Method for repairing defects of a greenhouse

Abstract

The application discloses a greenhouse defect repairing method, and relates to the technical field of agricultural implementation and maintenance, and comprises the following steps: S1, constructing a three-dimensional digital map of the greenhouse, generating a predetermined moving path through a path planning algorithm, controlling a moving platform to move in the greenhouse according to the predetermined route, and continuously collecting greenhouse film images through a visual system; S2, identifying a damaged area based on the greenhouse film images and determining the position coordinates in the greenhouse space; S3, analyzing the image features of the damaged area and generating a repair instruction; S4, controlling a mechanical arm to carry a repair head to the position coordinates, executing the repair instruction, and completing the repairing operation on the damaged area; according to the application, a deep learning semantic segmentation model and a three-dimensional visual positioning technology are used, the system can automatically identify various damages and calculate the accurate physical position and geometric features of the damages like an expert; and based on a preset repair strategy rule library, the system can objectively and consistently intelligently match the optimal repairing strategy and process parameters for different damages.

Description

Technical Field

[0002] This invention relates to the field of agricultural implementation and maintenance technology, specifically to a method for repairing defects in greenhouses. Background Technology

[0003] Greenhouse films are easily damaged during use due to wind, aging, hail, etc. If they are not repaired in time, they will affect the heat preservation and moisture retention effect. At present, repair mainly relies on manual inspection and manual repair, which is inefficient, dangerous at height, and the repair quality varies from person to person. Existing technologies include several greenhouse repair devices, such as the intelligent automatic greenhouse film repair device disclosed in patent CN212386052U, which repairs damage using hot-melt plastic. However, these devices primarily focus on improving the repair actuator itself or providing repair frames and specialized tapes for easy manual operation. They generally suffer from low automation, lack of proactive damage detection capabilities, and reliance on manual experience for repair strategies. They cannot achieve full-process autonomy from "problem detection" to "problem resolution," and lack the "intelligent" characteristics of making intelligent decisions and implementing closed-loop quality control based on damage characteristics. Therefore, we propose a method for repairing greenhouse defects. Summary of the Invention

[0004] To address the aforementioned technical problems, a method for repairing defects in greenhouses is provided. This technical solution solves the problems of lacking the ability to proactively detect damage, relying on manual experience for repair strategies, and lacking the "intelligent" characteristics of making intelligent decisions and implementing closed-loop quality control based on damage characteristics.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for repairing defects in a greenhouse, comprising: S1. Construct a 3D digital map of the greenhouse, generate a predetermined movement path through a path planning algorithm, control the mobile platform to move inside the greenhouse along the predetermined route, and continuously acquire images of the greenhouse film through the onboard vision system. S2. Based on the greenhouse film image, identify the damaged area and determine its location coordinates within the greenhouse space; S3. Analyze the image features of the damaged area and generate repair instructions; S4. Control the robotic arm to carry the repair head to the position coordinates, execute the repair command, and complete the repair operation on the damaged area; S5. Complete the repair operation, monitor the repaired area, and determine whether it is qualified. If it is not qualified, return to the decision process of S3, generate a new repair instruction and execute it; if it is qualified, record the repair result.

[0006] Preferably, step S1 specifically includes: The mobile platform is equipped with LiDAR, visual sensors and integrated navigation system to perform mapping operations. Based on simultaneous positioning and map building algorithms, it integrates multi-source data to generate a three-dimensional digital map of the greenhouse. The multi-source data consists of three-dimensional geometric point cloud data obtained by LiDAR scanning the environment, textured images and corresponding depth information acquired by a depth camera, and timestamps and spatial pose labels added to the acquired data by the integrated navigation system. Based on multi-source data, a three-dimensional digital map containing the greenhouse film surface, support frame and internal facility structure is finally constructed by fusion and global optimization through synchronous positioning and mapping algorithms. Based on a 3D digital map, a predetermined movement path covering the target area is calculated using a path planning algorithm.

[0007] Preferably, the predetermined movement path specifically includes: During the movement along the predetermined path, the real-time positioning and path-following control of the mobile platform are achieved by fusing data from wheel encoders and inertial measurement units and matching them with a three-dimensional digital map. The vision system uses an equidistant triggering mechanism to acquire images and adaptively adjusts the camera gimbal angle according to the platform's real-time pose to ensure continuous and vertical alignment with the greenhouse film surface; the acquired images are synchronously encapsulated with the platform's spatial coordinates and camera pose data at the time of acquisition.

[0008] Preferably, step S2 specifically includes the following steps: S21. Preprocessing is performed on the encapsulated acquired image by illumination normalization, Gaussian filtering, and image registration when using multispectral images. S22. Input the preprocessed image into a pre-trained deep learning semantic segmentation model and output the pixel set of the damaged area; S23. Perform feature quantization on the segmented damaged area, calculate its geometric features and extract its centroid pixel coordinates as representative image positions. S24. Using the pre-calibrated camera intrinsic parameter matrix, back-project the original centroid pixel coordinates and normalize them to a two-dimensional plane in the camera coordinate system. S25. Combining the depth value corresponding to the damage point measured by the depth sensor, the two-dimensional plane is converted into a three-dimensional spatial point in the camera coordinate system. ; S26, Based on three-dimensional spatial points Subtract the extrinsic matrix In Perform a reverse translation operation to obtain a transition coordinate. The extrinsic parameter matrix is ​​provided by the synchronous positioning and mapping system at the time of image acquisition. For rotation matrix, It is a translation vector; S27, the transition coordinates Left-multiplying the transpose of the rotation matrix R Perform the reverse rotation operation and output the three-dimensional coordinates. ; S28, Output three-dimensional coordinates The coordinate system is aligned and verified with the 3D digital map of the greenhouse, and the 3D position coordinates of the damaged area in physical space are output.

[0009] Preferably, the deep learning semantic segmentation model specifically includes: The preprocessed greenhouse film image is input into the loaded semantic segmentation model. The model performs forward propagation based on an encoder-decoder architecture, extracting features through the encoder and restoring spatial details through the decoder, and outputting an initial segmentation mask corresponding to the spatial size of the input image, where each pixel value represents the probability of it belonging to the damage category. Based on the confidence threshold set for the initial segmentation mask, it is converted into a binary mask. The binary mask is then processed using a connected component analysis algorithm to identify and mark the independent broken connected regions. For each identified broken connected region, its centroid pixel coordinates, minimum bounding rectangle parameters, pixel area, and category label are calculated sequentially; and the contour geometric features of the region, including equivalent diameter, aspect ratio, and contour irregularity, are further calculated. The calculated centroid pixel coordinates, circumscribed rectangle parameters, pixel area, category label, and contour geometric features are encapsulated into a structured data object, which serves as the pixel set of the damaged area.

[0010] Preferably, step S3 specifically includes: A repair strategy rule base is constructed, with the following specific rules: If the damage type is a hole and the equivalent diameter is less than the first threshold, then a point patching strategy is matched and associated with the first set of process control parameters; if the damage type is a crack and the aspect ratio is greater than the second threshold, then a strip patching strategy is matched and associated with the second set of process parameters; if the damage area is greater than a specific threshold or is an irregular sheet, then a surface patching strategy is matched. Based on the pixel set of the damaged area, quantitative analysis is performed to calculate the quantitative feature set of equivalent diameter, aspect ratio, and contour irregularity. The quantified feature set is matched with the preset repair strategy rule base to determine the repair strategy; Based on the repair strategy and the set of quantitative features, a set of corresponding process control parameters is automatically determined. The specific process control parameters include: hot melt temperature, plastic welding rod extrusion speed, pressure value of the pressing roller, and the dwell or movement speed of the repair tool on the damaged area; The determined repair strategy, the decided process control parameters, and the generated path instructions are encapsulated into a structured repair instruction package containing strategy instructions, parameter instructions, and path instructions.

[0011] Preferably, the path instructions specifically include: Based on the pixel set of the damaged area, the contour geometric features are calculated and matched with the repair strategy to automatically plan the motion trajectory of the end of the repair head. If the matching strategy is strip filling, then based on the circumscribed rectangle parameters of the damaged area, a unidirectional straight line filling path or a reciprocating polyline filling path along its main axis is generated. If the matching strategy is surface matching, then based on the contour shape of the damaged area, an annular path that gradually shrinks inward from the contour boundary, a spiral path that diverges outward from the internal starting point, or a multi-channel parallel grating scanning path that covers the entire area will be adaptively generated. The planning of the motion trajectory further includes the calculation of the path point spacing, the travel speed, and the start and stop points of the repair tool, wherein the path point spacing is determined based on the effective working width of the repair tool.

[0012] Preferably, step S4 specifically includes: Receive the three-dimensional coordinates Based on the path instructions in the repair instruction package, the robot arm's motion trajectory is planned according to the robot arm's current pose, the location of the damage, and the constraints of the greenhouse environment. The robotic arm moves along the trajectory and, through joint control and end-effector pose feedback, stabilizes the tool center point of the repair head above the damaged location. It adjusts the end-effector posture of the repair head by adjusting the normal direction of the greenhouse film surface to achieve perpendicularity to the surface to be repaired. The repair head is stably positioned, and the following job cycle is automatically executed in sequence using the strategy instructions and parameter instructions in the repair instruction package.

[0013] Preferably, the automatic execution of the following work cycle specifically includes: During the cleaning phase, the brushing device is activated to remove dust and moisture around the damaged area; During the pretreatment stage, the hot air unit is activated to controllably heat the damaged edges, causing the greenhouse film substrate to melt appropriately to improve adhesion. During the repair execution phase, if the strategy is point repair, the repair head is controlled to stay at the center of the damage and complete heating, extrusion and fixed-point pressing according to the set parameters; if the strategy is strip repair or surface repair, the repair head is controlled to move strictly along the path command and simultaneously and continuously perform heating, extrusion and roller pressing operations during the movement to achieve uniform filling and dense bonding of materials. In the post-processing stage, a brief pressure holding or secondary heating and smoothing operation is performed. After completion, the repair head is lifted away from the repair surface.

[0014] Preferably, step S5 specifically includes: After the repair operation is completed, the repair head is briefly left in place to ensure the initial curing of the material, and the integrated sensor immediately performs initial imaging of the repaired area; The quality inspection process begins, where cameras capture and verify images to analyze the continuity of the bond between the repair material and the original greenhouse film. Ultrasonic thickness gauges are used to measure physical parameters such as the thickness and flatness of the repaired area, thus completing the collection of multiple quality indicators. The central processing unit automatically compares the collected indicator data with the preset quality qualification threshold. If all key indicators meet the requirements, the repair is deemed qualified; if any indicator fails to meet the requirements, it is deemed unqualified. The system performs structured archiving of the complete data packet for this repair, forming a traceable repair log; if it is determined to be unqualified, the re-repair process is automatically triggered: the unqualified detection result is fed back to the decision process in step S3, and a new repair instruction is generated and executed; At the end of the task, the system automatically compiles and generates a comprehensive repair report, and updates the status markers corresponding to the repaired areas in the 3D digital map of the greenhouse.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a mobile platform for autonomous inspection and a robotic arm for automatic execution, enabling 24 / 7 uninterrupted unmanned operation and greatly improving the efficiency and response speed of greenhouse maintenance. It also proposes utilizing deep learning semantic segmentation models and 3D visual positioning technology, allowing the system to automatically identify various types of damage and calculate their precise physical location and geometric features, much like an expert. Based on a pre-set repair strategy rule base, the system can objectively and consistently match the optimal repair strategy and process parameters for different types of damage, overcoming the subjectivity and instability of manual judgment and ensuring the scientific and optimal nature of the repair plan. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the Cartographer algorithm of this invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, a method for repairing defects in a greenhouse includes: S1. Construct a 3D digital map of the greenhouse, generate a predetermined movement path through a path planning algorithm, control the mobile platform to move inside the greenhouse along the predetermined route, and continuously acquire images of the greenhouse film through the onboard vision system. Step S1 specifically includes: The mobile platform is equipped with LiDAR, visual sensors and integrated navigation system to perform mapping operations. Based on simultaneous positioning and map building algorithms, it integrates multi-source data to generate a three-dimensional digital map of the greenhouse. The multi-source data consists of three-dimensional geometric point cloud data obtained by LiDAR scanning the environment, textured images and corresponding depth information acquired by a depth camera, and timestamps and spatial pose labels added to the acquired data by the integrated navigation system. Based on multi-source data, a three-dimensional digital map containing the greenhouse film surface, support frame and internal facility structure is finally constructed by fusion and global optimization through synchronous positioning and mapping algorithms. Based on a 3D digital map, a predetermined movement path covering the target area is calculated using a path planning algorithm.

[0019] The predetermined movement path specifically includes: During the movement along the predetermined path, the real-time positioning and path-following control of the mobile platform are achieved by fusing data from wheel encoders and inertial measurement units and matching them with a three-dimensional digital map. The vision system uses an equidistant triggering mechanism to acquire images and adaptively adjusts the camera gimbal angle according to the platform's real-time pose to ensure continuous and vertical alignment with the greenhouse film surface; the acquired images are synchronously encapsulated with the platform's spatial coordinates and camera pose data at the time of acquisition.

[0020] The equidistant triggering mechanism's trigger distance is determined by both the physical width of the camera's field of view and the system's preset necessary image overlap rate; the calculation formula is expressed as follows: Trigger distance = field of view coverage width of a single frame image from the camera × (1 - preset image overlap rate).

[0021] The preset image overlap rate is a key system parameter (typically set to 20% to 50%), which ensures that there is sufficient common area between two adjacent frames for subsequent image stitching or independent damage verification. For example, in a typical embodiment, the camera's field of view on the target greenhouse membrane surface is 0.5 meters wide, and the system's preset image overlap rate requirement is 30%.

[0022] According to the formula above, the vision system automatically triggers an image acquisition every time the mobile platform advances by 0.5 meters × (1-30%) = 0.35 meters. This mechanism is tightly coupled with the mobile platform's real-time positioning system, ensuring that high-quality continuous image sequences can be acquired at this spatial interval regardless of whether the platform moves at a constant or variable speed. This provides a complete, comprehensive, and coherent data foundation for subsequent image analysis.

[0023] Based on the above content and... Figure 2 The simultaneous localization and mapping (SMR) algorithm specifically includes: To achieve high-precision environmental modeling, the Google Cartographer algorithm was selected as the core of simultaneous localization and mapping (SLAM). This algorithm is particularly suitable for greenhouse scenes with regular geometric features due to its stability and good adaptability to structured environments. The entire mapping process is a loop involving decision-making. First, the system constructs a local sub-map: the LiDAR on the mobile platform continuously acquires environmental point cloud data, which is preprocessed by voxel filtering and then matched with the currently active "sub-map" through the Ceres scanning matcher to solve for the optimal pose. The frame of data is then inserted into the sub-map to continuously update the local map. The key decision point is that after each frame of data is inserted, the system will determine whether the current sub-map is complete. The judgment criteria are usually whether the number of scan frames contained in the sub-map has reached the preset limit, or whether the robot has moved more than a certain distance from the center of the sub-map. If the submap is not completed, the system continues to receive new laser scan frames and repeats the above front-end matching and insertion process, focusing on expanding and improving the current local map; If a subgraph is determined to be complete, the system archives it and immediately activates the backend optimization process. This backend thread works in parallel, and its core is loop closure detection: the system compares the new real-time scan data with all completed historical subgraphs that may be spatially related. Once a loop closure is identified using an efficient branch and bound method, a strong constraint is generated; subsequently, the system initiates global pose graph optimization, using the SPA method to correct the trajectory drift error accumulated at the front end. Ultimately, all sub-maps that have undergone global optimization and calibration are seamlessly stitched and merged, outputting a globally consistent, accurate, and reliable 3D point cloud map, namely the "greenhouse 3D digital map" upon which this patented method relies, providing a unified spatial reference for subsequent autonomous navigation and damage location.

[0024] S2. Based on the greenhouse film image, identify the damaged area and determine its location coordinates within the greenhouse space; The specific steps of step S2 are as follows: S21. Preprocessing is performed on the encapsulated acquired image by illumination normalization, Gaussian filtering, and image registration when using multispectral images. S22. Input the preprocessed image into a pre-trained deep learning semantic segmentation model and output the pixel set of the damaged area; S23. Perform feature quantization on the segmented damaged area, calculate its geometric features and extract its centroid pixel coordinates as representative image positions. S24. Using the pre-calibrated camera intrinsic parameter matrix, back-project the original centroid pixel coordinates and normalize them to a two-dimensional plane in the camera coordinate system. S25. Combining the depth value corresponding to the damage point measured by the depth sensor, the two-dimensional plane is converted into a three-dimensional spatial point in the camera coordinate system. ; S26, Based on three-dimensional spatial points Subtract the extrinsic matrix In Perform a reverse translation operation to obtain a transition coordinate. The extrinsic parameter matrix is ​​provided by the synchronous positioning and mapping system at the time of image acquisition. For rotation matrix, It is a translation vector; S27, the transition coordinates Left-multiplying the transpose of the rotation matrix R Perform the reverse rotation operation and output the three-dimensional coordinates. ; S28, Output three-dimensional coordinates The coordinate system is aligned and verified with the 3D digital map of the greenhouse, and the 3D position coordinates of the damaged area in physical space are output.

[0025] The deep learning semantic segmentation model specifically includes: The preprocessed greenhouse film image is input into the loaded semantic segmentation model. The model performs forward propagation based on an encoder-decoder architecture, extracting features through the encoder and restoring spatial details through the decoder, and outputting an initial segmentation mask corresponding to the spatial size of the input image, where each pixel value represents the probability of it belonging to the damage category. Based on the confidence threshold set for the initial segmentation mask, it is converted into a binary mask. The binary mask is then processed using a connected component analysis algorithm to identify and mark the independent broken connected regions. For each identified broken connected region, its centroid pixel coordinates, minimum bounding rectangle parameters, pixel area, and category label are calculated sequentially; and the contour geometric features of the region, including equivalent diameter, aspect ratio, and contour irregularity, are further calculated. The calculated centroid pixel coordinates, circumscribed rectangle parameters, pixel area, category label, and contour geometric features are encapsulated into a structured data object, which serves as the pixel set of the damaged area.

[0026] S3. Analyze the image features of the damaged area and generate repair instructions; Step S3 specifically includes: A repair strategy rule base is constructed, with the following specific rules: If the damage type is a hole and the equivalent diameter is less than the first threshold, then a point patching strategy is matched and associated with the first set of process control parameters; if the damage type is a crack and the aspect ratio is greater than the second threshold, then a strip patching strategy is matched and associated with the second set of process parameters; if the damage area is greater than a specific threshold or is an irregular sheet, then a surface patching strategy is matched. Based on the pixel set of the damaged area, quantitative analysis is performed to calculate the quantitative feature set of equivalent diameter, aspect ratio, and contour irregularity. The quantified feature set is matched with the preset repair strategy rule base to determine the repair strategy; Based on the repair strategy and the set of quantitative features, a set of corresponding process control parameters is automatically determined. The specific process control parameters include: hot melt temperature, plastic welding rod extrusion speed, pressure value of the pressing roller, and the dwell or movement speed of the repair tool on the damaged area; The determined repair strategy, the decided process control parameters, and the generated path instructions are encapsulated into a structured repair instruction package containing strategy instructions, parameter instructions, and path instructions.

[0027] The path instructions specifically include: Based on the pixel set of the damaged area, the contour geometric features are calculated and matched with the repair strategy to automatically plan the motion trajectory of the end of the repair head. If the matching strategy is strip filling, then based on the circumscribed rectangle parameters of the damaged area, a unidirectional straight line filling path or a reciprocating polyline filling path along its main axis is generated. If the matching strategy is surface matching, then based on the contour shape of the damaged area, an annular path that gradually shrinks inward from the contour boundary, a spiral path that diverges outward from the internal starting point, or a multi-channel parallel grating scanning path that covers the entire area will be adaptively generated. The planning of the motion trajectory further includes the calculation of the path point spacing, the travel speed, and the start and stop points of the repair tool, wherein the path point spacing is determined based on the effective working width of the repair tool.

[0028] Based on the above, the repair strategy rule base specifically includes: The repair strategy rule base solidifies expert knowledge and experimental data into quantifiable decision logic; the system automatically matches the optimal repair scheme based on the precise characteristics of the damage; the decision logic of a specific embodiment is illustrated below: When the system identifies the damage type as "small hole" and its equivalent diameter is less than 2 cm, it automatically matches the point repair strategy and calls a set of corresponding process parameters, such as hot melt temperature 260°C, extrusion speed 2 mm / s, pressing pressure 3 N, and adopts a fixed-point operation path. If the damage is identified as a “slender crack” and meets the conditions of aspect ratio greater than or equal to 5 and equivalent diameter less than 1 cm, then a strip patching strategy is matched, with parameters such as temperature 280°C, extrusion speed 3 mm / s, pressure 4 N, and a “Z” shaped filling path is planned along the crack direction. For "large holes" with an equivalent diameter between 2 and 10 cm, a surface patching strategy is adopted, using a temperature of 270°C, a speed of 4 mm / s and a pressure of 5 N to perform concentric circular path coverage; for large-area tears, aging areas or damage with an equivalent diameter greater than 10 cm, the system defaults to a spiral filling surface patching strategy, using a temperature of 265°C to perform repair from the inside out. In actual decision-making, the central processing unit compares the above rules sequentially. Once the features are completely matched, the query is terminated immediately, and a complete repair instruction package containing strategies, parameters, and paths is generated. This rule base based on feature quantification is the core of realizing intelligent and standardized automation of repair operations.

[0029] S4. Control the robotic arm to carry the repair head to the position coordinates, execute the repair command, and complete the repair operation on the damaged area; Step S4 specifically includes: Receive the three-dimensional coordinates Based on the path instructions in the repair instruction package, the robot arm's motion trajectory is planned according to the robot arm's current pose, the location of the damage, and the constraints of the greenhouse environment. The robotic arm moves along the trajectory and, through joint control and end-effector pose feedback, stabilizes the tool center point of the repair head above the damaged location. It adjusts the end-effector posture of the repair head by adjusting the normal direction of the greenhouse film surface to achieve perpendicularity to the surface to be repaired. The repair head is stably positioned, and the following job cycle is automatically executed in sequence using the strategy instructions and parameter instructions in the repair instruction package.

[0030] The automatic execution of the following job cycle specifically includes: During the cleaning phase, the brushing device is activated to remove dust and moisture around the damaged area; During the pretreatment stage, the hot air unit is activated to controllably heat the damaged edges, causing the greenhouse film substrate to melt appropriately to improve adhesion. During the repair execution phase, if the strategy is point repair, the repair head is controlled to stay at the center of the damage and complete heating, extrusion and fixed-point pressing according to the set parameters; if the strategy is strip repair or surface repair, the repair head is controlled to move strictly along the path command and simultaneously and continuously perform heating, extrusion and roller pressing operations during the movement to achieve uniform filling and dense bonding of materials. In the post-processing stage, a brief pressure holding or secondary heating and smoothing operation is performed. After completion, the repair head is lifted away from the repair surface.

[0031] S5. Complete the repair operation, monitor the repaired area, and determine whether it is qualified. If it is not qualified, return to the decision process of S3, generate a new repair instruction and execute it; if it is qualified, record the repair result.

[0032] Step S5 specifically includes: After the repair operation is completed, the repair head is briefly left in place to ensure the initial curing of the material, and the integrated sensor immediately performs initial imaging of the repaired area; The quality inspection process begins, where cameras capture and verify images to analyze the continuity of the bond between the repair material and the original greenhouse film. Ultrasonic thickness gauges are used to measure physical parameters such as the thickness and flatness of the repaired area, thus completing the collection of multiple quality indicators. The central processing unit automatically compares the collected indicator data with the preset quality qualification threshold. If all key indicators meet the requirements, the repair is deemed qualified; if any indicator fails to meet the requirements, it is deemed unqualified. The system performs structured archiving of the complete data packet for this repair, forming a traceable repair log; if it is determined to be unqualified, the re-repair process is automatically triggered: the unqualified detection result is fed back to the decision process in step S3, and a new repair instruction is generated and executed; At the end of the task, the system automatically compiles and generates a comprehensive repair report, and updates the status markers corresponding to the repaired areas in the 3D digital map of the greenhouse.

[0033] The aforementioned quality qualification thresholds are as follows: To ensure controllable and reliable repair quality, the system presets clear and quantifiable quality acceptance thresholds, which serve as the sole standard for automatically determining whether a repair is acceptable. These thresholds are specifically reflected in three core dimensions: First, in terms of appearance quality, pixel-level analysis of the verification images is required to ensure that the average pixel width of the gap between the repair material and the original greenhouse film is less than 2 pixels. Secondly, in terms of physical properties, using an ultrasonic thickness gauge, the average thickness of the repaired area should not deviate from the average thickness of the surrounding intact greenhouse film by more than ±15%. Finally, regarding the surface morphology, the flatness difference between the repaired area and the surrounding film surface is required to be less than 0.5 mm, as measured by a high-precision displacement sensor. During the quality review phase, the central processing unit automatically compares the real-time collected test data with the aforementioned preset thresholds; only when all indicators meet the requirements will the system determine that the repair is "qualified", thus realizing the objectivity, standardization and automation of quality acceptance.

[0034] The present invention will now be described in detail with reference to a specific embodiment: In a solar greenhouse with a length of 80 meters and a span of 10 meters, an inspection and repair system as described in this invention is deployed. After the system is powered on for the first time, the operator sets the inspection area to the entire inner surface of the greenhouse film through the control terminal. The mobile platform (model: custom AGV chassis, equipped with Velodyne VLP-16 LiDAR and Intel Realsense D455 depth camera) starts from the starting point at the end of the greenhouse and automatically travels along the preset preliminary exploration path. During the travel, the system calls the Cartographer laser SLAM algorithm to fuse laser point cloud, IMU and wheel encoder data in real time. After about 25 minutes of traversing the entire greenhouse, a three-dimensional point cloud digital map of the greenhouse with centimeter-level accuracy is successfully constructed and optimized. The map clearly marks the curved surface of the arched greenhouse film, all steel frame structures and the gable walls at both ends.

[0035] After the map is built, the system automatically enters the daily inspection mode; the central controller plans a predetermined "bow"-shaped path covering the entire greenhouse area based on the 3D map; the mobile platform travels autonomously along the path at a speed of 0.3 meters per second; its onboard vision system (20-megapixel industrial camera, field of view 0.5 meters) strictly follows the equidistant triggering mechanism (trigger distance = 0.5 meters × (1-30%) = 0.35 meters) to acquire images; when the platform travels to the vicinity of the coordinates (40, 5, 2.5) in the middle of the greenhouse, the camera acquires a frame of greenhouse film image containing suspicious features.

[0036] The image frame was immediately transmitted to the industrial control computer (equipped with NVIDIA Jetson AGX Orin); the image preprocessing module first performed illumination equalization and Gaussian filtering, and then input it into the pre-trained deep learning semantic segmentation model (based on the U-Net architecture, trained on 5000 labeled images of greenhouse film damage, with a test set mIoU of 89%); the model output an initial segmentation mask, which, after thresholding and connected component analysis, confirmed the existence of an independent damaged area; the system calculated that the pixel area of ​​this area was 850 pixels, and the aspect ratio of the minimum bounding rectangle was approximately 7:1, classifying it as a "crack", and extracting its centroid pixel coordinates (u,v); combined with the precise extrinsic parameters of the camera at this moment (provided in real time by the SLAM system) and the depth value Z_c = 2.48 meters measured by the depth camera, the precise three-dimensional position coordinates of the damage in the greenhouse world coordinate system were calculated as (40.12, 5.03, 2.52) through inverse perspective transformation; The central decision-making unit receives the damage characteristics: type = "crack", equivalent diameter (after conversion) ≈ 0.8 cm, aspect ratio ≈ 7. The system then queries the repair strategy rule base for matching. According to the rule: "type = 'crack' and aspect ratio ≥ 5 and equivalent diameter < 1 cm", a successful match is found, and the decision is to adopt the "strip patching" strategy. The system automatically generates the corresponding process parameters: hot melt temperature 280°C, plastic welding rod extrusion speed 3 mm / s, and pressing roller pressure 4 Newtons. Simultaneously, the path planning module generates a "Z"-shaped filling path along the length of the crack, with a spacing 1.5 times the width of the welding rod, based on the direction of the circumscribed rectangle of the crack. The above strategy, parameters, and path are encapsulated into a structured repair instruction package.

[0037] Control commands are sent to a six-axis collaborative robotic arm (model: UR5e). The robotic arm plans a collision-free path, moves the multi-functional repair head above the target coordinates (40.12, 5.03, 2.52), and adjusts its end effector according to the normal of the greenhouse film surface to ensure the repair head is vertically aligned with the surface. Then, it automatically executes the work cycle according to the instruction package: First, a micro air pump is activated to blow away dust from the crack; next, a hot air gun preheats the crack edges at 280°C for 2 seconds; then, the repair head moves strictly along a "Z" shaped path, simultaneously extruding molten repair material and compacting it in real-time with a pressure of 4 Newtons by a pressing roller; finally, after a 1-second pressure hold, the repair head is raised. After the repair head is raised, its integrated macro camera immediately performs high-definition imaging of the repair point, while an ultrasonic thickness gauge performs point measurements. After the data is transmitted back, the central processing unit automatically determines the following: image analysis shows no continuous gaps at the joint, and the average pixel width is 1.2 pixels (<2 pixels is acceptable); thickness measurement shows the average thickness of the repaired area is 0.185 mm, with a deviation of -7.5% from the surrounding film thickness of 0.20 mm (within ±15% is acceptable); flatness measurement shows a drop of 0.3 mm (<0.5 mm is acceptable). All indicators meet the preset quality acceptance threshold, and the system determines this repair to be "acceptable," storing the complete log of this operation (including coordinates, features, strategies, parameters, and quality data) in the database. Simultaneously, the status is updated to "repaired" at the corresponding location on the greenhouse's 3D digital map. After the quality review is completed, the system sends instructions to the mobile platform to continue along the predetermined path and perform subsequent inspection and repair tasks until the entire area is covered. This embodiment fully demonstrates the system's fully automated closed-loop process from "perception" to "decision," "execution," and "verification," showcasing the high intelligence and reliability of the invention.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for repairing defects in a greenhouse, characterized in that, include: S1. Construct a 3D digital map of the greenhouse, generate a predetermined movement path through a path planning algorithm, control the mobile platform to move inside the greenhouse along the predetermined route, and continuously acquire images of the greenhouse film through the onboard vision system. S2. Based on the greenhouse film image, identify the damaged area and determine its location coordinates within the greenhouse space; S3. Analyze the image features of the damaged area and generate repair instructions; S4. Control the robotic arm to carry the repair head to the position coordinates, execute the repair command, and complete the repair operation on the damaged area; S5. Complete the repair operation, monitor the repaired area, and determine whether it is qualified. If it is not qualified, return to the decision process of S3, generate a new repair instruction and execute it. If the repair is successful, the repair result will be recorded.

2. The method for repairing defects in a greenhouse according to claim 1, characterized in that, Step S1 specifically includes: The mobile platform is equipped with LiDAR, visual sensors and integrated navigation system to perform mapping operations. Based on simultaneous positioning and map building algorithms, it integrates multi-source data to generate a three-dimensional digital map of the greenhouse. The multi-source data consists of three-dimensional geometric point cloud data obtained by LiDAR scanning the environment, textured images and corresponding depth information acquired by a depth camera, and timestamps and spatial pose labels added to the acquired data by the integrated navigation system. Based on multi-source data, a three-dimensional digital map containing the greenhouse film surface, support frame and internal facility structure is finally constructed by fusion and global optimization through synchronous positioning and mapping algorithms. Based on a 3D digital map, a predetermined movement path covering the target area is calculated using a path planning algorithm.

3. The method for repairing defects in a greenhouse according to claim 2, characterized in that, The predetermined movement specifically includes: During the movement along the predetermined path, the real-time positioning and path-following control of the mobile platform are achieved by fusing data from wheel encoders and inertial measurement units and matching them with a three-dimensional digital map. The vision system uses an equidistant triggering mechanism to acquire images and adaptively adjusts the camera gimbal angle according to the platform's real-time pose to ensure continuous and vertical alignment with the greenhouse film surface; the acquired images are synchronously encapsulated with the platform's spatial coordinates and camera pose data at the time of acquisition.

4. The method for repairing defects in a greenhouse according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Preprocessing is performed on the encapsulated acquired image by illumination normalization, Gaussian filtering, and image registration when using multispectral images. S22. Input the preprocessed image into a pre-trained deep learning semantic segmentation model and output the pixel set of the damaged area; S23. Perform feature quantization on the segmented damaged area, calculate its geometric features and extract its centroid pixel coordinates as representative image positions. S24. Using the pre-calibrated camera intrinsic parameter matrix, back-project the original centroid pixel coordinates and normalize them to a two-dimensional plane in the camera coordinate system. S25. Combining the depth value corresponding to the damage point measured by the depth sensor, the two-dimensional plane is converted into a three-dimensional spatial point in the camera coordinate system. ; S26, Based on three-dimensional spatial points Subtract the extrinsic matrix In Perform a reverse translation operation to obtain a transition coordinate. The extrinsic parameter matrix is ​​provided by the synchronous positioning and mapping system at the time of image acquisition. For rotation matrix, It is a translation vector; S27, the transition coordinates Left-multiplying the transpose of the rotation matrix R Perform a reverse rotation operation and output three-dimensional coordinates. ; S28, Output three-dimensional coordinates The coordinate system is aligned and verified with the 3D digital map of the greenhouse, and the 3D position coordinates of the damaged area in physical space are output.

5. A method for repairing defects in a greenhouse according to claim 4, characterized in that, The deep learning semantic segmentation model specifically includes: The preprocessed greenhouse film image is input into the loaded semantic segmentation model. The model performs forward propagation based on an encoder-decoder architecture, extracting features through the encoder and restoring spatial details through the decoder, and outputting an initial segmentation mask corresponding to the spatial size of the input image, where each pixel value represents the probability of it belonging to the damage category. Based on the confidence threshold set for the initial segmentation mask, it is converted into a binary mask. The binary mask is then processed using a connected component analysis algorithm to identify and mark the independent broken connected regions. For each identified broken connected region, its centroid pixel coordinates, minimum bounding rectangle parameters, pixel area, and category label are calculated sequentially; and the contour geometric features of the region, including equivalent diameter, aspect ratio, and contour irregularity, are further calculated. The calculated centroid pixel coordinates, circumscribed rectangle parameters, pixel area, category label, and contour geometric features are encapsulated into a structured data object, which serves as the pixel set of the damaged area.

6. The method for repairing defects in a greenhouse according to claim 1, characterized in that, Step S3 specifically includes: A repair strategy rule base is constructed, with the following specific rules: If the damage type is a hole and the equivalent diameter is less than the first threshold, then a point patching strategy is matched and associated with the first set of process control parameters; if the damage type is a crack and the aspect ratio is greater than the second threshold, then a strip patching strategy is matched and associated with the second set of process parameters; if the damage area is greater than a specific threshold or is an irregular sheet, then a surface patching strategy is matched. Based on the pixel set of the damaged area, quantitative analysis is performed to calculate the quantitative feature set of equivalent diameter, aspect ratio, and contour irregularity. The quantified feature set is matched with the preset repair strategy rule base to determine the repair strategy; Based on the repair strategy and the set of quantitative features, a set of corresponding process control parameters is automatically determined. The specific process control parameters include: hot melt temperature, plastic welding rod extrusion speed, pressure value of the pressing roller, and the dwell or movement speed of the repair tool on the damaged area; The determined repair strategy, the decided process control parameters, and the generated path instructions are encapsulated into a structured repair instruction package containing strategy instructions, parameter instructions, and path instructions.

7. A method for repairing defects in a greenhouse according to claim 6, characterized in that, The path instructions specifically include: Based on the pixel set of the damaged area, the contour geometric features are calculated and matched with the repair strategy to automatically plan the motion trajectory of the end of the repair head. If the matching strategy is strip filling, then based on the circumscribed rectangle parameters of the damaged area, a unidirectional straight line filling path or a reciprocating polyline filling path along its main axis is generated. If the matching strategy is surface matching, then based on the contour shape of the damaged area, an annular path that gradually shrinks inward from the contour boundary, a spiral path that diverges outward from the internal starting point, or a multi-channel parallel grating scanning path that covers the entire area will be adaptively generated. The planning of the motion trajectory further includes the calculation of the path point spacing, the travel speed, and the start and stop points of the repair tool, wherein the path point spacing is determined based on the effective working width of the repair tool.

8. A method for repairing defects in a greenhouse according to claim 1, characterized in that, Step S4 specifically includes: Receive the three-dimensional coordinates Based on the path instructions in the repair instruction package, the robot arm's motion trajectory is planned according to the robot arm's current pose, the location of the damage, and the constraints of the greenhouse environment. The robotic arm moves along the trajectory and, through joint control and end-effector pose feedback, stabilizes the tool center point of the repair head above the damaged location. It adjusts the end-effector posture of the repair head by adjusting the normal direction of the greenhouse film surface to achieve perpendicularity to the surface to be repaired. The repair head is stably positioned, and the following job cycle is automatically executed in sequence using the strategy instructions and parameter instructions in the repair instruction package.

9. A method for repairing defects in a greenhouse according to claim 8, characterized in that, The automatic execution of the following job cycle specifically includes: During the cleaning phase, the brushing device is activated to remove dust and moisture around the damaged area; During the pretreatment stage, the hot air unit is activated to controllably heat the damaged edges, causing the greenhouse film substrate to melt appropriately to improve adhesion. During the repair execution phase, if the strategy is point repair, the repair head is controlled to stay at the center of the damage and complete heating, extrusion and fixed-point pressing according to the set parameters; if the strategy is strip repair or surface repair, the repair head is controlled to move strictly along the path command and simultaneously and continuously perform heating, extrusion and roller pressing operations during the movement to achieve uniform filling and dense bonding of materials. In the post-processing stage, a brief pressure holding or secondary heating and smoothing operation is performed. After completion, the repair head is lifted away from the repair surface.

10. A method for repairing defects in a greenhouse according to claim 1, characterized in that, Step S5 specifically includes: After the repair operation is completed, the repair head is briefly left in place to ensure the initial curing of the material, and the integrated sensor immediately performs initial imaging of the repaired area; The quality inspection process begins, where cameras capture and verify images to analyze the continuity of the bond between the repair material and the original greenhouse film. Ultrasonic thickness gauges are used to measure physical parameters such as the thickness and flatness of the repaired area, thus completing the collection of multiple quality indicators. The central processing unit automatically compares the collected indicator data with the preset quality qualification threshold. If all key indicators meet the requirements, the repair is deemed qualified; if any indicator fails to meet the requirements, it is deemed unqualified. The system archives the complete data packet for this repair in a structured manner to form a traceable repair log; if it is determined to be unqualified, the re-repair process is automatically triggered: the unqualified detection result is fed back to the decision process in step S3, and a new repair instruction is generated and executed; at the end of the task, the system automatically summarizes and generates a comprehensive repair report, and updates the status identifier corresponding to the repair area in the greenhouse 3D digital map.

Images