channel weights to enhance the response of key feature channels; the spatial attention submodule performs pooling and concatenation and 1×1 convolution on the channel-weighted feature map to learn spatial weights, highlight the features of the target region, and effectively extract the shape features (rectangular outline, aspect ratio, etc.), texture features (battery cell arrangement, etc.) and damage features (cracks, holes, etc.) of the solar panel in the preprocessed image.
[0076] The classification and recognition subunit is trained based on a dataset constructed using images of solar panels, and classifies the images using the image features extracted by the feature extraction subunit to obtain classification results for "solar panels" and "non-solar panels" and their corresponding damage levels.
[0077] In this embodiment, a solar panel image dataset containing over 100,000 samples is first constructed, covering different lighting conditions, damage levels (0-4), placement postures, and backgrounds. After data augmentation processing such as random flipping, rotation, and scaling, the dataset is expanded to over 300,000 samples and divided into training, validation, and test sets in a 7:2:1 ratio. The YOLOX model parameters are initialized using transfer learning, and the first 10 layers of the backbone network are frozen for iterative training. Hyperparameters are adjusted using the validation set, and the final output is the classification results of "solar panel" and "non-solar panel" and their corresponding damage levels. The damage level identification error is less than 1 level.
[0078] Combination Figure 3The fusion recognition unit addresses the issue of insufficient feature extraction in traditional target detection models for photovoltaic module identification and localization. Drawing inspiration from YOLOv7, it constructs a novel backbone network integrating ELAN and MP structures. This architecture effectively enhances the model's feature capture and representation capabilities for photovoltaic module targets through multi-branch parallel feature extraction and cross-layer feature reuse mechanisms, comprising:
[0079] Input layer: In the photovoltaic module inspection scenario, the preprocessed images are first standardized, including normalizing pixel values and adjusting image size, so that the data meets the requirements of model training and inference. To adapt to the characteristics of diverse photovoltaic module models and different installation angles on the production line, the input layer adopts a multi-scale training strategy to dynamically adjust the image resolution, allowing the model to learn target features at different scales.
[0080] Backbone Layer: The backbone layer adopts the CSPDarknet structure, which divides the feature extraction process into multiple stages through the concept of cross-stage local networks. Each stage consists of residual blocks and cross-stage connections. This design reduces computational cost and enhances the gradient propagation efficiency of the network, avoiding the gradient vanishing problem during deep network training. In photovoltaic module detection, the backbone network extracts low-level features such as the frame, cell arrangement, and junction box of the photovoltaic module step by step through multi-layer convolutional operations.
[0081] In photovoltaic module testing, the backbone network extracts underlying features such as the frame, cell arrangement, and junction box of the photovoltaic module step by step through multi-layer convolution operations.
[0082] Feature fusion layer: The feature fusion layer achieves the fusion of features at different scales through a path aggregation network;
[0083] The feature fusion layer is based on a PAN structure and adopts a bidirectional feature fusion strategy of bottom-up and top-down to enhance the semantic expression of low-level features and make up for the insufficient positioning accuracy of high-level features.
[0084] Prediction Layer: In photovoltaic module inspection, the prediction layer analyzes each preset anchor box through convolution operations to predict the probability that the target belongs to different photovoltaic module categories, and outputs the position coordinates of the target bounding box.
[0085] To adapt to the differences in photovoltaic module shapes, the prediction layer adopts a dynamic anchor frame mechanism, which automatically adjusts the anchor frame size and proportion based on training data, so that the model can accurately locate photovoltaic module targets of various shapes such as photovoltaic panels and solar cells of different specifications.
[0086] Finally, the prediction layer removes duplicate prediction boxes using a non-maximum suppression algorithm, outputting the optimal detection result. To further enhance the identification and localization capabilities of photovoltaic module targets, the YOLOX model is improved by introducing CBAM, employing depthwise separable convolution to optimize the network structure, and combining it with a novel backbone network. The specific process is as follows: Figure 2 As shown.
[0087] The three-dimensional positioning unit preprocesses and performs stereo matching on the left and right eye images acquired by the binocular vision system in the multi-view image acquisition unit to obtain the preliminary three-dimensional coordinates of the target.
[0088] The depth information collected by the lidar is used to correct the target's initial three-dimensional coordinates;
[0089] Perform coordinate transformation to change the target's coordinates from the image coordinate system to the robot arm coordinate system.
[0090] In this implementation, a binocular vision system is formed by the main camera and one of the side cameras to calculate the initial three-dimensional coordinates of the target; the lidar is fixed on the top of the mobile chassis to collect the depth information of the target and correct the binocular vision positioning results; the three-dimensional positioning unit has a built-in coordinate transformation module to convert the target's three-dimensional coordinates from the image coordinate system to the robot arm's base coordinate system.
[0091] Specifically, the hardware configuration in this embodiment includes: a binocular vision system consisting of a main camera and a left or right side camera, with the optical axes of the two cameras parallel and the baseline distance fixed at 80cm to ensure sufficient parallax calculation accuracy;
[0092] 3D coordinate calculation: First, stereo matching is performed on the preprocessed left and right eye images. The SGBM algorithm is used, the matching window size is set to 9×9, and the disparity range is 0-128. By calculating the disparity of the corresponding pixels, combined with the intrinsic and extrinsic parameters of the binocular system, the preliminary 3D coordinates of the target are calculated using the principle of triangulation.
[0093] Functional localization: Primarily collects depth information of the target and corrects the binocular vision localization results. LiDAR is unaffected by lighting conditions and can accurately measure target distance even in strong light, low light, or when the target surface texture is unclear.
[0094] Fusion strategy: A weighted fusion algorithm is adopted, with the binocular vision positioning result having a weight of 0.6 and the lidar positioning result having a weight of 0.4; when the disparity matching confidence of the binocular vision positioning is lower than 0.8, the lidar weight is automatically increased to 0.7 to ensure the stability of the positioning result; the three-dimensional coordinates of the target are output after fusion.
[0095] Coordinate transformation:
[0096] Transformation Algorithm: A homogeneous coordinate transformation algorithm is adopted, which realizes the coordinate transformation through a 4×4 homogeneous transformation matrix. The core formula is: ,in The coordinates are in the robot arm's base coordinate system. Let be the transformation matrix of the camera relative to the robot arm's base coordinate system. This is the transformation matrix from the image coordinate system to the camera coordinate system;
[0097] Preset calibration parameters: Preset the calibration parameters of each camera and LiDAR relative to the robot arm's base coordinate system, including position coordinates and attitude angles. These parameters are obtained through calibration using a laser tracker with an accuracy of ±0.1mm.
[0098] Real-time conversion process: First, the 3D coordinates in the image coordinate system obtained by fusion positioning are converted into the camera coordinate system. Then, according to the transformation matrix of the camera relative to the robot arm base coordinate system, the coordinates in the camera coordinate system are converted into the coordinates in the robot arm base coordinate system. During the conversion process, the motion error of the robot arm is compensated in real time. The current joint angle is obtained through the robot arm controller, and the transformation matrix parameters are corrected to ensure the conversion accuracy.
[0099] In addition, this embodiment also includes a data synchronization unit, which is electrically connected to the control systems of the fusion recognition unit, the three-dimensional positioning unit, and the retrieval system, respectively. It adopts a timestamp synchronization mechanism to associate the recognition results output by the fusion recognition unit with the coordinate information output by the three-dimensional positioning unit, generate standardized data frames, and transmit them to the control system to provide data support for the robotic arm's grasping action.
[0100] This invention also provides a method for intelligent recycling and identification of decommissioned photovoltaic modules based on a robotic arm, comprising the following steps:
[0101] Acquire multi-view images of close-range targets within the robotic arm's operating range;
[0102] Preprocess the target images acquired by the multi-view image acquisition unit based on environmental data;
[0103] The preprocessed image undergoes feature extraction and recognition, and the target classification result is output.
[0104] 3D localization based on the identified target;
[0105] The fusion recognition results are associated with the coordinate information of the 3D positioning, and a standardized data frame is generated and transmitted to the control end of the robotic arm.
[0106] This solution also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.
[0107] This solution also provides a computer-storable medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above method.
[0108] The embodiments described above are merely one implementation method of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A smart recycling and identification system for decommissioned photovoltaic modules based on a robotic arm, characterized in that, It includes a multi-view image acquisition unit, an environment adaptive preprocessing unit, a fusion recognition unit, a 3D positioning unit, and a data synchronization unit; The multi-view image acquisition unit is used to acquire close-range target images from multiple perspectives within the working range of the robotic arm; The environment adaptive preprocessing unit is used to preprocess the target image acquired by the multi-view image acquisition unit according to the environmental data; The fusion recognition unit is used to extract and recognize features from the preprocessed image and output the target classification result. The three-dimensional positioning unit performs three-dimensional positioning based on data from the multi-view image acquisition unit and the fusion recognition unit. The data synchronization unit integrates the recognition results output by the fusion recognition unit with the coordinate information output by the three-dimensional positioning unit, generates a standardized data frame, and then transmits it to the control end of the robotic arm.
2. The intelligent recycling and identification system for decommissioned photovoltaic modules based on a robotic arm as described in claim 1, characterized in that, The multi-view image acquisition unit includes a main camera and a follow-up camera set at multiple angles; Used for acquiring panoramic and long-distance target images from multiple angles; Each of the cameras is equipped with an adaptive fill light, which automatically adjusts the fill light power according to the ambient light intensity.
3. The intelligent recycling and identification system for decommissioned photovoltaic modules based on a robotic arm according to claim 1, characterized in that, The environment adaptive preprocessing unit uses a light intensity sensor to collect ambient light intensity data at certain time intervals, and automatically switches the corresponding preprocessing parameter template according to the data to adapt to low light environment, normal light environment and strong light environment, and preprocesses the collected target image. The preprocessing parameter template includes Retinex algorithm scale parameters, filter window size, and edge enhancement threshold.
4. The intelligent recycling and identification system for decommissioned photovoltaic modules based on a robotic arm according to claim 3, characterized in that, The preprocessing procedure of the environment adaptive preprocessing unit for the acquired target image includes: First, the original RGB image is converted to HSV color space through illumination equalization processing based on the Retinex algorithm. The multi-scale Retinex algorithm is used only for the luminance channel to separate the illumination component and the reflection component and make adaptive adjustments. Finally, the grayscale stretching is fused with the original H and S channels to solve the problem of local dark or bright images caused by uneven outdoor lighting. Next, a denoising process based on a combination of median filtering and Gaussian filtering is performed, using a composite strategy of "median first, then Gaussian" to achieve smooth denoising while preserving image edge details; Finally, contour extraction is performed based on edge enhancement algorithms. Combining the Sobel and Laplacian operators, the gradient magnitude map is first calculated in the horizontal and vertical directions. Then, the Laplacian operator enhances the gradient magnitude map a second time. The edge region is separated from the background by adaptive threshold segmentation, generating a clear contour mask map, which provides clear target contour information for subsequent feature extraction.
5. The intelligent recycling and identification system for decommissioned photovoltaic modules based on a robotic arm according to claim 1, characterized in that, The fusion recognition unit includes a feature extraction subunit and a classification recognition subunit; The feature extraction subunit is constructed based on the improved YOLOX algorithm. The improved YOLOX algorithm embeds the CBAM attention mechanism module in the C3 module of the YOLO backbone network. This module is located after each residual block and achieves dual weighting through the channel attention submodule and the spatial attention submodule in sequence. The channel attention submodule adopts a parallel structure of "global average pooling + global max pooling" and combines two fully connected layers to learn channel weights, thereby strengthening the response of key feature channels. The spatial attention submodule performs pooling and 1×1 convolution on the channel-weighted feature map to learn spatial weights, highlight the features of the target region, and effectively extract the shape features, texture features and damage features of the solar panel in the preprocessed image. The classification and recognition subunit is trained on a dataset constructed using images of solar panels. It uses image features extracted by the feature extraction subunit to classify the data and obtain classification results for "solar panels" and "non-solar panels" and their corresponding damage levels.
6. The intelligent recycling and identification system for decommissioned photovoltaic modules based on a robotic arm according to claim 5, characterized in that, The fusion recognition unit includes, in sequence: Input layer: In the photovoltaic module inspection scenario, the preprocessed images are first standardized, including normalizing pixel values and adjusting image size, so that the data meets the requirements of model training and inference. To adapt to the characteristics of diverse photovoltaic module models and different installation angles on the production line, the input layer adopts a multi-scale training strategy to dynamically adjust the image resolution, allowing the model to learn target features at different scales. Backbone layer: The backbone layer adopts the CSPDarknet structure, which divides the feature extraction process into multiple stages through the idea of cross-stage local networks. Each stage consists of residual blocks and cross-stage connections. In photovoltaic module testing, the backbone network extracts underlying features such as the frame, cell arrangement, and junction box of the photovoltaic module step by step through multi-layer convolution operations. Feature fusion layer: The feature fusion layer achieves the fusion of features at different scales through a path aggregation network; The feature fusion layer is based on the PAN structure and adopts a bidirectional feature fusion strategy of bottom-up and top-down to enhance the semantic expression of low-level features and make up for the insufficient positioning accuracy of high-level features. Prediction Layer: In photovoltaic module inspection, the prediction layer analyzes each preset anchor box through convolution operations to predict the probability that the target belongs to different photovoltaic module categories, and outputs the position coordinates of the target bounding box. To adapt to the differences in photovoltaic module shapes, the prediction layer adopts a dynamic anchor frame mechanism, which automatically adjusts the anchor frame size and proportion based on training data, so that the model can accurately locate photovoltaic module targets of various shapes such as photovoltaic panels and solar cells of different specifications. Finally, the prediction layer removes duplicate prediction boxes using a non-maximum suppression algorithm, outputting the optimal detection result.
7. The intelligent recycling and identification system for decommissioned photovoltaic modules based on a robotic arm according to claim 1, characterized in that, The three-dimensional positioning unit preprocesses and performs stereo matching on the left and right eye images acquired by the binocular vision system in the multi-view image acquisition unit to obtain the preliminary three-dimensional coordinates of the target. The depth information collected by the lidar is used to correct the target's initial three-dimensional coordinates; Perform coordinate transformation to change the target's coordinates from the image coordinate system to the robot arm coordinate system.
8. A method for intelligent recycling and identification of decommissioned photovoltaic modules based on a robotic arm, characterized in that, Includes the following steps: Acquire multi-view images of close-range targets within the robotic arm's operating range; Preprocess the target images acquired by the multi-view image acquisition unit based on environmental data; The preprocessed image undergoes feature extraction and recognition, and the target classification result is output. 3D localization based on the identified target; The fusion recognition results are associated with the coordinate information of the 3D positioning, and a standardized data frame is generated and transmitted to the control end of the robotic arm.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method of claim 8.
10. A computer-storable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 8.