Light-weight photovoltaic module infrared defect detection method

Abstract

The invention discloses a light-weight photovoltaic module infrared defect detection method, which comprises the following steps of: 1, acquiring a private infrared thermal image data set by matching an unmanned aerial vehicle with an infrared camera, and acquiring a public infrared thermal image data set on a Roboflow website; step 2, marking defects for the private infrared thermal image data set, performing data enhancement, and dividing the private infrared thermal image data set into a training set, a verification set and a test set; 3, introducing a spatial perception and local enhancement convolution module (SLConv), a spatial attention and dynamic modulation convolution module (SDConv) and an image enhancement mask module (IEM) on the basis of the YOLOv11; 4, obtaining a lightweight high-precision model through private data set training; and step 5, verifying the robustness on the public data set, and deploying a fixed infrared acquisition detection system to verify the generalization of the model so as to realize real-time intelligent monitoring of the defects of the photovoltaic power station.

Description

TECHNICAL FIELD

[0001] The present application relates to photovoltaic module defect detection technology, in particular to a light photovoltaic module infrared defect detection method. BACKGROUND

[0002] In recent years, with the growing demand for clean energy worldwide, photovoltaic power generation systems have been widely used. However, during long-term operation, photovoltaic modules are susceptible to environmental factors, manufacturing defects, and aging, resulting in defects such as micro-cracks, virtual soldering of solder joints, hidden cracks, and hot spots. These defects not only reduce power generation efficiency, but also can cause local overheating of the module, posing a safety hazard such as fire. Therefore, developing efficient and accurate photovoltaic module defect detection is of great significance to ensure the safe and stable operation of photovoltaic power stations.

[0003] Infrared thermal imaging technology is widely used in defect detection because it can directly reflect the temperature distribution on the surface of the photovoltaic module under non-contact conditions. Compared with visible light imaging, infrared images have better stability under changing light conditions and can effectively reveal thermal anomaly areas. However, infrared images generally have insufficient texture information, low contrast, and are sensitive to noise, especially when detecting small-scale defects or low-contrast defects, the accuracy and robustness of traditional image processing methods and general object detection algorithms are still insufficient.

[0004] In recent years, deep learning technology has made significant progress in object detection, among which the YOLO series has attracted widespread attention due to its high detection accuracy and real-time performance. However, when directly applying the standard YOLO model to photovoltaic infrared defect detection, problems such as small target missed detection, blurred edge features, and sensitivity to complex background interference may occur. Therefore, there is an urgent need for a photovoltaic module infrared defect detection method and system that can enhance feature extraction ability and improve small target detection accuracy under low texture and low contrast conditions, and is suitable for real-time deployment.

[0005] The information provided in the background section merely helps to understand the overall background of the application and should not be considered as admitting or implying that these contents have become prior art to the present application. SUMMARY

[0006] The present application aims to overcome the problems of high small target missed detection rate, sensitivity to complex background, and insufficient real-time performance in photovoltaic module infrared defect detection in the prior art, and proposes a light photovoltaic module infrared defect detection method and system to achieve high-precision and real-time detection of low-texture, low-contrast, and small-scale defects.

[0007] To achieve the above-mentioned purpose, the present application provides a light photovoltaic module infrared defect detection method, comprising the following steps:

[0008] Step 1: Collect private infrared thermal image dataset with UAV and infrared camera, and obtain public infrared thermal image dataset from Roboflow website;

[0009] Step 2: Label defects on private infrared thermal image dataset, perform data augmentation, and divide training set, validation set and test set;

[0010] Step 3: Introduce spatial perception and local enhancement convolution module (SLConv), spatial attention and dynamic modulation convolution module (SDConv) and image enhancement mask module (IEM) based on YOLOv11;

[0011] Step 4: Obtain lightweight and high-precision model through private dataset training;

[0012] Step 5: Verify robustness on public dataset and deploy in fixed infrared detection system to verify model generalization.

[0013] Preferably, in the above technical solution, the private infrared thermal image data in step 1 is collected by an infrared thermal imager carried by a UAV at a photovoltaic power station site, and the public dataset is selected from an online public data platform with photovoltaic module infrared defect annotation information.

[0014] Preferably, in the above technical solution, when labeling defects on the private infrared thermal image dataset in step 2, LabelMe tool is used to complete the labeling work, and the defect types include two types of hot spots (Hot Spots) and partial hot spots (Partial Hot Spots), wherein hot spots refer to the presence of concentrated high temperature in local area, and partial hot spots refer to the presence of temperature anomaly but not obvious high temperature concentration area.

[0015] Preferably, the data augmentation of step 2 includes random rotation in the range of -45° to +45°, and random brightness adjustment in the range of -25% to +25%.

[0016] Preferably, in step 2, the collected infrared thermal image dataset is divided into training set, validation set and test set, wherein the number of images in the training set, validation set and test set is 11400, 119 and 80 respectively.

[0017] Preferably, in step 3, the SLConv module includes a main convolution branch, a channel attention branch and a local detail branch: the main convolution branch is used to extract global features, the channel attention branch is used to highlight hot anomaly area, and the local detail branch is used to capture subtle texture features; three features are fused through learnable weight, so as to enhance edge structure and detail features. The calculation process of this module is as follows:

[0018] F = BN (Conv (X))

[0019] F sse = F Conv (Conv (F)) 1×1 (GAP (F))

[0020] F local = F BN (DWConv (F)) 3×3 (GAP (F))

[0021] F SL = F SiLU (ω1 F + ω2 F + ω3 F) sse local

[0022] where X is the input of the module, F SL is the output of the module. Conv(·) represents the standard two-dimensional convolution operation, BN(·) represents batch normalization, GAP(·) represents global average pooling, Conv 1×1 (·) represents a 1x1 convolution, DWConv 3×3 (·) represents a 3x3 depthwise separable convolution, σ(·) is a Sigmoid activation function, ω1, ω2, ω3 are learnable fusion weights normalized by softmax to meet SiLU(·) is a Sigmoid weighted linear unit activation function.

[0023] Preferably, in step 3, the SDConv module introduces a dynamic feature modulation mechanism combining channel attention and spatial attention, generates an adaptive attention mask using global channel statistics and performs residual fusion to improve the attention degree to small-scale defect regions; its calculation process is as follows:

[0024] F att = F SA (CA (F))

[0025]

[0026] F SD = F + λ (M c ⊙ F att )

[0027] where F is the input of the module, F SD is the output of the module. CA(·) represents a channel attention module. SA(·) represents a spatial attention module. represents the global average pooling response of the cth channel. α is a learnable nonlinear scaling factor. w c ,b c are learnable weight and bias parameters for each channel.

[0028] M c ∈ [0, 1]​​B×C×1×1 is the learned attention mask. Lambda is the fusion coefficient for balancing attention enhanced features. Synbol represents element-wise multiplication.

[0029] Preferably, in step 3, the IEM module guides the network to focus on the hot anomaly area in the shallow feature stage, and the calculation formula is:

[0030] M = sigma (Conv 3×3 (X))

[0031] X' = M * (beta * X) + (1-M) * (alpha * X)

[0032] Where M is in [0, 1] B×1×H×W is the learned attention mask. Lambda is the fusion coefficient for balancing attention enhanced features. Synbol represents element-wise multiplication.

[0033] Preferably, in step 4, the model training batch size is 80, the training period is 100, the default hyperparameter setting of YOLOv11 is adopted, and the precision, recall, mAP@50 and mAP@50:95 four indexes are selected in the evaluation stage to evaluate the performance.

[0034] Preferably, in step 5, the fixed infrared detection system includes an infrared thermal imaging acquisition device, a data processing terminal and a detection platform, the infrared thermal imaging acquisition device is used to acquire infrared image data of the photovoltaic module, the data processing terminal is used to execute the defect detection algorithm and output the detection result, and the detection platform is used for equipment support and environmental adaptation.

[0035] Compared with the prior art, the present application has the following beneficial effects

[0036] The photovoltaic module infrared defect detection method based on the improved YOLOv11 structure, innovatively proposes the SLConv module, which is composed of a backbone convolution branch, a channel attention branch (SSE) and a local detail enhancement branch, can perform multi-scale and multi-channel fusion extraction on the edge structure and detail features of hot spots and local hot spot areas in the infrared thermal image. By introducing the learnable branch fusion weight, the spatial perception ability and edge contour capture ability of the model to the small defects are effectively enhanced, and the background noise interference is suppressed.

[0037] The photovoltaic module infrared defect detection method based on the improved YOLOv11 structure, innovatively proposes the SDConv module, which combines channel attention and spatial attention mechanisms, and introduces a dynamic feature modulation strategy, generates an adaptive attention mask according to global channel statistical information, and optimizes the feature map by weighting, thereby strengthening the response of small-scale and low-contrast defect areas and suppressing the interference of complex background. This module can significantly improve the recall rate and detection accuracy when detecting difficult-to-detect targets such as partial hot spots.

[0038] The present application is based on an improved YOLOv11 structure photovoltaic module infrared defect detection method, which innovatively proposes an IEM module. The module introduces a learnable foreground enhancement and background suppression mechanism in the shallow layer of the network. Through two learnable coefficients, the response of the potential defect area is enhanced and the response of the background area is attenuated, so as to highlight the abnormal hot area and suppress the redundant background signal in the early stage of feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a general flow chart of a lightweight photovoltaic module infrared defect detection method according to the present application;

[0040] Figure 2 is a SSD-YOLOv11 model framework diagram;

[0041] Figure 3 is a SLConv module structure diagram;

[0042] Figure 4 is a SDConv module structure diagram;

[0043] Figure 5 is an IEM module structure diagram;

[0044] Figure 6 is a comparison diagram of images after data enhancement;

[0045] Figure 7 is a comparison diagram of four evaluation indexes of the method of the present application and other methods on a private data set

[0046] Figure 8 is a fixed infrared detection system display diagram DETAILED DESCRIPTION

[0047] The embodiments of the present application will be specifically described below in combination with the drawings. It should be particularly pointed out that the protection scope of the present application is not limited to these specific embodiments.

[0048] Unless otherwise specifically stated, the terms "comprise", "comprises", "comprising", "include", "includes", "including", and the like used in the specification and claims, such as "contain", "contains", "containing", and the like, are to be construed as encompassing the components or parts listed thereafter, and do not exclude other components or parts not listed.

[0049] As shown in Figures 1 to 8 A lightweight photovoltaic module infrared defect detection method according to the specific embodiment of the present application includes the following steps:

[0050] Step 1: Collect private infrared thermal image data sets by unmanned aerial vehicle with infrared camera, and obtain public infrared thermal image data sets on the Roboflow website;

[0051] Infrared thermal image dataset was obtained by DJI drone combined with Hikvision infrared camera in Changji City, Xinjiang

[0052] Step two, label the defects of private infrared thermal image dataset, and divide the training set, validation set and test set after data enhancement;

[0053] LabelMe tool was used for data annotation, and three experienced photovoltaic power station maintenance workers participated in the annotation process, such as Figure 6 Data enhancement includes random rotation in-45° to +45° range and random brightness adjustment in-25% to +25% range, and the number of images in training set, validation set and test set is 11400, 119 and 80 respectively.

[0054] Step three, introduce spatial perception and local enhancement convolution module (SLConv), spatial attention and dynamic modulation convolution module (SDConv) and image enhancement mask module (IEM) based on YOLOv11;

[0055] (1) As shown in Figure 2 , IEM module is added before the backbone network, SLConv replaces b0, b1, b3 and b5 layers in the backbone network, and DConv module replaces b7 layer in the backbone network

[0056] (2) As shown in Figure 3 , we proposed SLConv module, which includes main convolution branch, channel attention branch and local detail branch: main convolution branch is used to extract global features, channel attention branch is used to highlight thermal anomaly area, and local detail branch is used to capture subtle texture features; Three features are fused through learnable weights, so as to enhance edge structure and detail features. The calculation process of this module is as follows:

[0057] F = BN (Conv (X))

[0058] F sse = F·σ(Conv 1×1 (GAP(F)))

[0059] F local = ReLU (BN (DWConv 3×3 (F)))

[0060] F SL = SiLU (ω1·F+ω2·F sse +ω3·F local )

[0061] Where X is the input of the module, F SLThis is the output of the module. `Conv(·)` represents the standard 2D convolution operation, `BN(·)` represents batch normalization, and `GAP(·)` represents global average pooling. 1×1 (·) represents a 1×1 convolution, DWConv 3×3 (·) denotes a 3×3 depthwise separable convolution, σ(·) is the Sigmoid activation function, and ω1, ω2, ω3 are learnable fusion weights, normalized by softmax to satisfy... SiLU(·) is the Sigmoid weighted linear unit activation function.

[0062] (3) Figure 4 As shown, we propose the SDConv module, which introduces a dynamic feature modulation mechanism combining channel attention and spatial attention. It utilizes global channel statistics to generate adaptive attention masks and performs residual fusion to enhance attention to small-scale defect regions. The computation process is as follows:

[0063] F att =SA(CA(F))

[0064]

[0065] F SD =F+λ·(M c ⊙F att )

[0066] Where F is the input of the module, F SD This is the output of the module. CA(·) represents the channel attention module. SA(·) represents the spatial attention module. This represents the global average pooling response of the c-th channel. α is a learnable nonlinear scaling factor. c ,b c These are the learnable weights and bias parameters for each channel.

[0067] M c ∈[0,1] B×C×1×1 This is the learned attention mask. λ is the fusion coefficient used to balance the attention-enhancing features. ⊙ denotes element-wise multiplication.

[0068] (3) Figure 5 As shown, we propose an IEM module that guides the network to focus on thermal anomaly regions during the shallow feature stage. Its calculation formula is as follows:

[0069] M=σ(Conv 3×3 (X))

[0070] X'=M·(β·X)+(1-M)·(α·X)

[0071] Where M∈[0,1]B×1×H×W The learned guiding mask is denoted by α and β, which are scalar coefficients, and X' is the modulated feature map.

[0072] Step 4: Train a lightweight, high-precision model using a private dataset;

[0073] The model training batch size is 80, the training epochs are 100, and the default hyperparameter settings of YOLOv11 are used, such as... Figure 7 As shown, in the evaluation phase, four metrics were selected for performance evaluation: precision, recall, mAP@50, and mAP@50:95. The model achieved scores of 0.694, 0.665, 0.680, and 0.331 on the four metrics, respectively, all of which were better than the comparison model.

[0074] Step 5: Verify robustness on a public dataset and deploy the model on a fixed infrared sampling and detection system to verify its generalization ability.

[0075] like Figure 8 As shown, the fixed infrared detection system consists of three parts: an infrared thermal imaging acquisition device, a data processing terminal, and a detection platform. The infrared thermal imaging acquisition device is used to acquire infrared image data of photovoltaic modules, the data processing terminal is used to execute defect detection algorithms and output detection results, and the detection platform is used for equipment support and environmental adaptation.

[0076] The foregoing description of specific exemplary embodiments of the present invention is primarily for illustration and demonstration, and not intended to limit the invention to the specific forms disclosed. Based on this, it is evident that those skilled in the art can make various adjustments or improvements based on these teachings. The purpose of selecting and describing these specific embodiments is to clearly demonstrate the core ideas of the present invention and its practical applications, so that those skilled in the art can implement various different exemplary solutions based on the present invention, and make various adjustments and optimizations. The scope of protection of the present invention is defined by the appended claims and their equivalent variations.

Claims

1. A lightweight infrared defect detection method for photovoltaic modules, characterized in that, Comprise: Step 1, private infrared thermal image data set is collected by unmanned aerial vehicle equipped with infrared camera, and public infrared thermal image data set is obtained from Roboflow website; Step 2, the private infrared thermal image data set is labeled with defects, and the data is enhanced and divided into training set, verification set and test set; Step 3, introduce spatial perception and local enhancement convolution module (SLConv), spatial attention and dynamic modulation convolution module (SDConv) and image enhancement mask module (IEM) based on YOLOv11; Step 4, obtain a lightweight and high-precision model through private data set training; Step 5, verify the robustness on public data set, and deploy in fixed infrared detection system to verify the generalization of the model.

2. The method of claim 1, wherein the method is used for detecting infrared defects of a lightweight photovoltaic module. The private infrared thermal image data in step 1 is collected by unmanned aerial vehicle equipped with infrared thermal imager, and the public data set is obtained from online data platform.

3. The method of claim 1, wherein the method is characterized by: Step 2, the private infrared thermal image data set is labeled with defects: the data labeling work is completed by using LabelMe tool, and the defects are divided into two categories: hot spot and local hot spot.

4. The method of claim 1, wherein the method is used for infrared defect detection of a lightweight photovoltaic module. The data enhancement in step 2 includes random rotation in the range of-45° to +45°, and random brightness adjustment in the range of-25% to +25%.

5. The method of claim 1, wherein the method is used for infrared defect detection of a lightweight photovoltaic module. Step 2, the training set, verification set and test set are divided: the images of training set, verification set and test set are 11400, 119 and 80 respectively.

6. The method of claim 1, wherein the method is used for infrared defect detection of a lightweight photovoltaic module. Step 3, SLConv module is proposed, which convolves and batch normalizes the input feature map to obtain the basic feature; the channel attention coefficient is generated by global average pooling and one-dimensional convolution, and is multiplied with the basic feature element by element to obtain the channel attention enhanced feature; at the same time, the basic feature is subjected to depth separable convolution, batch normalization and ReLU activation to obtain the local detail feature; the basic feature, channel attention enhanced feature and local detail feature are weighted and summed according to the learnable weight, and the module output feature is obtained through SiLU activation function.

7. The method of claim 1, wherein, Step 3, SDConv module is proposed, which sequentially processes the input feature map through channel attention and spatial attention to obtain enhanced attention feature; Calculate the global average response value of each channel in the spatial dimension, and perform nonlinear scaling and weighted bias operation on the response value to obtain the channel level attention mask; The mask is elementarily weighted and fused with the enhanced feature, and the module output feature is obtained by adding the residual connection and the original input feature.

8. The method of claim 1, wherein the method is used for infrared defect detection of a lightweight photovoltaic module. Step 3, IEM module is proposed, which convolves the input feature map, generates a guide mask with a value range of 0 to 1 through Sigmoid function; the guide mask is used to weight and fuse the input feature scaled by the first scalar coefficient and the input feature scaled by the second scalar coefficient, respectively, to enhance the response of potential defect area and suppress the response of background area in the feature layer, and obtain the module output feature.

9. The method of claim 1, wherein the method is used for infrared defect detection of a lightweight photovoltaic module. Step 4, the model training includes: (1) the training batch size is 80, and the training process is 100 cycles; (2) YOLOv11 default hyperparameter setting is adopted; (3) four indicators are used for model evaluation: precision, recall, mAP@50 and mAP@50:

95.

10. The method of claim 1, wherein the method is used for infrared defect detection of a lightweight photovoltaic module. The step 5 fixed infrared acquisition detection system is composed of an infrared thermal imaging acquisition device, a data processing terminal and a detection platform.

Images