Methods for Identifying and Locating Pitting and Cracks in Deep-Sea Submarine Solar Photovoltaic Panels

Abstract

This invention discloses a method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels. The method uses acquired images of the photovoltaic panel surface and a constructed model for identifying pitting corrosion and cracks. The model's loss function consists of three parts: edge detection loss with weighted coefficients, segmentation loss, and generalized Dice loss. A continuous adaptive gradient flow is introduced, and a stability hyperparameter is used to update the model's training parameters. When locating pitting corrosion and cracks, a sliding detection box method is used to detect images at multiple scales. At each detection box location, the trained model classifies the image within the detection box to obtain candidate detection boxes containing pitting corrosion and cracks. After removing redundancy and overlap, the detection box most likely representing the true damage is retained, thus determining the location of the damage.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection technology, specifically relating to a method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels. Background Technology

[0002] With the increasing demand for clean energy, the application of offshore solar photovoltaic (PV) panels is becoming increasingly widespread. However, due to factors such as corrosion prevention, these panels are typically installed at a high height, making direct manual inspection by maintenance vessels impossible. The harsh marine environment causes various types of damage to the surface of these panels, particularly pitting and cracking, severely impacting their lifespan and power generation efficiency. Because of the need for corrosion protection, offshore PV panels are usually installed at a high position, preventing direct manual inspection by maintenance vessels. This urgently necessitates an automated inspection technology capable of accurately identifying and locating defects such as pitting and cracks on the surface of PV panels to ensure their normal operation and power generation efficiency.

[0003] Existing methods for inspecting marine solar panels rely on unmanned vessels or drones equipped with cameras for image acquisition, followed by manual or simple algorithmic image processing. However, these methods still suffer from insufficient recognition accuracy and inaccurate positioning. To improve inspection accuracy and work efficiency, a new image recognition and positioning method is urgently needed. Summary of the Invention

[0004] The present invention aims to overcome the shortcomings of the prior art by providing a method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels.

[0005] This invention is achieved through the following technical solution:

[0006] A method for identifying and locating pitting corrosion and cracks in deep-sea submersible solar photovoltaic panels, the specific steps of which are as follows:

[0007] Step 1: Acquire images of the photovoltaic panel surface;

[0008] Step 2: Perform data cleaning, normalization, standardization, labeling, missing value handling, and class balancing preprocessing on the dataset of photovoltaic panel surface images collected in Step 1.

[0009] Step 3: Construct a model for identifying pitting and crack defects in photovoltaic panels;

[0010] Step 3.1: Construct the loss function. The total loss function consists of three parts with weighted coefficients: edge detection loss, segmentation loss, and generalized Dice loss function.

[0011] Step 3.2: Design the optimizer: Introduce a continuous adaptive gradient flow and combine it with stability hyperparameters to update the model training parameters;

[0012] Step 3.3: Train the model using the training data from Step 2. In each iteration, input the training data into the model, calculate the loss using the loss function from Step 3.1, and update the model parameters using the optimizer from Step 3.2. Continuously adjust the model to reduce the loss and enable the model to learn the patterns and rules in the data.

[0013] Step 4: Locate the pitting and crack positions;

[0014] Step 4.1: Use the sliding detection box method to detect images. At each detection box location, use the model trained in Step 3 to classify the image within the detection box to obtain candidate detection boxes containing pitting and cracks.

[0015] Step 4.2: Scale the image at different ratios to generate multi-scale images, and repeat the sliding detection box detection process in Step 4.1 to detect pitting and crack areas of different sizes.

[0016] Step 4.3: Remove redundant and overlapping detection boxes from the candidate detection boxes obtained in Steps 4.1 and 4.2, and finally retain the detection boxes that are most likely to represent the real damage to determine the location of the damage.

[0017] In the above technical solution, in step 1, a reflector is installed around the photovoltaic panel, and the reflector is mounted on an adjustable angle adjustment mechanism. The angle of the reflector is adjusted by controlling the angle adjustment mechanism. A camera is set below the reflector to receive the image of the photovoltaic panel surface reflected by the reflector.

[0018] In the above technical solution, the edge detection loss It is used to capture pitting and crack edges in images. It is calculated by an edge detection algorithm based on gradient information or edge features of the image. It is used to measure the difference between the gradient of the segmented image edge and the real edge.

[0019] ;

[0020] in, The model predicts the image at the location. The gradient value at that point; : Real label image at location The gradient value at that point.

[0021] In the above technical solution, the segmentation loss The pitting and crack segmentation capabilities of images are optimized by calculating the difference between the predicted image and the ground truth label.

[0022] ;

[0023] in, : Real label image at location The value at; The model predicts the image at the location. The probability value at that location.

[0024] In the above technical solution, the generalized Dice loss It is used to handle imbalanced classification problems and measures the overlap between the predicted and actual regions.

[0025] ;

[0026] in, The model predicts the image at the location. Pixel values; : Real label image at location Pixel values; : Summation of all pixel positions in the image.

[0027] In the above technical solution, step 3.2 includes the following steps:

[0028] Step 3.2.1: Define the gradient update rule:

[0029] ;

[0030] In time Gradient squared update;

[0031] The exponential decrease coefficient scheduling controls the contribution of previous estimates to the current estimate. It is a vector-valued function with a value range of [value range missing]. Between, and defined in time superior;

[0032] Current parameter The square of the gradient of the loss function;

[0033] :time The cumulative estimate of the squared gradient;

[0034] Step 3.2.2: Define the second-time update rules:

[0035] ;

[0036] :time The cumulative estimate of the squared parameter update;

[0037] In time For parameters The update step size is calculated using the formula in step 3.2.3;

[0038] :time The cumulative estimate of the squared parameter update;

[0039] Step 3.2.3: Define parameter update rules:

[0040] ;

[0041] :time Update the parameters and apply them to the current parameters. ;

[0042] Numerical stability parameter to prevent numerical problems such as division by zero during updates;

[0043] Numerical stability term, used to stabilize gradient flow;

[0044] Current parameter The gradient of the loss function is used to calculate the update step size.

[0045] In the above technical solution, in step 4.3, for multiple candidate detection boxes... , The total number of detection boxes, each with a confidence score. Confidence score Based on the probability value of the detection box The calculation involves using a non-maximum suppression algorithm to remove redundant and overlapping detection boxes based on a set threshold. The process is as follows:

[0046] ①. Based on confidence score Sort the detection boxes from largest to smallest;

[0047] ②. Select the detection box with the highest confidence level. As a reserved frame;

[0048] ③. Calculation Intersection over Union (IoU) with other detection boxes ;

[0049] ④. If Greater than the set threshold Remove the detection box. ;

[0050] ⑤. Repeat steps ② to ④ until all detection boxes have been processed.

[0051] In the above technical solution, the positioning results are recorded and stored, and the data is transmitted to the monitoring center or relevant maintenance personnel's equipment via wireless communication so that repair and maintenance work can be arranged in a timely manner.

[0052] The advantages and beneficial effects of this invention are as follows:

[0053] This invention automates the inspection of offshore solar photovoltaic panels. By periodically collecting images and positioning results, it can determine the location of damage such as pitting and cracks on the surface of the photovoltaic panels and issue maintenance alerts in advance to help maintenance personnel carry out maintenance work in a timely manner. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to the present invention.

[0055] For those skilled in the art, other related figures can be obtained from the above figures without any creative effort. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below with reference to specific embodiments.

[0057] This invention presents a method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels. The specific steps are as follows:

[0058] Step 1: Acquire images of the photovoltaic panel surface.

[0059] In this embodiment, a reflector combined with a camera is used to acquire images of the photovoltaic panel surface. Specifically, a reflector is installed around the photovoltaic panel, and the reflector is mounted on an adjustable angle adjustment mechanism. By controlling the angle adjustment mechanism, the angle of the reflector can be adjusted, allowing the reflector to capture images from different angles as needed, ensuring coverage of the entire photovoltaic panel area.

[0060] A camera is installed below the reflector to receive images of the photovoltaic panel surface reflected by the reflector.

[0061] The program is set up to acquire images at regular intervals, and the acquisition frequency is determined according to actual needs, such as acquiring images once per hour or at a specific time each day. This method allows for the timely acquisition of images of the photovoltaic panel surface without contact with the panel, facilitating regular inspections.

[0062] Step 2: Perform preprocessing operations such as data cleaning, normalization, standardization, labeling, handling missing values, and class balancing on the dataset of photovoltaic panel surface images collected in Step 1 to ensure data quality and usability, making the data suitable for model training.

[0063] Step 3: Construct a model for identifying pitting and crack defects in photovoltaic panels.

[0064] This invention introduces an improved loss function and optimizer, which facilitates model training and convergence. The specific steps are as follows.

[0065] Step 3.1: Construct the loss function.

[0066] The total loss function consists of three parts with assigned weights: edge detection loss, segmentation loss, and generalized Dice loss function;

[0067] ;

[0068] The total loss function represents the final objective of model optimization.

[0069] Edge detection loss, : Segmentation loss, Generalized Dice loss;

[0070] The weighting coefficients of the edge detection loss. : Weighting coefficients of the segmentation loss : Weighting coefficients of Dice loss.

[0071] The edge detection loss This is used to capture pitting and crack edges in an image. It is calculated using edge detection algorithms (such as the Sobel operator, Canny operator, etc.) based on gradient information or image edge features, and is used to measure the difference between the gradient of the segmented image edge and the real edge.

[0072] ;

[0073] in, The model predicts the image at the location. The gradient value at that point; : Real label image at location The gradient value at that point.

[0074] The segmentation loss The pitting and crack segmentation capabilities of images are optimized by calculating the difference between the predicted image and the ground truth label.

[0075] ;

[0076] in, : Real label image at location The value at that location (0 or 1, indicating whether it is a pitting or crack area); The model predicts the image at the location. The probability value at the location (representing the probability of being predicted as a pitting or cracked area).

[0077] The generalized Dice loss This is used to handle imbalanced classification problems. It can effectively measure the overlap between the predicted and the true regions, and performs better, especially when the target region is small or sparse.

[0078] ;

[0079] in, The model predicts the image at the location. The pixel value (0 or 1, indicating whether it is a pitting or crack area); : Real label image at location The pixel value (0 or 1, indicating whether it is a pitting or crack area); : Summation of all pixel positions in the image.

[0080] Step 3.2: Constructing the optimizer: By introducing a continuous adaptive gradient flow and combining it with stability hyperparameters, the model training parameters are updated. The specific steps are as follows.

[0081] Step 3.2.1: Define the gradient update rule:

[0082] ;

[0083] In time Gradient squared update;

[0084] The exponential decrease coefficient scheduling controls the contribution of previous estimates to the current estimate. It is a vector-valued function with a value range of [value range missing]. Between, and defined in time superior;

[0085] Current parameter The square of the gradient of the loss function;

[0086] :time The cumulative estimate of the squared gradient.

[0087] Step 3.2.2: Define the second-time update rules:

[0088] ;

[0089] :time The cumulative estimate of the squared parameter update;

[0090] In time For parameters The update step size is calculated using the formula in step 3.2.3;

[0091] :time The cumulative estimate of the squared parameter update.

[0092] Step 3.2.3: Define parameter update rules:

[0093] ;

[0094] :time Update the parameters and apply them to the current parameters. ;

[0095] : Numerical stability term, to prevent numerical problems such as division by zero (usually a very small positive number) during updates;

[0096] : Numerical stability term, used to stabilize gradient flow (usually a very small positive number);

[0097] Current parameter The gradient of the loss function is used to calculate the update step size.

[0098] Step 3.2: After building the above model, train the model using the training data from Step 2. In each iteration, input the training data into the model, calculate the loss using the loss function from Step 3.1, and update the model parameters using the optimizer from Step 3.2. Continuously adjust the model to reduce the loss and enable the model to learn the patterns and rules in the data.

[0099] Step 4: Locate the pitting and crack locations.

[0100] Step 4.1: Use the sliding detection box method to detect the image and determine whether there is pitting or cracks.

[0101] First, set an appropriate detection box size and movement step size. Starting from the top left corner of the image, slide across the image with a certain step size. At each detection box position, use the model trained in step 3 to classify the image within the detection box and determine whether there is pitting or cracks.

[0102] Step 4.1.1: Position the sliding detection box.

[0103] Assuming the image size is The detection box size is ,in It is the height of the detection frame. It is the brightness of the detection frame, with a step size of 1. and Then the current position of the sliding detection box This can be expressed by the following formula:

[0104] ;

[0105] : The coordinates of the top left corner of the detection box;

[0106] : The index of the current detection box position;

[0107] : Sliding step size in the horizontal and vertical directions;

[0108] : Height and width of the detection frame.

[0109] Step 4.1.2: Classify the image within the detection box.

[0110] Position of each sliding detection box The trained defect recognition model is used to classify images and outputs the probability value of whether they contain pitting or cracks. This value is expressed by the following formula:

[0111] ;

[0112] Detection box The classification probability of a location indicates whether the area contains pitting or cracks;

[0113] : A well-trained defect recognition model;

[0114] Current detection box position The image region above, with a size of .

[0115] probability value Detection boxes with a probability greater than the set probability threshold are considered candidate detection boxes with pitting or cracks.

[0116] Step 4.2: Scale the image at different ratios to generate multi-scale images. Repeat the sliding detection box process from Step 4.1 to detect pitting and crack regions of different sizes. This operation effectively enhances the model's accuracy and robustness in detecting damage of different sizes, thereby improving its ability to detect pitting and crack regions of varying sizes.

[0117] Step 4.3: Remove redundant and overlapping detection boxes from the candidate detection boxes obtained in Steps 4.1 and 4.2, and finally retain the detection boxes that are most likely to represent the real damage, thereby determining the location of the damage.

[0118] For multiple candidate detection boxes , The total number of detection boxes, each with a confidence score. Confidence score Probability value of the detection box The calculation involves using a non-maximum suppression algorithm to remove redundant and overlapping detection boxes based on a set threshold. The process is as follows:

[0119] ①. Based on confidence score Sort the detection boxes from largest to smallest;

[0120] ②. Select the detection box with the highest confidence level. As a reserved frame;

[0121] ③. Calculation Intersection over Union (IoU) with other detection boxes ;

[0122] ④. If Greater than the set threshold Remove the detection box. ;

[0123] ⑤. Repeat steps ② to ④ until all detection boxes have been processed.

[0124] The intersection-union ratio (IUGR) is used to measure the degree of overlap between two detection boxes, and its calculation formula is as follows:

[0125] ;

[0126] Detection box and The crossover ratio between them;

[0127] Detection box and The area of ​​their intersection;

[0128] Detection box and The area of ​​the union of the sets.

[0129] Finally, the location results are recorded and stored, and the data can be transmitted to the monitoring center or relevant maintenance personnel's equipment via wireless communication or other means, so that repair and maintenance work can be arranged in a timely manner.

[0130] Experimental verification

[0131] This invention collected 5000 images, including 3000 images without erosion or cracks and 2000 images with pitting and cracks. These image datasets were randomly divided into training, validation, and test sets, with the test and validation sets each comprising 15% of the dataset (30% total), and the relative proportion of each image type in each subset being the same as the total dataset. The model evaluation metrics are as follows:

[0132] Accuracy: Used to measure the proportion of samples correctly classified by the model, calculated using the following formula:

[0133] ;

[0134] Among them, TP is a true positive example (the model predicts it to be a positive example and it is actually a positive example), TN is a true negative example (the model predicts it to be a negative example and it is actually a negative example), FP is a false positive example (the model predicts it to be a positive example but it is actually a negative example), and FN is a false negative example (the model predicts it to be a negative example but it is actually a positive example).

[0135] F1 score: The harmonic mean of precision and recall. Precision represents the proportion of correctly predicted positive instances, while recall represents the proportion of actual positive instances correctly predicted by the model. The formula for calculating the F1 score is:

[0136] ;

[0137] ;

[0138] ;

[0139] The area under the receiver operating characteristic curve (ROCAUC) is used to evaluate the performance of the model at different classification thresholds. The ROC curve depicts the relationship between the true positive rate (TPR, which is equivalent to recall) and the false positive rate (FPR). The larger the AUC value, the better the model adapts to different thresholds.

[0140] Area under the precision-recall curve (P-RAUC): Similar to ROCAUC, it is used to evaluate the performance of a binary classification model at different thresholds, especially suitable for class imbalance. The P-RAUC curve shows the relationship between precision and recall. The higher the AUC value, the better the model performance.

[0141] ;

[0142] Referring to Table 1, the accuracy of the method presented in this invention is significantly higher than that of the CNN model, indicating that the proposed method performs better in terms of classification correctness. The F1 score of the method presented in this invention is also superior to that of CNN, indicating that it performs better in balancing precision and recall. The ROC AUC of the method presented in this invention is 0.95, higher than that of CNN (0.91), indicating that the method presented in this invention performs better at various classification thresholds. The P-RAUC of the method presented in this invention is 0.92, significantly higher than that of CNN (0.87), indicating that it has stronger detection capabilities under class imbalance conditions.

[0143] The present invention has been described above by way of example. It should be noted that any simple modifications, alterations or other equivalent substitutions that can be made by those skilled in the art without creative effort without departing from the core of the present invention fall within the protection scope of the present invention.

Claims

1. A method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels, characterized in that: Includes the following steps: Step 1: Acquire images of the photovoltaic panel surface; Step 2: Perform data cleaning, normalization, standardization, labeling, missing value handling, and class balancing preprocessing on the dataset of photovoltaic panel surface images collected in Step 1. Step 3: Construct a model for identifying pitting and crack defects in photovoltaic panels; Step 3.1: Construct the loss function. The total loss function consists of three parts with weighted coefficients: edge detection loss, segmentation loss, and generalized Dice loss function. Step 3.2: Design the optimizer: Introduce a continuous adaptive gradient flow and combine it with stability hyperparameters to update the model training parameters; Step 3.2 includes the following steps: Step 3.2.1: Define the gradient update rule: g ′ (t): Gradient squared update at time t; ρ(t): The exponential decrease coefficient scheduler controls the contribution of the previous estimate to the current estimate. It is a vector-valued function with a value range between [0,1] and is defined at time t. The square of the gradient of the loss function under the current parameters w(t); g(t): The cumulative estimate of the squared gradient over time t; Step 3.2.2: Define the second-time update rules: h ′ (t)=(1-ρ(t))(w ′ (t) 2 -h(t)); h ′ (t): The cumulative estimate of the squared parameter update over time t; w ′ (t): The update step size of parameter w(t) at time t, calculated using the formula in step 3.2.3; h(t): The cumulative estimate of the squared parameter updates over time t; Step 3.2.3: Define parameter update rules: w ′ (t): The parameter update at time t is applied to the current parameter w(t); ε: Numerical stability term, to prevent division by zero issues during updates; δ: Numerical stability term, used to stabilize gradient flow; : The gradient of the loss function under the current parameters w(t), used to calculate the update step size; Step 3.3: Train the model using the training data from Step 2. In each iteration, input the training data into the model, calculate the loss using the loss function from Step 3.1, and update the model parameters using the optimizer from Step 3.

2. Continuously adjust the model to reduce the loss and enable the model to learn the patterns and rules in the data. Step 4: Locate the pitting and crack locations; Step 4.1: Use the sliding detection box method to detect images. At each detection box location, use the model trained in Step 3 to classify the image within the detection box to obtain candidate detection boxes containing pitting and cracks. Step 4.2: Scale the image at different ratios to generate multi-scale images, and repeat the sliding detection box detection process in Step 4.1 to detect pitting and crack areas of different sizes. Step 4.3: Remove redundant and overlapping detection boxes from the candidate detection boxes obtained in Steps 4.1 and 4.2, and finally retain the detection boxes that are most likely to represent the real damage to determine the location of the damage.

2. The method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to claim 1, characterized in that: In step 1, a reflector is installed around the photovoltaic panel, and the reflector is mounted on an adjustable angle adjustment mechanism. The angle of the reflector is adjusted by controlling the angle adjustment mechanism. A camera is set below the reflector to receive the image of the photovoltaic panel surface reflected by the reflector.

3. The method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to claim 1, characterized in that: The edge detection loss L edge It is used to capture pitting and crack edges in images. It is calculated by an edge detection algorithm based on gradient information or edge features of the image. It is used to measure the difference between the gradient of the segmented image edge and the real edge. in, The model predicts the gradient value of the image at position (m,n); : The gradient value of the real labeled image at position (m,n).

4. The method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to claim 1, characterized in that: The segmentation loss L seg The pitting and crack segmentation capabilities of images are optimized by calculating the difference between the predicted image and the ground truth label. L seg =-∑ m,n [y gt (m,n)log y pred (m,n)+(1-y gt (m,n))log(1-y pred (m,n))]; Among them, y gt (m,n): The value of the real label image at position (m,n); y pred (m,n): The probability value of the image predicted by the model at position (m,n).

5. The method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to claim 1, characterized in that: The generalized Dice loss L Dice It is used to handle imbalanced classification problems and measures the overlap between the predicted and actual regions. Where A(m,n): the pixel value of the image predicted by the model at position (m,n); : The pixel value of the real label image at position (m,n); ∑ m,n : Summation of all pixel positions in the image.

6. The method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to claim 1, characterized in that: In step 4.3, for multiple candidate detection boxes B1, B2, ..., B Q Q represents the total number of bounding boxes, and each bounding box has a confidence score S(B). q ), confidence score S(B) q The probability value p of the detection box is used for calculation; a non-maximum suppression algorithm is used to remove redundant and overlapping detection boxes based on a set threshold. The process is as follows: ①. According to the confidence score S(B) q Sort the detection boxes from largest to smallest; ②. Select the detection box B with the highest confidence level. q As a reserved frame; ③. Calculate B q Intersection over Union (IoU) with other detection boxes q B j ); ④. If IoU(B q B qq () greater than the set threshold θ IoU Then remove detection box B. qq ; ⑤. Repeat steps ② to ④ until all detection boxes have been processed.

7. The method for identifying and locating pitting corrosion and cracks in deep-sea underwater solar photovoltaic panels according to claim 1, characterized in that: The location results are recorded and stored, and then transmitted to the monitoring center or relevant maintenance personnel's equipment via wireless communication.

Images