Lightweight YOLO network and method for solar panel damage detection

Abstract

The present application relates to a kind of light YOLO network and method for solar panel damage detection.The network includes: backbone network is configured as feature extraction, it includes embedding multiscale dilation convolution C3k2_MDC module, the C3k2_MDC module is configured as expanding receptive field and extracting multiscale feature;Neck network is configured as feature fusion, it includes the C3k2_AFFM module integrated attention mechanism, the C3k2_AFFM module is configured as performing weighted fusion to the multiscale feature;Detection head is configured as based on the feature output damage detection result after fusion, it includes multi-feature sharing detection head MSDH.The network and method proposed in the present application can accelerate convergence speed, reduce parameter quantity, and significantly improve detection precision and operating efficiency.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and spacecraft health monitoring technology, and in particular to a lightweight YOLO network and method for detecting damage to solar panels. Background Technology

[0002] As small satellite missions become increasingly complex, solar panels, as a critical component of their energy systems, are directly affected by their operational status, impacting power supply security. During their time in orbit, solar panels may be subjected to impacts from space debris and electrical short circuits, leading to damage such as surface cracks, delamination, or other issues. In severe cases, this can threaten the stable operation of the entire satellite system. Therefore, conducting damage detection on solar panels is of great significance for ensuring their operational safety.

[0003] In recent years, damage detection methods have been continuously developing. Angeletti et al. proposed a multi-damage detection method based on LSTM networks, combining finite element simulation to generate acceleration and piezoelectric sensor data, which can be used for health monitoring of flexible structures such as solar panels. Iannelli et al. applied deep learning detection methods to large aerospace antennas, using on-orbit attitude data to identify and locate structural damage, providing a new approach for on-orbit condition assessment of flexible components such as solar panels. Xia Zhenxin et al. used laser ultrasonic imaging and multi-feature fusion parameter methods to achieve visualized damage assessment of industrial pure titanium under different plastic deformation states. Although the above methods perform well in damage detection based on physical modeling and sensor signals, their dependence on telemetry data, structural modeling, or experimental devices makes it difficult to meet the needs of rapid and intuitive image-level on-orbit damage detection. In recent years, the concept of edge computing has driven the development of embedded AI chips. These chips are gradually being integrated into small satellite platforms, providing basic computing power support for on-orbit image processing. However, due to limitations in chip resources, it is still difficult to deploy computationally complex networks. Therefore, designing a lightweight, high-precision, and easily deployable solar panel damage detection method has become an important research direction.

[0004] Existing image-level damage detection methods are mainly divided into two categories: traditional image processing methods and deep learning detection methods. The former relies on manually designed image features and detection rules, such as edge extraction, thermal response analysis, and region segmentation, offering good interpretability and computational controllability, but lacking generalization ability in complex backgrounds. In contrast, deep learning-based target detection methods have stronger adaptability and detection accuracy. Tang Xinliang et al. improved the Faster R-CNN network, combining multi-scale fusion and rotation region proposal strategies to enhance the detection performance of bolted parts in complex environments. Li et al. constructed a Res-Mask R-CNN network based on surface strain field information to accurately detect various damages in spacecraft composite laminates, demonstrating good generalization ability. While the aforementioned two-stage detection methods exhibit excellent accuracy, their high computational complexity limits their recognition speed in complex images compared to one-stage detection methods. In the field of target damage detection, one-stage detection methods, represented by the YOLO series, are gaining increasing attention due to their high efficiency and good real-time performance. Pei Xiaofang et al. introduced an improved CBAM attention mechanism and a Bi-FPN feature fusion network into YOLOv5, effectively improving the accuracy and robustness of damage detection in complex scenes. An Mengjun et al. introduced multi-scale attention wavelet convolution and edge convolution attention mechanisms to enhance the network's ability to detect small targets in complex backgrounds without increasing the number of parameters. Su Jun et al. proposed the GLF-YOLOX method, which integrates a dual-branch structure of CNN and Transformer, introduces global and local enhanced attention mechanisms and a Transformer-based detection head, demonstrating outstanding performance in detection accuracy.

[0005] Although the above detection methods have achieved good results, they still have problems in solar panel damage detection, such as slow convergence speed, large number of parameters, rare sample accuracy fluctuations, and redundant overlapping boxes. Summary of the Invention

[0006] The objective of this invention is to provide a lightweight YOLO network and method for solar panel damage detection. Through the network and / or method, the problems of slow convergence speed, large number of parameters, and high computational cost in existing damage detection methods can be solved, thereby accelerating the convergence speed, reducing the number of parameters, and significantly improving detection accuracy and operating efficiency.

[0007] In a first aspect of the invention, the aforementioned task is accomplished by a lightweight YOLO network for solar panel damage detection, the network comprising:

[0008] The backbone network, configured for feature extraction, includes a C3k2_MDC module with embedded multi-scale dilated convolutions, which is configured to expand the receptive field and extract multi-scale features.

[0009] A neck network, configured for feature fusion, includes a C3k2_AFFM module with an integrated attention mechanism, the C3k2_AFFM module being configured to perform weighted fusion of the multi-scale features; and

[0010] The detection head is configured to output damage detection results based on fused features, and the detection head includes a multi-feature shared detection head (MSDH).

[0011] Furthermore, the backbone network also includes a depthwise separable convolution DSConv module and a grouped local convolution GLConv module.

[0012] Furthermore, the DSConv module includes a depthwise convolutional (DWConv) module and a pointwise convolutional (PWConv) module; and / or

[0013] The GLConv module generates the main feature map through standard convolution, performs linear transformation using a 3×3 convolution kernel to generate a redundant feature map, and finally concatenates the main feature map and the redundant feature map to obtain the final output feature map.

[0014] Furthermore, the C3k2_MDC module includes at least two dilated convolution kernels with different dilation rates, including 3 and 5.

[0015] Furthermore, feature fusion is performed using the C3k2_AFFM module, including:

[0016] Channel segmentation is performed on the input feature map to obtain the first branch feature map and the second branch feature map;

[0017] The bottleneck structure is used to process the first branch feature map;

[0018] The second branch feature map is subjected to average pooling downsampling, lightweight self-attention computation, and upsampling restoration operations, and then fused with the processing result of the first branch feature map; and

[0019] The fused features are then fed into a lightweight feedforward enhancement network to perform further feature extraction.

[0020] Furthermore, the lightweight self-attention computation includes:

[0021] The input features are mapped to query vectors, key vectors, and value vectors.

[0022] Q,K,V=Split(Conv 1×1 (X))

[0023] Where Q represents the query vector, K represents the key vector, and V represents the value vector; Split means dividing the feature matrix obtained from the convolution operation according to the channel dimension to generate multiple sub-feature matrices; Conv 1×1 (X) indicates that a convolution operation with a kernel size of 1×1 is performed on the input feature matrix X;

[0024] Calculate the attention score matrix.

[0025]

[0026] Where Attention(Q,K,V) represents the attention score matrix, d k K represents the dimension of the key vector K. T This represents the transpose of the key vector K; softmax represents the normalization operation, converting the correlation calculation results into weight coefficients; and

[0027] The attention output and input features are fused through residual connections.

[0028] Furthermore, the damage detection results based on the fused features include:

[0029] By using a lightweight initial convolution module, the number of channels is unified for the fused features;

[0030] The output feature map is obtained by sharing a convolutional module;

[0031] The bounding box regression is obtained using the shared bounding box regression module; and

[0032] Use the shared classification module to obtain category predictions;

[0033] The damage detection results include the bounding box regression and the category prediction.

[0034] Furthermore, the PIoUv2 loss function is used as the loss function for bounding box regression, wherein the PIoUv2 loss function L PIoUv2 for,

[0035]

[0036] q = e -p ,q∈(0,1],

[0037]

[0038] L PIoUv2 =3m(λq)·L PIoU

[0039] Where dw1 and dw2 represent the absolute difference between the predicted bounding box and the target bounding box in the width direction, dh3 and dh4 represent the absolute difference between the predicted bounding box and the target bounding box in the height direction, and w gt and h gt L represents the width and height of the target bounding box, respectively. IoU L represents the intersection-union loss function between the predicted bounding box and the target bounding box. PIoU Indicates in L IoU The loss function obtained by introducing a penalty factor P on the basis is q, which represents the quality of the anchor box, P represents the penalty factor, m(x) represents the attention function, and λ represents the hyperparameter that controls the behavior of the attention function.

[0040] In a second aspect of the invention, the aforementioned task is also addressed by a method for detecting damage to solar panels, the method comprising the following steps:

[0041] Collect images of the solar panels to be inspected and perform preprocessing;

[0042] The preprocessed image is input into the lightweight YOLO network according to any one of claims 1 to 8 to perform processing; and

[0043] Output the damage detection results.

[0044] In a third aspect of the invention, an electronic device is also provided, comprising the aforementioned lightweight YOLO network for solar panel damage detection.

[0045] The technical solution provided by this invention has the following advantages:

[0046] 1. The lightweight YOLO network for solar panel damage detection proposed in this invention integrates the MDC module through the C3k2 module in the backbone network, which effectively expands the receptive field of the network and reduces the number of parameters; at the same time, the DSConv and GLConv modules are used to replace the standard convolution, which compresses the network parameters and improves the feature extraction capability.

[0047] 2. The lightweight YOLO network for solar panel damage detection proposed in this invention introduces a multi-branch feature fusion module C3k2_AFFM with integrated attention mechanism in the neck structure. Combined with average pooling and upsampling modules, it alleviates the problem of information redundancy and conflict during feature fusion, and improves the extraction efficiency and accuracy of damage target information.

[0048] 3. The lightweight YOLO network for solar panel damage detection proposed in this invention features a shared detection head MSDH. By sharing convolution parameters, it improves parameter utilization efficiency while significantly reducing the number of parameters.

[0049] 4. The lightweight YOLO network for solar panel damage detection proposed in this invention employs a dynamically adjustable PIoUv2 damage function to enhance attention to medium-quality anchor frames, alleviate the problems of gradient saturation of high-quality anchor frames and information redundancy of low-quality anchor frames, and accelerate the convergence speed of the network. Attached Figure Description

[0050] To further illustrate the above and other advantages and features of the various embodiments of the present invention, a more specific description of the various embodiments of the present invention will be presented with reference to the accompanying drawings. It is to be understood that these drawings depict only typical embodiments of the invention and are therefore not intended to limit its scope. In the drawings, identical or corresponding parts will be indicated by identical or similar reference numerals for clarity.

[0051] Figure 1 A schematic diagram of the structure of a lightweight YOLO network according to an embodiment of the present invention is shown;

[0052] Figure 2 A schematic diagram of the structure of a DSConv module according to an embodiment of the present invention is shown;

[0053] Figure 3 A schematic diagram of the structure of a GLConv module according to an embodiment of the present invention is shown;

[0054] Figure 4 A schematic diagram of the structure of the C3k2_MDC module according to an embodiment of the present invention is shown;

[0055] Figure 5 A schematic diagram of the structure of the C3k2_AFFM module according to an embodiment of the present invention is shown;

[0056] Figure 6 A schematic diagram of the structure of the shared detection head MSDH according to an embodiment of the present invention is shown;

[0057] Figure 7 A schematic diagram of a simulated damage dataset for a solar panel according to an embodiment of the present invention is shown;

[0058] Figure 8 This diagram illustrates a comparison of mAP50 curves for different networks according to an embodiment of the present invention.

[0059] Figure 9 This diagram illustrates a comparison of the detection performance of different networks according to an embodiment of the present invention; and

[0060] Figure 10 A schematic diagram of a solar panel damage detection system based on lightweight YOLO, according to an embodiment of the present invention, is shown. Detailed Implementation

[0061] In the following description, the invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be practiced without one or more specific details or with other alternatives and / or additional methods or components. In other instances, well-known structures or operations are not shown or described in detail so as not to obscure the inventive points of the invention. Similarly, for illustrative purposes, specific numbers and configurations are set forth to provide a comprehensive understanding of embodiments of the invention. However, the invention is not limited to these specific details.

[0062] In this specification, references to "an embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. The phrase "in one embodiment" appearing throughout this specification does not necessarily refer to the same embodiment in all instances.

[0063] It should be noted that the embodiments of the present invention describe the method steps in a specific order; however, this is only for illustrating the specific embodiment and not for limiting the order of the steps. On the contrary, in different embodiments of the present invention, the order of the steps can be adjusted according to actual needs.

[0064] In this invention, the modules of the system according to the invention can be implemented using software, hardware, firmware, or a combination thereof. When a module is implemented using software, its function can be implemented through computer program flow. For example, the module can be implemented using code segments (such as code segments in languages ​​like C and C++) stored in a storage device (such as a hard disk, memory, etc.), wherein the corresponding function of the module can be implemented when the code segment is executed by a processor. When a module is implemented using hardware, its function can be implemented by setting a corresponding hardware structure. For example, the module's function can be implemented by hardware programming a programmable device such as a field-programmable gate array (FPGA), or by designing an application-specific integrated circuit (ASIC) that includes multiple transistors, resistors, capacitors, and other electronic devices. When a module is implemented using firmware, the module's function can be written into a read-only memory such as an EPROM or EEPROM in the form of program code, and the corresponding function of the module can be implemented when the program code is executed by a processor. In addition, some functions of the module may need to be implemented by separate hardware or by working in cooperation with the hardware. For example, the detection function is implemented by the corresponding sensor (such as a proximity sensor, accelerometer, gyroscope, etc.), the signal transmission function is implemented by the corresponding communication device (such as a Bluetooth device, infrared communication device, baseband communication device, Wi-Fi communication device, etc.), the output function is implemented by the corresponding output device (such as a display, speaker, etc.), and so on.

[0065] To address the problems of slow convergence speed, large number of parameters, and high computational cost in existing damage detection methods, this invention provides a lightweight YOLO network for solar panel damage detection, which can accelerate convergence speed, reduce the number of parameters, and significantly improve detection accuracy and operating efficiency.

[0066] YOLO (You Only Look Once) divides an image into multiple grids, with each grid detecting targets whose center point lies within its region, and outputting a fixed number of bounding boxes and their corresponding categories. YOLOv11 improves upon YOLOv8 by refining the backbone network and neck structure, introducing two key modules: the C3k2 module and the C2PSA module. The C3k2 module effectively reduces redundant computation and improves feature extraction efficiency and expressive power through feature channel partitioning and cross-layer information fusion strategies. The C2PSA module introduces a pyramid slice attention mechanism (Partial Self-Attention, PSA) and combines it with convolutional operations to compress and reconstruct attention features, reducing computational complexity and enhancing the correlation and fusion effect between features at different scales. Simultaneously, depthwise convolution (DWConv) is introduced into the detection head, significantly reducing network parameters while effectively improving feature extraction capabilities. To improve the accuracy and efficiency of solar panel damage detection and achieve network lightweighting, this invention makes a series of targeted improvements based on YOLOv11 n (YOLOv11 n is a lightweight version of YOLOv11). The improved model is called the LDD-YOLO model, or LDD-YOLO (Lightweight Damage Detection YOLO) network, and its network structure is as follows: Figure 1 As shown.

[0067] like Figure 1 As shown, the LDD-YOLO network includes a backbone network 101, a neck network 102, and a head 103.

[0068] The backbone network 101 is configured for feature extraction, including a C3k2_MDC module with embedded multi-scale dilated convolutions, the C3k2_MDC module being configured to expand the receptive field and extract multi-scale features.

[0069] Convolutional Neural Networks (CNNs) have achieved remarkable results in object detection tasks. However, traditional CNNs generate a large number of redundant features, leading to increased consumption of computational resources and memory access. To reduce the computational overhead of the network and improve its deployment efficiency on embedded devices, this invention introduces two lightweight convolutional modules in the feature extraction stage of the backbone network: the depthwise separable convolution (DSConv) module and the grouped local convolution (GLConv) module. The DSConv module includes a depthwise convolution (DWConv) module and a pointwise convolution (PWConv) module. DWConv performs convolution independently on each input channel to extract spatial features; each input channel corresponds to a convolution kernel, and convolution operations are performed only in the two-dimensional space of that channel, without mixing information across channels. This reduces computational cost while preserving spatial feature extraction capabilities. PWConv is used for cross-channel information integration to achieve feature fusion; it uses a 1×11×1 convolutional kernel to linearly combine the output channels of the depthwise convolution, achieving channel reweighting and fusion; through cross-channel operations, it achieves the combination and transformation of features from different channels, enhancing expressive power. The structure of the DSConv module is as follows: Figure 2 As shown. Compared to the standard convolution module 201 (Convolution, Conv), the DSConv module (including the depthwise convolution DWConv module 202 and the pointwise convolution PWConv module 203) can significantly reduce the network parameter size and floating-point operations per second (FLOPs) while maintaining the same output feature map size and number of channels. The FLOPs ratio of the DSConv module to Conv is as follows:

[0070]

[0071] Where H represents the height of the input feature map, W represents the width of the input feature map, KxK represents the size of the convolution kernel, and C... in Indicates the number of input channels, C out Indicates the number of output channels.

[0072] Furthermore, the GLConv module proposed in this invention draws inspiration from the design of GhostConv. The GLConv module first generates a main feature map using partial standard convolutions. Then, it performs a low-computational-cost linear transformation on this main feature map using lightweight 3×3 convolutional kernels to generate a redundant feature map. Finally, it concatenates the main feature map and the redundant feature map to obtain the final output feature map, thereby effectively reducing the dependence on the complete set of convolutional kernels. This method not only reduces computational complexity and the number of parameters but also maintains the network's expressive power, achieving efficient feature representation. The structure of the GLConv module is as follows: Figure 3 As shown. Compared to the GhostConv module, the GLConv module uses smaller convolutional kernels in the feature generation process, further reducing computational cost and improving feature representation ability while maintaining detection performance. The FLOPs of the GLConv module are F. GLConv F GLConv Calculated by the following formula,

[0073]

[0074] Its FLOPs ratio relative to Conv can be expressed as:

[0075]

[0076] Where H represents the height of the input feature map, W represents the width of the input feature map, and C... in Indicates the number of input channels, C out Indicates the number of output channels, s represents the number of cheap transformation operations, and K h ×K w D represents the size of the convolution kernel. h ×D w This indicates the kernel size for linear operations.

[0077] The DSConv and GLConv modules work together to compress parameters as follows: in the basic feature generation stage, the DSConv module decomposition strategy is used to first complete the efficient extraction of main features; in the redundant feature supplementation stage, the GLConv module's 3x3 convolution kernel generation method is used to only perform lightweight expansion on a portion of the main features output by the DSConv module; through this combination, the main feature extraction and redundant feature generation each adopt the optimal lightweight computation path, thereby significantly reducing the overall parameter size and computational cost while maintaining feature expressive power.

[0078] In YOLOv11, the C3k2 module enhances feature extraction capabilities by introducing more bottleneck structures, but this also leads to higher channel information redundancy, thus reducing the network's expressive efficiency. To address this, this invention proposes an improved C3k2-MDC module, which enhances the multi-scale feature extraction capabilities of the backbone network while reducing the number of parameters. This module discards activation functions and normalization layers in convolution operations, retaining only a single two-dimensional convolution (Conv2d), reducing information loss and channel redundancy caused by nonlinear operations. The structure of the C3k2-MDC module is as follows: Figure 4 As shown.

[0079] In one embodiment of the present invention, the C3k2_MDC module includes at least two dilated convolutional kernels with different dilation rates, which may include 3 and 5. For example, the C3k2_MDC module uses dilated convolutions with dilation rates of 3 and 5 to effectively expand the receptive field and enhance the ability to capture contextual information without increasing computational complexity. C3k2-MDC balances network lightweightness with multi-scale feature representation, demonstrating stronger adaptability to detection tasks of targets at different scales in complex backgrounds.

[0080] Let the input feature map be x, the convolution kernel be w, and the kernel size be kxk. Then the output of the C3k2_MDC module using dilated convolution at position i is:

[0081]

[0082] Where, r j d represents the relative position offset of kernel element j on the input feature map; d is the dilation rate, which represents the step size between sampling points.

[0083] The receptive field size of a regular convolution (dilation rate d = 1) is:

[0084] k + (k-1) × (d-1) = k

[0085] When d=1, it is equivalent to no holes being inserted, and the convolution kernel elements are closely arranged.

[0086] With dilated convolution, the receptive field size becomes:

[0087]

[0088] This is because d−1 holes are inserted between every two adjacent elements in the convolution kernel.

[0089] For a convolutional kernel with k=3, when the dilation rate d=3, the receptive field size is 7, meaning the kernel covers a 7×7 area, which is much larger than a typical 3×3 kernel, but the number of parameters and computational cost are the same. For a convolutional kernel with k=3, when the dilation rate d=5, the receptive field size is 11, meaning the kernel covers an 11×11 area, a significant expansion.

[0090] The neck network is configured for feature fusion, including a C3k2_AFFM module with an integrated attention mechanism, which is configured to perform weighted fusion on the multi-scale features.

[0091] Damaged areas on solar panels typically appear as well-defined, relatively independent medium-scale targets, but their diverse textures and edge features make detection challenging. To enhance the network's ability to represent such targets, this invention deploys the C3k2_AFFM module with an integrated attention mechanism in the medium-resolution feature fusion branch, thereby improving the network's detection accuracy for damaged solar panel targets.

[0092] The structure of the C3k2_AFFM module is as follows: Figure 5 As shown, the C3k2_AFFM module employs a multi-branch fusion design. It divides the input features into two parts along the channel direction using a channel chunking operation, resulting in a first-branch feature map and a second-branch feature map. The first-branch feature map is processed using a bottleneck structure (Bottleneck / C3k) to progressively extract deeper features, expanding the effective receptive field and enhancing the network's ability to express complex image structures. The second-branch feature map undergoes average pooling downsampling, lightweight self-attention computation, and upsampling restoration, and is then fused with the processing results of the first-branch feature map. By introducing a self-attention mechanism and drawing inspiration from the Transformer architecture, it enhances the non-linear modeling capability of features through a lightweight attention mechanism based on depthwise separable convolution (ConvLite-Attention, CLATION) and a lightweight feed-forward enhanced network (LFEN). To reduce the memory overhead of the attention mechanism, the module incorporates a downsampling and then restoration design strategy. This module first downsamples the input feature map using 2D average pooling (AvgPool 2d) to compress the spatial dimensionality and reduce computation. Then, attention calculations are performed at low resolution, followed by upsampling to restore the original size, and finally fusion with the residual features. This fusion result serves as the input to LFEN for further extraction of deeper features. Finally, the output of LFEN is fused with its input through residual connections, enhancing feature representation and training stability.

[0093] In the feature fusion stage, pooling is used to spatially reduce the dimensionality of the input high-resolution feature map, significantly reducing its width and height. This effectively reduces redundant information in the solar panel damage features, lowers the computational complexity of subsequent convolutional and feedforward networks, thereby alleviating the computational pressure from high-resolution images and improving the efficiency and quality of feature fusion. The downsampled low-resolution features enter the AFFM module in the fusion stage. This module has a built-in attention mechanism that dynamically focuses on the damage feature region, effectively compensating for the spatial detail loss caused by pooling downsampling. At the same time, the feedforward network enhances the expressive power of the features through nonlinear mapping, strengthening the extraction and discriminative power of semantic information, thereby achieving efficient fusion and enhancement of multi-scale features. The feature map processed by AFFM is upsampled to restore the original spatial resolution, enabling it to be effectively fused with other high-resolution features of the neck network, ensuring the integrity and spatial consistency of multi-scale information. This strategy fully utilizes the advantage of pooling downsampling in significantly reducing computational cost, combined with the built-in attention mechanism of the AFFM module and the efficient feature compensation of the feedforward network, effectively preserving the details and key structural information in the high-resolution solar panel image. While significantly improving the network inference speed and resource utilization efficiency, it ensures the accuracy and integrity of feature fusion.

[0094] Specifically, to introduce an attention mechanism based on channel partitioning and weighted fusion, the input features are first... Features are obtained through 1×1 convolution operations. Then, further segmentation is performed along the channel dimension to obtain the query vector (Query, Q), key vector (Key, K), and value vector (Value, V), that is:

[0095] Q,K,V=Split(Conv 1×1 (X))

[0096] Where d q d k d v These represent the dimensions of the Q, K, and V vectors, respectively; Split indicates that the feature matrix obtained from the convolution operation is divided according to the channel dimension to generate multiple sub-feature matrices; Conv 1×1 (X) represents a 1×1 convolution operation performed on the input feature matrix X to achieve linear mapping and feature transformation of the input channels; the aim is to extract multi-view representations from the input features for attention mechanism design. Next, the attention score matrix is ​​calculated using dot products:

[0097]

[0098] Where Attention(Q,K,V) represents the attention score matrix, d k K represents the dimension of the key vector K.T represents the transpose of the key vector K; softmax represents the normalization operation, which converts the correlation calculation result into weight coefficients to reflect the importance of each feature in the attention calculation.

[0099] This score matrix reflects the degree of similarity between positions in the feature sequence, which helps to capture global contextual dependencies and improve the ability to identify the edge and distinguish the texture of the target.

[0100] To enhance gradient backpropagation capability, the module employs a residual connection method to fuse the attention output with the feature, which is expressed as:

[0101]

[0102] However, when the input feature size is (Where B is the batch size, C is the number of channels, and H and W are the height and width of the feature map, respectively). According to the formula for calculating the attention score matrix, the computational complexity of the self-attention mechanism is... Memory overhead increases quadratically with memory size, becoming a key bottleneck in deployment. To alleviate this problem, this invention introduces a two-dimensional average pooling operation with a kernel size of 2 before the attention mechanism, increasing the feature map spatial resolution from... Shrink to The video memory overhead is caused by Reduced to:

[0103]

[0104] In theory, this design reduces memory usage to a fraction of the original requirement, effectively alleviating the resource burden caused by the attention module. Simultaneously, the smaller feature maps significantly reduce intermediate activations and gradient tensors generated during backpropagation, further reducing memory usage.

[0105] Traditional feedforward neural networks (FFNs) perform the same operation at each pixel location, lacking spatial information interaction capabilities. This invention constructs LFEN, which introduces a spatial interaction mechanism based on FFN, further optimizing the representative features extracted by CLAttention while reducing computational overhead. LFEN first uses 1×1 convolutions and the GELU activation function in the expanded hidden feature space to achieve cross-channel information interaction. Then, the features are divided into two sub-branches. right 3×3 convolutions and GELU activations are applied to enhance its local context modeling capabilities. Finally, the processed... and The concatenation process involves feature fusion via 1×1 convolutions, restoring the channel count to its original size. This process is defined as follows:

[0106]

[0107] in, C(·) represents the feature concatenation operation, Conv2d 3×3 (·) indicates a convolutional layer.

[0108] The detection head is configured to output damage detection results based on fused features. The detection head includes a multi-feature shared detection head (MSDH). The YOLOv11 detection head employs a decoupled structure, performing bounding box regression and class prediction separately using multi-scale feature maps, enhancing its ability to perceive targets of different sizes. Each scale feature map is processed by an independent convolutional module, which helps improve detection accuracy but also introduces parameter redundancy and increased computational cost. Furthermore, even if the relative sizes of targets are similar in feature maps at different scales, the parameters of each convolutional module cannot be shared, reducing parameter utilization efficiency. To address these issues, this invention designs an MSDH detection head. This structure first unifies the number of channels in the multi-scale feature maps, then introduces a shared convolutional module for serial operation, effectively reducing redundant computation and compressing the number of network parameters. In addition, this structure can dynamically adjust the number of channels in the shared convolutional layers to flexibly control computational overhead, improving deployment flexibility while ensuring detection performance.

[0109] The structure of the shared detection head MSDH is as follows: Figure 6 As shown, MSDH consists of a lightweight initial convolutional module (DWConv), a shared convolutional module (Shared_Conv), a shared bounding box regression module (SharedConv_Reg), and a shared classification module (SharedConv_Cls). DWConv adjusts the number of channels in the input feature map, providing a uniform input dimension for the shared convolutional modules. Shared_Conv contains a single Conv and an integrated group normalization (GN) ConvGN. Compared to batch normalization, GN performs normalization independently of the batch size, thus offering greater stability in small-batch or dynamic-batch scenarios. Experiments have shown that it can improve the localization and classification performance of the detector head. After initial channel adjustment, this module processes three feature maps of different scales sequentially, reducing the number of parameters while further enhancing their feature representation capabilities. SharedConv_Reg and SharedConv_Cls are responsible for generating the bounding box regression and class prediction outputs, respectively.

[0110] The MSDH workflow is as follows: A lightweight initial convolutional module is used to unify the number of channels on the fused features; a shared convolutional module is used to obtain the output feature map; a shared bounding box regression module is used to obtain bounding box regression; and a shared classification module is used to obtain the class prediction. The damage detection result includes both the bounding box regression and the class prediction.

[0111] For example, the MSDH workflow is as follows: the input is features from different scales. Figure X The number of channels for Y, Z, and Z was uniformly adjusted to h. Based on features... Figure X For example, first, a 1x1 convolutional layer is used to compress the channels, and then the output features are processed. Figure X 1. Its channel number is adjusted to h, that is:

[0112] X1 = DWConv 1×1 (X, h)

[0113] Next, features Figure X 1. Further processing by Shared_Conv to output features Figure X 2, that is

[0114] X2 = Conv GN 3×3 (Conv 3×3 (X1))

[0115] Finally, features Figure X 2. The bounding box regression B and class prediction C are obtained by passing the data through SharedConv_Reg and SharedConv_Cls respectively, i.e.:

[0116] B = Conv2d 1×1 (X2,64)

[0117] C = Conv2d 1×1 (Conv 1×1 (X2),n)

[0118] Where n is the number of detected categories. The processing of feature maps Y and Z is the same as that of X. The bounding box regression and category prediction results are integrated into the final detection output.

[0119] The CIoU loss function comprehensively considers the overlap between the predicted bounding box and the target bounding box, the distance between their center points, and the difference in aspect ratio during the regression process. Its calculation can be expressed as:

[0120]

[0121] Where d represents the Euclidean distance between the center points of the predicted bounding box and the target bounding box, c is the diagonal length of the minimum bounding box, v represents the aspect ratio penalty, and (w, h) and (wgt h gt ) represent the aspect ratios of the predicted bounding box and the target bounding box, respectively.

[0122] CIoU introduces a penalty term while balancing IoU and aspect ratio, but it still has limitations in scale-sensitive object detection tasks. Its aspect ratio term struggles to distinguish between bounding boxes of different sizes and can easily lead to overexpansion of anchor boxes during regression, affecting the network's convergence speed and regression accuracy. To address this limitation, this invention replaces the original CIoU with PIoUv2. PIoUv2 incorporates a penalty factor adapted to the target size and dynamically adjusts the magnitude and direction of the gradient based on the quality difference between the anchor box and the target box. Simultaneously, a non-monotonic attention mechanism enhances attention to medium-quality anchor boxes, effectively mitigating gradient saturation in high-quality anchor boxes and information redundancy in low-quality anchor boxes, thereby accelerating network convergence and improving detection accuracy.

[0123] The PIoUv2 loss function is used as the loss function for bounding box regression. PIoUv2 for,

[0124]

[0125] q = e -p ,q∈(0,1],

[0126]

[0127] L PIoUv2 =3m(λq)·L PIoU

[0128] Where dw1 and dw2 represent the absolute difference between the predicted bounding box and the target bounding box in the width direction, dh3 and dh4 represent the absolute difference between the predicted bounding box and the target bounding box in the height direction, and w gt and h gt L represents the width and height of the target bounding box, respectively; IoU The Cross-Union Ratio (CUI) loss function represents the ratio of the predicted bounding box to the target bounding box, used to measure the degree of matching between the predicted and target bounding boxes; L PIoU Indicates in L IoU The loss function obtained by introducing a penalty factor P is used to reduce the weight of low-quality anchor boxes in the loss calculation, so as to improve the overall training effect; q represents the quality of the anchor box, P represents the penalty factor, m(x) represents the attention function, and λ represents the hyperparameter that controls the behavior of the attention function.

[0129] When q = 1, zero-error matching between the predicted bounding box and the target bounding box is achieved. As P increases, q gradually decreases, reflecting the reduced importance of low-quality anchor boxes. By introducing a non-monotonic attention mechanism based on anchor box quality, the network's ability to model medium-quality anchor boxes is enhanced, which helps improve the stability and optimization efficiency of bounding box regression.

[0130] The C3k2_MDC module enhances feature extraction capabilities and strengthens response to small targets. It employs multi-scale dilated convolutions (with dilated kernels of varying dilation rates) to adjust the receptive field size, capturing both local details and broader contextual features. This design reduces computational cost while effectively enhancing the network's ability to extract small solar panel damage areas, enabling more accurate identification of small targets in complex backgrounds.

[0131] The C3k2_AFFM module efficiently fuses multi-scale features, enhancing semantic and detailed information. C3k2_AFFM employs pooling downsampling, an attention mechanism, a feedforward network, and upsampling to effectively integrate features at different scales, dynamically focusing on damaged areas, enhancing the fusion of weak feature information related to solar panel damage, reducing redundant computation, and balancing feature integrity with computational efficiency.

[0132] The MSDH module adopts a multi-scale detector head design with shared convolutional kernels. By sharing parameters, it achieves efficient extraction and utilization of multi-scale features. This design significantly reduces the number of network parameters and computational overhead, making the network more lightweight.

[0133] These three modules complement each other. C3k2_MDC enhances feature extraction and effectively reduces the number of parameters through multi-scale dilated convolutions; C3k2_AFFM achieves efficient fusion and detail compensation of multi-scale features; and MSDH adopts a shared convolutional kernel design to further reduce the number of parameters and achieve accurate multi-scale damage detection. The three form a closed-loop optimization synergy, comprehensively improving the overall performance of the LDD-YOLO network in small target recognition, weak feature enhancement, and resource-constrained environments.

[0134] The following example illustrates the application of the lightweight YOLO network proposed in this invention for solar panel damage detection.

[0135] The experiments of this invention were conducted on a computing platform running the Windows 11 operating system, and network training was performed on an NVIDIA GeForce RTX 3090 GPU. The software environment included CUDA 12.4, Python 3.10.16, and PyTorch 2.6.0. No pre-trained weights were used during training, and SGD was used as the optimizer. Specific experimental parameters are shown in Table 1.

[0136] Table 1 Experimental parameters

[0137] Training rounds 300 Batch size 16 Image size 640x640 Network size depth: 0.25 width: 0.25 Initial learning rate 0.01 Learning rate momentum 0.937 Weight decay coefficient 0.0005

[0138] Given the scarcity and difficulty in acquiring real-world images of solar panel damage in orbit, this invention constructs a simulated solar panel damage dataset, collecting 2346 images containing different damage scenarios. These images were acquired using visible light and infrared cameras, covering various damage morphologies caused by simulated impacts and short circuits, specifically as follows: Figure 7 As shown. To enhance the network's adaptability to different input conditions and further improve its generalization performance, this invention employs data augmentation strategies including vertical flipping, horizontal flipping, and image translation. Regarding dataset partitioning, the dataset is divided into training, validation, and test sets in a 7:1:2 ratio, used for network training, parameter tuning, and performance evaluation, respectively, to ensure the effectiveness of network learning and the reliability of evaluation results.

[0139] To objectively evaluate the performance of the proposed LDD-YOLO model (i.e., the LDD-YOLO network), this invention selects several evaluation metrics, including F1 score, average precision (AP), mean average precision (mAP), giga-float operations (GFLOPs), number of parameters, and frames per second (FPS). The F1 score measures the network's balance between precision (P) and recall (R); mAP reflects the overall detection accuracy; GFLOPs and number of parameters evaluate the network's computational complexity and lightweight nature; and FPS measures the network's inference capability. The calculation formulas for these metrics are as follows:

[0140]

[0141] Where TP represents the number of positive samples correctly classified by the network, FP represents the number of negative samples misclassified as positive, FN represents the number of positive samples misclassified as negative, N represents the total number of target categories detected by the network, and AP is the integral of the PR curve.

[0142] To verify the effectiveness and mutual influence of various improvements in the LDD-YOLO network for solar panel damage detection, this invention uses YOLOv11 n as the base network. Under the same experimental conditions, various modules are gradually introduced and combined for ablation experiments to evaluate the contribution of different modules to the network. The ablation experiment results are shown in Table 2.

[0143] Table 2 Ablation Experiment Table

[0144]

[0145] As shown in Table 2, the introduction of GLConv and DSConv modules into the baseline network improved all performance metrics. Specifically, mAP50 increased by 0.2%, and the number of parameters decreased by approximately 5%, indicating that the improved convolutional structure effectively improved computational efficiency while maintaining accuracy. Introducing the improved C3k2_MDC module into the backbone network reduced GFLOPs by 0.3, increased FPS by 5.9%, and reduced the number of parameters by 5%, demonstrating superior computational efficiency. Adding the C3k2_AFFM module to the neck network improved F1 and mAP@0.5 by 0.7% and 1.0%, respectively, effectively enhancing the fusion capability of multi-scale features. Introducing the improved MSDH detector head, with a 13.2% decrease in the number of parameters and a 0.3 decrease in GFLOPs, improved F1, mAP50, and FPS by 0.8%, 0.8%, and 7.7%, respectively, validating the effectiveness of this detector head in network simplification and accuracy optimization. After adopting the PioUv2 loss function, mAP50 improved by 0.8%, demonstrating its effectiveness in optimizing regression accuracy. Integrating the PioUv2 loss function with the improved backbone network reduced network parameters by 15.5%, GFLOPs by 0.8, and mAP50 by 0.7%, indicating a good synergistic effect in feature extraction capability and network lightweighting. Based on this, introducing the detection head MSDH further reduced network parameters and GFLOPs to 29.1% and 1.1, respectively, significantly improving computational efficiency. Finally, integrating all improvements into the LDD-YOLO network resulted in a 26.3% reduction in parameters and a 0.9 reduction in GFLOPs, while improving mAP50, F1, and FPS by 1.1%, 0.4%, and 10.0%, respectively, achieving a comprehensive improvement in both accuracy and efficiency. Experimental results show that the proposed algorithm significantly reduces network complexity while maintaining detection accuracy.

[0146] To further verify the effectiveness of the proposed improved algorithm, a comparative experiment was conducted with the current mainstream lightweight networks YOLOv5n, YOLOv8n, YOLOv10n, and YOLOv11n. The results are shown in Table 3 and... Figure 8 As shown.

[0147] Table 3 Comparison Experiment Table

[0148]

[0149] As shown in Table 3, the LDD-YOLO network achieves a good balance between detection accuracy, network complexity, and inference speed. In terms of detection accuracy, the mAP50 reaches 0.971, outperforming all comparison networks and improving upon YOLOv5n, YOLOv8n, YOLOv10n, and YOLOv11n by 0.7%, 0.5%, 0.6%, and 1.1%, respectively. Regarding network complexity, the LDD-YOLO network has 1.90M parameters and a computational GFLOPs of 5.4, significantly lower than YOLOv5n, YOLOv8n, YOLOv10n, and YOLOv11n, with parameter reductions of 24%, 6.9%, 29.4%, and 26.3%, respectively. GFLOPs are reduced by 1.7, 2.7, 2.8, and 0.9, respectively. With a network size of only 3.94MB and an inference speed of 484 FPS, it demonstrates good deployment capabilities. Furthermore, from... Figure 8 As can be seen, the LDD-YOLO network converges quickly, exhibits significant performance improvement in the early stages of training, and shows a stable curve with minimal fluctuations in the later stages of training, demonstrating good convergence and stability. Experimental results show that the LDD-YOLO network has strong advantages in detection accuracy, computational efficiency, and training performance, making it more suitable for deployment on resource-constrained spaceborne platforms.

[0150] To more intuitively analyze the performance of the LDD-YOLO network in solar panel damage detection, test data covers typical scenarios including cracks and defects caused by space debris impacts, as well as thermal anomalies caused by short circuits. Detection was performed under the same conditions as YOLOv5n, YOLOv8n, YOLOv10n, and YOLOv11n, with a focus on analyzing the performance of each network in terms of target localization accuracy, category recognition accuracy, and false positives.

[0151] Figure 9 This diagram illustrates a comparison of the detection performance of different networks according to an embodiment of the present invention. Figure 9 The detection results include those for the YOLOv5n network (901), YOLOv8n network (902), YOLOv10n network (903), YOLOv11n network (904), and LDD-YOLO network (905). Figure 9As shown, the YOLOv8n network can effectively locate cracks, defects, and thermal anomalies, but it suffers from duplicate detection, has weak detection capability for small targets, and its accuracy is somewhat reduced. The YOLOv10n network improves the accuracy of small target detection, but it misses minor cracks and also exhibits duplicate detection. While the YOLOv5n and YOLOv11n networks do not exhibit duplicate detection, the problem of missing minor cracks remains significant, and the YOLOv11n network also suffers from over-detection. In contrast, the LDD-YOLO network can more accurately detect small targets and minor cracks, and it exhibits no duplicate detection, demonstrating higher stability and robustness, making it particularly suitable for damage detection tasks on solar panels.

[0152] To address the issues of missed detections, false detections, and excessive network parameters in current solar panel damage detection methods, this invention proposes a lightweight YOLO network for solar panel damage detection, primarily implemented using an LDD-YOLO network. This network introduces a C3k2_MDC module into the backbone to expand the receptive field and compress the parameter size. Simultaneously, it replaces standard convolutions with DSConv and GLConv modules, effectively reducing computational complexity while enhancing feature extraction capabilities. A multi-branch feature fusion module C3k2_AFFM with an integrated attention mechanism is introduced into the neck structure, combined with average pooling and upsampling modules to improve the efficiency and accuracy of damage target information extraction. A shared detection head, MSDH, is designed in the detection head to further improve parameter utilization efficiency and reduce network size. Finally, a dynamically adjustable PIOUv2 damage function guides the network to focus more on medium-quality candidate boxes, accelerating convergence.

[0153] In the solar panel damage detection task, the improved LDD-YOLO model, i.e. the LDD-YOLO network, compared with the YOLOv11 n network, improves mAP by 1.1% to 97.1%, reduces the number of parameters by 26.3%, reduces GFLOPs by 0.9, and improves FPS by 10%. While balancing detection accuracy and network lightweighting, it demonstrates superior performance in detecting small targets and minor cracks.

[0154] In one embodiment of the present invention, a method for detecting damage to solar panels is also provided. In another embodiment, the method for detecting damage to solar panels can be executed by a computer. Figure 10 As shown, the method includes the following steps:

[0155] Step 1001: Collect images of the solar panels to be detected and perform preprocessing;

[0156] Step 1002: Input the preprocessed image into the lightweight YOLO network for solar panel damage detection to perform processing; and

[0157] Step 1003: Output the damage detection results.

[0158] In one embodiment of the invention, the invention also provides an electronic device comprising the aforementioned lightweight YOLO network for solar panel damage detection.

[0159] In one embodiment of the present invention, another electronic device is also provided, comprising: a processor, a graphics card with an artificial intelligence chip, and a memory. The memory is configured to store machine-readable instructions, the graphics card is configured to train the lightweight YOLO network for solar panel damage detection, and the processor is configured to execute the machine-readable instructions. When the processor and / or the graphics card executes the machine-readable instructions, it performs the following processing steps: collecting images of the solar panel to be detected and performing preprocessing; inputting the preprocessed image into the lightweight YOLO network for solar panel damage detection as described above to perform processing; and outputting the damage detection result.

[0160] The graphics card used can preferably be a model with a GPU computing power higher than 5.0. Since the amount of data to be trained is large, providing a graphics card configuration can significantly improve the training speed.

[0161] The memory includes various media capable of storing machine-readable instructions, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0162] It is understood that, in addition to the memory and processor mentioned above, the above-mentioned electronic devices also include other hardware and software components not listed in this specification. The specific components can be determined according to the model of the specific data processing device in different application scenarios, and will not be listed and described in detail in this specification.

[0163] Although various embodiments of the present invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined according to the technical solutions of the invention and their equivalents.

Claims

1. A lightweight YOLO network for solar panel damage detection, characterized in that, include: The backbone network is configured for feature extraction and includes a C3k2_MDC module with embedded multi-scale dilated convolutions. The C3k2_MDC module is configured to include at least two dilated convolution kernels with different dilation rates, including 3 and 5, to expand the receptive field and extract multi-scale features. The neck network, configured for feature fusion, includes a C3k2_AFFM module with an integrated attention mechanism, which is configured to perform weighted fusion on the multi-scale features; as well as The detection head is configured to output damage detection results based on fused features, and the detection head includes a multi-feature shared detection head (MSDH). The C3k2_AFFM module is used to perform feature fusion, including: Channel segmentation is performed on the input feature map to obtain the first branch feature map and the second branch feature map; The bottleneck structure is used to process the first branch feature map; The second branch feature map is subjected to average pooling downsampling, lightweight self-attention computation, and upsampling restoration operations, and then fused with the processing result of the first branch feature map; and The fused features are then fed into a lightweight feedforward enhancement network to perform further feature extraction.

2. The lightweight YOLO network according to claim 1, characterized in that, The backbone network also includes a depthwise separable convolution DSConv module and a grouped local convolution GLConv module.

3. The lightweight YOLO network according to claim 2, characterized in that: The DSConv module includes a depthwise convolutional (DWConv) module and a pointwise convolutional (PWConv) module; and / or The GLConv module generates the main feature map through standard convolution, performs linear transformation using a 3×3 convolution kernel to generate a redundant feature map, and finally concatenates the main feature map and the redundant feature map to obtain the final output feature map.

4. The lightweight YOLO network according to claim 1, characterized in that, The lightweight self-attention computation includes: The input features are mapped to query vectors, key vectors, and value vectors. in, Represents the query vector. Represents the key vector. Represents a value vector; Split indicates that the feature matrix obtained from the convolution operation is divided according to the channel dimension to generate multiple sub-feature matrices; This indicates that a convolution operation with a kernel size of 1×1 is performed on the input feature matrix X; Calculate the attention score matrix. in, Represents the attention score matrix. Represents the key vector Dimension size, Represents the key vector The transpose of ; softmax represents the normalization operation, converting the correlation calculation results into weight coefficients; and The attention output and input features are fused through residual connections.

5. The lightweight YOLO network according to claim 1, characterized in that, The damage detection results based on the fused features include: By using a lightweight initial convolution module, the number of channels is unified for the fused features; The output feature map is obtained by sharing a convolutional module; The bounding box regression is obtained using the shared bounding box regression module; and Use the shared classification module to obtain category predictions; The damage detection results include the bounding box regression and the category prediction.

6. The lightweight YOLO network according to claim 1, characterized in that, The PIoUv2 loss function is used as the loss function for bounding box regression. for, in, and This represents the absolute difference between the predicted bounding box and the target bounding box in the width direction. and This represents the absolute difference between the predicted bounding box and the target bounding box in the height direction. and These represent the width and height of the target bounding box, respectively. The intersection-union ratio (IUU) loss function represents the ratio between the predicted bounding box and the target bounding box. Indicates in The loss function obtained by introducing a penalty factor P on the basis of the above is This indicates the quality of the anchor point frame. Indicates the penalty factor. Represents the attention function. The hyperparameters that control the behavior of the attention function.

7. A method for detecting damage to solar panels, characterized in that, Includes the following steps: Collect images of the solar panels to be inspected and perform preprocessing; The preprocessed image is input into the lightweight YOLO network according to any one of claims 1 to 6 to perform processing; as well as Output the damage detection results.

8. An electronic device, characterized in that, Includes the lightweight YOLO network according to any one of claims 1 to 6.

Images