A Two-Stage Hybrid Heuristic for Crack Path Detection in Solar Cells

Abstract

This invention discloses a two-stage hybrid heuristic method for crack path detection in solar cells. The method includes: acquiring an electroluminescence image; generating a crack potential map (CPM) that suppresses background noise and enhances crack features using cascaded morphological processing; dividing the CPM into a coarse-grained region of interest (ROI) grid; searching for a globally optimized ROI path that satisfies spatial continuity constraints in the grid space using a first heuristic algorithm; generating a constrained pixel search space based on the regions contained in this path and after region expansion processing; and finally, searching for the optimal crack path within the constrained space using a second heuristic algorithm based on a movement cost that is non-linearly negatively correlated with the pixel potential value. This invention effectively reduces the search space and, combined with a cost function modified for crack features, achieves high-precision, topologically complete crack path extraction without the need for a high-performance GPU.

Description

Technical Field

[0001] This invention relates to the field of swarm intelligence optimization algorithms, and more particularly to a two-stage hybrid heuristic method for crack path detection in solar cells. Background Technology

[0002] Solar photovoltaic (PV) is a key component of renewable energy. However, solar cells (usually made of crystalline silicon) are highly susceptible to microcracks due to mechanical and thermal stresses during manufacturing, transportation, installation, and long-term service. These cracks are among the most common defects in solar cell modules; they impede current conduction, form electrically isolated regions, lead to power degradation, and can even trigger hotspot effects, severely impacting the power generation efficiency, reliability, and lifespan of the photovoltaic module.

[0003] Electroluminescence (EL) imaging is currently recognized by industry and academia as one of the most effective and highest-resolution non-destructive testing techniques for microcracks. EL images are created by applying a forward bias voltage to a solar cell, causing it to emit light, which is then captured by a high-sensitivity camera. In EL images, intact areas exhibit uniform brightness, while defective areas such as cracks and broken grids appear as dark lines, dark areas, or non-luminescent regions due to increased recombination rates or current interruption.

[0004] Traditionally, crack detection using EL images relies on manual visual inspection. However, this method is highly subjective, inefficient, prone to fatigue, and suffers from high false negative and false negative rates for complex and minute cracks. With the expansion of the photovoltaic industry and the increasing speed of automated production lines, the development of fast, accurate, and automated crack detection algorithms has become an urgent need.

[0005] To address this, existing technologies have proposed various automated detection solutions:

[0006] Deep Learning (CNN)-Based Segmentation Schemes: In recent years, deep learning methods, represented by Convolutional Neural Networks (CNNs), have dominated the detection of defects in electroluminescence (EL) images. For example, S. Deitsch et al., in their paper "Automatic classification of defective photovoltaic module cells in electroluminescence images" (Solar Energy, 2019, 185: 455–468), disclosed a semantic segmentation model based on the U-Net architecture for automatically segmenting crack pixels from EL images. Similarly, researchers have also used various CNN architectures such as ResNet and SqueezeNet for defect classification and segmentation, achieving high detection accuracy.

[0007] Problems: Although this type of method has high accuracy, it has two main technical bottlenecks: (1) High computational complexity: CNN models (especially deep networks such as U-Net and ResNet) contain millions of parameters. Their training and inference processes require high-performance graphics processing units (GPUs), resulting in high computational costs and relatively long inference times. This makes them unsuitable for all online detection scenarios that require low latency and high throughput. (2) Data dependence: The performance of deep learning models is highly dependent on large-scale, high-quality, and balanced labeled datasets. However, in actual production, obtaining pixel-level accurately labeled crack datasets is costly, and crack samples (dark features) are usually extremely imbalanced compared to the background (bright features), which makes model training difficult.

[0008] Schemes based on traditional image processing and heuristic algorithms: Before the rise of deep learning, researchers attempted to use traditional image processing (such as filtering and edge detection) combined with machine learning (such as SVM) or heuristic algorithms (such as PSO and GA) for detection. For example, AH Aghamohammadi et al., in their paper "Solar cell panel crack detection using Particle Swarm Optimization algorithm" (2011 International Conference on Pattern Analysis and Intelligence Robotics, DOI: 10.1109 / ICPAIR.2011.5976888), disclosed a method for edge detection of EL images using the Particle Swarm Optimization (PSO) algorithm, and then used fuzzy logic to classify the extracted edge features. Other studies, such as [references to studies on crack classification], used PSO to optimize the hyperparameters of Support Vector Machines (SVM) to improve the accuracy of crack classification.

[0009] Problems: While these methods have low computational complexity, their core issue lies in the completeness of detection. As shown, algorithms like PSO are primarily used for auxiliary edge detection or parameter optimization, rather than direct path extraction. Cracks in EL images (especially at low contrast) are often discontinuous and blurry. Simple edge detection algorithms can only identify isolated, fragmented edge points, failing to topologically connect these points into a complete, continuous crack path. This results in low accuracy and a high false negative rate when dealing with complex cracks (such as mesh cracks and dendritic cracks).

[0010] In addition, there are technologies that use infrared (IR) thermal imaging for detection, such as those disclosed in US patent application US20170277966A1. However, infrared imaging is mainly sensitive to "hot spots" that have already generated significant power loss and heat generation. For early-stage microcracks that have not yet formed hot spots, its detection sensitivity and resolution are far lower than those of EL imaging.

[0011] In summary, there is a core technical contradiction in the existing technology for detecting cracks in solar cells: Deep learning-based solutions pursue high detection accuracy and completeness, but this leads to high computational complexity; while solutions based on traditional heuristic algorithms maintain low computational complexity, but their detection accuracy and completeness (especially path completeness) are low.

[0012] Therefore, there is an urgent need in this field for a new technical solution that can overcome the above-mentioned technical contradictions, namely, to achieve high-precision and high-completeness extraction of complex crack paths in EL images while maintaining low computational complexity (no need for expensive GPUs and fast inference speed). Summary of the Invention

[0013] The purpose of this invention is to provide a two-stage hybrid heuristic method for detecting crack paths in solar cells, in order to solve the technical problems existing in the prior art.

[0014] To achieve the above objectives, the present invention provides a two-stage hybrid heuristic method for crack path detection in solar cells, comprising the following steps:

[0015] (a) Acquire the electroluminescent image , This represents the grayscale value of the pixel at coordinates (x, y).

[0016] (b) Based on the electroluminescent image A crack potential map is generated through nonlinear morphological image processing. , wherein The pixel value in the image represents the potential for a crack to exist at that pixel location;

[0017] (c) The crack potential map Divided into There are ROI grids, among which and It is an integer greater than 1;

[0018] (d) Using the first heuristic algorithm, in the... In a grid space composed of ROI grids, search for and determine a globally optimized ROI path. The It is a spatially continuous grid index sequence of ROIs;

[0019] (e) Based on the aforementioned globally optimized ROI path In the Define a constrained pixel search space. The Based on the above The region is the set of pixels corresponding to the union of the contained ROI grids, and the region obtained by performing region expansion processing on the set of pixels.

[0020] (f) Employing the second heuristic algorithm in the constrained pixel search space Internally, a pixel-level optimal crack path is searched and determined based on the principle of minimizing movement cost. The cost of the movement is related to the The pixel potential values ​​in the data exhibit a non-linear negative correlation.

[0021] Further, step (b) generates the crack potential map. include:

[0022] (b1) The electroluminescent image The application has a first structural element Morphological closing operations are used to obtain an approximate background image. ;

[0023] (b2) By calculation The black hat transformation diagram is obtained. The It highlights the dark features in the image;

[0024] (b3) The black hat transformation diagram The application has a second structural element Morphological opening operations are used to obtain crack feature enhancement maps. ;

[0025] (b4) The above The crack potential map is obtained by performing normalization. .

[0026] Furthermore, the first structural element The size of the second structural element is greater than the preset maximum crack width and less than the scale of the background brightness change. It is a linear structuring element.

[0027] Furthermore, the first heuristic algorithm is a genetic algorithm, which introduces neighborhood constraints during initialization and evolution to ensure the spatial continuity of the ROI grid index sequence.

[0028] Furthermore, the genetic algorithm uses chromosomes. Representing a sequence of ROI paths, the fitness function of the genetic algorithm. It is determined to be a weighted combination of at least the following three items:

[0029] (i) The chromosome The average maximum crack potential of the included ROI mesh sequence ;

[0030] (ii) The chromosomes The reciprocal of the total length of the path represented ;

[0031] (iii) The chromosome The reciprocal of the smoothness of the represented path .

[0032] Furthermore, the aforementioned , and Calculated using the following formula: , , ,

[0033] in, For the chromosome The number of ROI grids in the middle For the first One ROI grid, for The coordinates of the center point, for and The Euclidean distance between them To preset the maximum path length, For vector sum vector exist The angle formed at the point; when hour, Let it be a constant of 1.

[0034] Furthermore, in step (e), the region expansion process is a morphological dilation operation, and the application has a preset width. structural elements .

[0035] Furthermore, the second heuristic algorithm is the A* algorithm.

[0036] Furthermore, the total cost function of the A* algorithm ,in For the current pixel node, for The parent node, From the starting point to The cumulative movement cost, From The heuristic cost to the destination, the Calculated using the following formula: ,

[0037] in, for and pixel distance between for The potential value of the point on the crack potential map. A preset constant is used to prevent division by zero; This is the cost scaling factor, and its value range is... This is used to adjust the sensitivity of path search to crack potential value, so as to prevent path straightening while suppressing background noise interference.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1) This invention utilizes coarse-grained global search to quickly locate the region (ROI) where the crack is located, while strictly limiting fine-grained pixel-level search to an extremely small area. This strategy reduces computational complexity by several orders of magnitude, allowing the algorithm to run in real-time on a CPU without the need for a GPU;

[0040] 2) This invention modifies the cost function of the second heuristic algorithm by introducing a cost scaling factor. This causes the crack potential value of the pixel to have a non-linear negative correlation (specifically, a reciprocal exponential relationship). This means that when the algorithm is looking for the "minimum cost path," it is actually looking for the "maximum crack probability path," thus ensuring that the extracted path is topologically continuous and closely follows the crack center.

[0041] 3) The present invention introduces smoothness constraints and continuity constraints in the global search stage, which effectively avoids path jaggedness caused by noise and makes the detection results more consistent with the physical laws of crack fracture mechanics. Attached Figure Description

[0042] Figure 1 This is a flowchart of the improved overall method of the present invention;

[0043] Figure 2 This is the crack potential map generation process of the present invention;

[0044] Figure 3 This invention relates to a multi-scale ROI mesh generation method;

[0045] Figure 4 This describes the workflow of the genetic algorithm in this invention.

[0046] Figure 5 This invention provides a process for constructing the constrained search space.

[0047] Figure 6 A flowchart for refining the path of the A* algorithm. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. The embodiments provided in this specification are intended to cover all the technical features described in the claims and to provide the best mode for implementing this technology.

[0049] Example 1:

[0050] This embodiment details the complete crack detection process of a two-stage hybrid heuristic method for crack path detection in solar cells. The method is designed to overcome the fracture problem encountered by traditional image processing operators (such as Canny and Sobel) when processing EL images, as well as the dependence of deep learning methods (such as U-Net) on expensive GPU hardware. This method achieves high-precision crack skeleton extraction on a standard CPU through a strategy of "macro-global search + micro-local refinement".

[0051] like Figure 1 As shown, the method specifically includes the following six detailed steps:

[0052] Step (a): High-quality acquisition of electroluminescence images

[0053] First, acquire electroluminescence images of the solar cell module. .

[0054] Apply a forward bias voltage (typically 1) to the photovoltaic module The injection of current (doubled open-circuit voltage) causes recombination of non-equilibrium carriers (electrons and holes) in the silicon substrate, emitting near-infrared light with a wavelength of approximately 1150 nm. Cracks, broken gates, or dark spots appear as dark lines or dark areas with significantly lower brightness than the surrounding areas in images, due to increased non-radiative recombination of carriers or obstructed current transport caused by lattice defects.

[0055] The acquired image is a grayscale matrix. A Cartesian coordinate system is established, where... Indicates column index (range) ), Indicates row index (range) In this embodiment, the preferred image resolution is... Pixels, with a bit depth of 12 or 16 bits, to preserve crack texture details in low contrast.

[0056] Step (b): Generation of Crack Potential Map (CPM) based on Cascade Morphology

[0057] The raw EL image has a complex background, often containing grain boundary noise, cell edge shadows, and brightness gradients caused by uneven series resistance. Directly searching for paths on the raw grayscale image is easily disturbed. Therefore, this step uses nonlinear morphological image processing to generate a clean crack potential map. .

[0058] The process consists of four detailed sub-steps:

[0059] 1. Similar background:

[0060] For the original image Apply morphological closing operations. The closing operation is defined as performing a dilation operation first, followed by an erosion operation. The size selection is crucial. It must be larger than the preset maximum crack width (typically 3-10 pixels) and smaller than the scale of slowly changing background lighting (such as half the width of a solar cell). This embodiment preferably uses... A rectangular structuring element of pixels. The dilation operation takes the maximum value within the structuring element's coverage area, causing the surrounding bright background to "flood" the tiny dark cracks; the subsequent erosion operation restores the edges of the large-scale background. The final result... It is an "idealized" background lighting image without any cracks.

[0061] 2. Black Hat Transformation: Calculate the difference . Represents background, This represents a defect superimposed on the background. Subtracting the two mathematically eliminates the uneven background lighting (low-frequency component), preserving only details in the image that are darker than the background (high-frequency component). At this point, The higher the pixel value, the darker the area is compared to the background.

[0062] 3. Feature enhancement: Besides cracks, it also contains a large amount of point noise (such as grain boundary contamination and bad spots). For Morphological opening operations are applied. Opening operations involve erosion followed by dilation. Prior geometric knowledge of cracks—"linear features"—is utilized. This embodiment constructs linear structuring elements. To simultaneously capture cracks in different directions, the specific operations are as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Horizontal structural elements of pixels and The vertical structuring elements of the pixels are opened, and the maximum value between the two is taken as the output. (Size smaller than...) Isolated noise points are completely erased and cannot be recovered during the corrosion stage; however, continuous crack features are preserved. This results in an enhanced crack feature map. .

[0063] 4. Normalization: ... The pixel values ​​are linearly mapped to Interval, generation . The value directly represents the "confidence" or "potential" of a pixel belonging to a crack. This indicates a very high probability that it is a crack.

[0064] Step (c): Multi-scale ROI mesh generation

[0065] To reduce the complexity of subsequent search algorithms, the high-dimensional pixel space is reduced to a low-dimensional grid space.

[0066] Will of The map is divided into A regional interest (ROI) grid is defined in this embodiment. Each grid ROI Corresponding to the original image A sub-region of a pixel. The search space starts from... The number of nodes suddenly dropped The presence of multiple nodes enables the genetic algorithm to converge in milliseconds.

[0067] Step (d): Global optimization path search based on neighborhood-constrained genetic algorithm

[0068] This step uses the first heuristic algorithm (Genetic Algorithm GA) to search for a rough crack path at the grid level. .

[0069] 1. Chromosome Encoding and Population Initialization (Detailed Implementation of Continuity Constraints): It must be "spatially continuous". To ensure this, this embodiment imposes strict restrictions on the underlying logic of the algorithm:

[0070] chromosome Encoded as a variable-length grid index sequence Each gene Includes grid coordinates Traditional random walks (GAs) typically generate genes randomly, which leads to discrete, invalid paths. This invention employs a "restricted random walk" strategy to initialize the population:

[0071] a. Randomly select the largest internal value. Grids with values ​​exceeding a threshold (e.g., 0.5) are used as the starting point. .

[0072] b. For the current node its next node Only from 8-neighborhood (i.e. Choose from (excluding itself).

[0073] c. In an 8-neighborhood, the probability of being selected is proportional to the potential value of that neighborhood grid. This simulates the physical process of crack propagation.

[0074] Through this mechanism, every path in the initial population is naturally continuous in physical space.

[0075] 2. Design and Calculation of Fitness Function: To select the path that most resembles a crack, a fitness function is defined. Weighted for three items: .

[0076] This embodiment sets weights. The calculations for each item are as follows:

[0077] Potential Adaptability : ,

[0078] Calculate the average value of the "brightest (highest potential)" pixel within each grid cell the path passes through. This value should be close to 1 if the path accurately covers the crack.

[0079] Length adaptability : ,

[0080] It is the geometric center of the grid. This is a preset constant (such as the diagonal length). This term, by taking its reciprocal, penalizes excessively long redundant paths, embodying the "Occam's Razor" principle.

[0081] Smoothness adaptability : ,

[0082] The path is at the node The angle between the forward vector and the backward vector at a given point.

[0083] If the crack extends straight. (180 degrees), The punishment is minimal.

[0084] o If the path involves an acute angle turn (such as a zig-zag). It is very large, and its adaptability drops sharply.

[0085] o Boundary handling: If the path is too short ( If an angle cannot be formed, then directly set... .

[0086] 3. Maintaining continuity during the evolutionary process:

[0087] The crossover operation uses a location-aware, single-point crossover. Cutting and swapping are only performed when two parent paths coincide or are adjacent at a grid location, ensuring the continuity of the child path. The mutation operation fine-tunes the path (e.g., by moving nodes...). (Move to the side), verification is required. and and Whether to maintain 8-neighborhood connectivity. If broken, trigger local repair logic (insert intermediate node) or roll back mutation.

[0088] Step (e): Constructing the constrained pixel search space: Only a coarse corridor is provided, and due to quantization errors in the meshing, cracks may be located at the mesh edges. This step aims to generate a search domain that is both inclusive of the cracks and sufficiently small. .

[0089] 1. Region extraction: Traversal Each ROI index in the image will correspond to a specific location in the original image. Pixel blocks are marked as foreground (set to 1), and the rest as background (set to 0), generating an initial binary mask. .

[0090] 2. Region expansion processing (core feature): For Perform the expansion. This embodiment employs a morphological expansion operation.

[0091] 3. Parameter settings: Use a width of Square structural element of pixels .

[0092] 4. Mechanism of Action: The dilation operation pushes the jagged boundaries of the mesh outward by 10 pixels, smoothing the search space boundaries and filling the gaps at mesh connections. This creates a fault-tolerant "buffer zone" that ensures that even if the crack center deviates from the mesh center, it is still contained within the search space.

[0093] Step (f): Micropath Refinement Based on Nonlinear Cost A* Algorithm

[0094] In the smallest Within a region (typically only 2-5% of the total image pixels), a second heuristic algorithm (A* algorithm) is run for pixel-level precise localization.

[0095] 1. Graph model construction: Each pixel in the graph is considered a node in the graph. .

[0096] The starting point is Select the first grid area The pixel with the largest value. The endpoint is... Select the last grid area The pixel with the largest value. Neighbor nodes are connected to each pixel and its 8-neighbor pixels.

[0097] 2. Design and Physical Significance of the Cost Function: The core of the A* algorithm lies in the cost function. This invention addresses the cumulative movement cost. Creative modifications were made: ,

[0098] Among them, the moving resistance factor Defined as: ,

[0099] In-depth parameter analysis: Geometric distance. Horizontal / vertical movement is 1, diagonal movement is 1.414. (Numerical stability term): Set as .when (For pure background) Prevent division by zero errors. (Cost scaling factor): This is a key non-linear adjustment parameter, which is limited to... Optimal value 2.0. Why is it non-linear? Assume... For the crack center pixel ( resistance factor For background noise pixels ( ), resistance factor Although background pixels may be geometrically closer, their passage cost is nearly a hundred times that of cracked pixels. This huge difference in cost creates a "high wall" or "deep valley" in the search space.

[0100] 3. Algorithm execution flow:

[0101] 1. Add the starting point to the priority queue.

[0102] 2. Loop until the queue is empty or the end point is found:

[0103] 3. Remove smallest node .

[0104] 4. Traversal Neighbors .like exist Not accessed:

[0105] 5. Calculate using the above formula Note that here... It includes nonlinear resistance.

[0106] 6. Calculate the heuristic cost (Euclidean distance from the current point to the destination).

[0107] 7. Update The parent node is ,Will Add to the queue.

[0108] 4. Output Results: After the search reaches its endpoint, a sequence of pixel coordinates is obtained by backtracking the parent node pointer. Due to the nonlinear design of the cost function, this path It automatically avoids low-potential backgrounds and adheres to the center line of high-potential cracks like a magnet. It can also intelligently cross tiny breakpoints in the middle of cracks (because although crossing breakpoints is costly, it is still less costly than going around a large section of background), thus achieving high-precision topology integrity extraction.

[0109] Through the detailed steps described above, this embodiment fully implements all the technical solutions described in the claims, and significantly improves the robustness of photovoltaic crack detection while ensuring computational efficiency.

[0110] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A two-stage hybrid heuristic method for solar cell crack path detection characterized in that, Includes the following steps: (a) obtaining an electroluminescent image , denotes the grey value of the pixel at the position of the coordinates (x, y); (b) Based on the electroluminescent image A crack potential map is generated through nonlinear morphological image processing. , wherein The pixel value in the image represents the potential for a crack to exist at that pixel location. (c) dividing the crack potential map into a number of ROI grids, wherein is an integer greater than 1;​​​ (d) using the first heuristic algorithm, searching and determining a globally optimized ROI path in a grid space composed of the ROI grid which is a spatially continuous ROI grid index sequence;​​ (e) determining a global optimized ROI path based on the global optimized ROI paths , the global optimized ROI path is determined based on the global optimized ROI paths , the global optimized ROI path is determined based on the global optimized ROI paths , the global optimized ROI path (f) Employing the second heuristic algorithm in the constrained pixel search space Internally, a pixel-level optimal crack path is searched and determined based on the principle of minimizing movement cost. The cost of the movement is related to the The pixel potential values ​​in the data exhibit a non-linear negative correlation.

2. The method according to claim 1, characterized in that, Step (b) generates the crack potential map. include: (b1) The electroluminescent image The application has a first structural element Morphological closing operations are used to obtain an approximate background image. ; (b2) By calculation The black hat transformation diagram is obtained. The It highlights the dark features in the image; (b3) The black hat transformation diagram The application has a second structural element Morphological opening operations are used to obtain crack feature enhancement maps. ; (b4) The above The crack potential map is obtained by performing normalization. .

3. The method according to claim 2, characterized in that, The first structural element The size of the second structural element is greater than the preset maximum crack width and less than the scale of the background brightness change. It is a linear structuring element.

4. The method according to claim 1, characterized in that, The first heuristic algorithm is a genetic algorithm, which introduces neighborhood constraints during initialization and evolution to ensure the spatial continuity of the ROI grid index sequence.

5. The method according to claim 4, characterized in that, The genetic algorithm uses chromosomes. Representing a sequence of ROI paths, the fitness function of the genetic algorithm. It is determined to be a weighted combination of at least the following three: (i) The chromosome The average maximum crack potential of the included ROI mesh sequence ; (ii) The chromosomes The reciprocal of the total length of the path represented ; (iii) The chromosome The reciprocal of the smoothness of the represented path .

6. The method according to claim 5, characterized in that, The , and Calculated using the following formula: ， ， ， in, For the chromosome The number of ROI grids in the middle For the first One ROI grid, for The coordinates of the center point, for and The Euclidean distance between them To preset the maximum path length, For vector sum vector exist The angle formed at the point; when hour, Let it be a constant of 1.

7. The method according to claim 1, characterized in that, In step (e), the region expansion process is a morphological dilation operation, and the application has a preset width. structural elements .

8. The method according to claim 1, characterized in that, The second heuristic algorithm is the A* algorithm.

9. The method according to claim 8, characterized in that, The total cost function of the A* algorithm ,in For the current pixel node, for The parent node, From the starting point to The cumulative movement cost, From The heuristic cost to the destination, the Calculated using the following formula: ， in, for and pixel distance between for The potential value of the point on the crack potential map. A preset constant is used to prevent division by zero; This is the cost scaling factor, and its value range is... .

Images