Insulator abnormal heating point detection method, device and equipment based on neural network

Abstract

The application provides a kind of insulator abnormal heating point detection method, device and equipment based on neural network, it is related to image processing technical field.The method comprises: obtaining the infrared image of transmission line insulator;The infrared image is input into the trained defect detection model, and the abnormal heating point defect of insulator in the infrared image is identified;Wherein, the defect detection model is based on YOLOV8 framework and is built, including main network, neck network and detection head, and the main network is the EfficientNetV2 backbone network with deformable attention mechanism.The application can realize the balance of lightweight and high performance of abnormal heating point detection, and significantly improve the accuracy and reliability of detection, and provide strong technical support for the intelligentization, preventive operation and maintenance of power line.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, and device for detecting abnormal heating points in insulators based on neural networks. Background Technology

[0002] Transmission line insulators are critical equipment in power systems. Exposed to complex and changing outdoor environments for extended periods, they are susceptible to increased local resistance due to contamination, aging, cracks, and other defects, leading to abnormal heating. This abnormal heating is a significant precursor to serious faults such as insulator flashover, line tripping, and even line breakage. Therefore, regular and efficient heating point detection of insulators is crucial for ensuring the safe and stable operation of the power grid.

[0003] Traditional inspection methods primarily rely on manual handheld infrared thermal imagers for ground or tower inspections. This approach is not only inefficient and costly, but also heavily influenced by human subjective experience, making it difficult to implement in complex terrain and adverse weather conditions, and hindering routine monitoring of large-scale power lines. With the development of drones and infrared imaging technology, automated inspection using drones has become a trend. However, transforming the massive amounts of infrared images collected into effective defect diagnosis remains a significant challenge. Existing deep learning-based automated inspection methods often encounter the following bottlenecks in practical deployments: First, the background of insulator infrared images is complex, containing interference from conductors, tower materials, vegetation, and even clouds, easily obscuring minute heat-generating features and leading to missed detections. Second, high-precision deep learning models typically have high computational complexity and a large number of parameters, making real-time inference difficult on edge devices with limited load capacity and computing resources, such as drones, thus limiting their field application. Furthermore, the size of insulator heat-generating points varies considerably, making single-scale feature extraction insufficient to effectively capture defects of different sizes simultaneously, affecting the robustness of the detection. Summary of the Invention

[0004] This invention provides a method, apparatus, and device for detecting abnormal heating points in insulators based on neural networks, in order to solve the problem that current methods for detecting abnormal heating in insulators are difficult to balance high detection accuracy, low computational complexity, and strong environmental adaptability.

[0005] In a first aspect, embodiments of the present invention provide a method for detecting abnormal heating points in insulators based on neural networks, comprising: Acquire infrared images of insulators of power transmission lines; Infrared images are input into a trained defect detection model to identify abnormal heating points in insulators within the infrared images. The defect detection model is built on the YOLOV8 framework and includes a backbone network, a neck network, and a detection head. The backbone network is an EfficientNetV2 backbone network with a deformable attention mechanism.

[0006] In one possible implementation, an infrared image is input into a trained defect detection model to identify abnormal heating points in the insulator within the infrared image, including: The infrared image is input into the backbone network to obtain the multi-scale feature map of the infrared image; The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image; The fused feature map is input into the detection head to obtain the abnormal heating point defect of the insulator in the infrared image.

[0007] In one possible implementation, the backbone network includes three Fused-MBConv modules, a first MBConv module, a second MBConv module, a third MBConv module, a DAttention module, and an SPPF module connected in sequence. Infrared images are input into the backbone network to obtain multi-scale feature maps of the infrared images, including: The infrared image is subjected to multi-scale feature extraction using three Fused-MBConv modules to obtain the initial feature map of the infrared image; The first MBConv module extracts spatial features and dynamically weights the initial feature map to obtain and output the first feature map. The second feature map is obtained and output by performing spatial feature extraction and dynamic weighting on the first feature map through the second MBConv module. The second feature map is obtained and output by performing spatial feature extraction and dynamic weighting, deformable attention weighting, and multi-scale fusion on the third MBConv module, DAttention module, and SPPF module.

[0008] In one possible implementation, the input features are subjected to deformable attention weighting via the DAttention module, including: The query tags are obtained by linearly projecting the input features; Input the query tag into the offset subnetwork to obtain the position offset; The positional offset is superimposed on the input features and bilinear interpolation is performed to obtain the relative positional deviation between the sampled features and the deformed points. By performing a linear projection on the sampled features, deformed keys and values ​​are obtained; Based on the relative positional deviation between the query marker and the deformation point, attention weighting is applied to the deformable keys and values ​​to obtain input features that undergo deformable attention weighting.

[0009] In one possible implementation, the C2f module in the neck network is replaced with a CSP module, which is a CSP module improved by partial convolution. The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image, including: The third feature map is upsampled and fused with the second processed feature map to obtain the first intermediate feature map; After the first intermediate feature map is convolved by the first CSPCC module, it is upsampled and fused with the first feature map to obtain the second intermediate feature map. After the second fused feature map is convolved by the second CSPCC module, the first fused feature map is obtained. After the first fused feature map is convolved by the first CBS module, it is fused with the first intermediate feature map to obtain the third intermediate feature map. The third intermediate feature map is convolved by the third CSPCC module to obtain the second fused feature map; After the second fused feature map is convolved by the second CBS module, it is fused with the third feature map to obtain the fourth intermediate feature map. The fourth intermediate feature map is convolved by the fourth CSPCC module to obtain the third fused feature map; Output the first fused feature map, the second fused feature map, and the third fused feature map.

[0010] In one possible implementation, the CSPCC module includes a ConvModule layer, a SPlit layer, multiple PConv layers, a Concat layer, and a Conv layer connected in sequence.

[0011] In one possible implementation, before inputting the infrared image into the trained defect detection model to identify the abnormal heating point defect in the insulator in the infrared image, the following steps are also included: Multiple infrared images of insulators with abnormal heating points were acquired, and the locations of the abnormal heating points were marked to form the original dataset; The original dataset was divided into a training set and a test set, and the infrared images of insulators in the training set were augmented using methods such as Mosaic, max pooling, photometric distortion, and geometric distortion. The initial defect detection model is trained using a loss function and an augmented training set to obtain a trained defect detection model.

[0012] Secondly, embodiments of the present invention provide a neural network-based insulator abnormal heating point detection device, comprising: The acquisition module is used to acquire infrared images of insulators of transmission lines; The detection module is used to input the infrared image into a trained defect detection model to identify abnormal heating points in the insulator in the infrared image. The defect detection model is built on the YOLOV8 framework and includes a backbone network, a neck network, and a detection head. The backbone network is an EfficientNetV2 backbone network with a deformable attention mechanism.

[0013] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0014] The present invention provides a neural network-based method, apparatus, and device for detecting abnormal heating points in insulators. By employing a design combining a computationally efficient backbone network with an attention mechanism, it achieves a balance between lightweight design and high performance at the model level. This allows complex deep learning models to be deployed on edge computing devices such as drones, enabling real-time, online intelligent identification of insulator defects and significantly improving the automation level and efficiency of inspections. Furthermore, by introducing a dynamically focusing attention mechanism, the model can adaptively suppress complex background noise, concentrating computational resources on analyzing the characteristics of the insulator itself and its key components. This allows for accurate localization of minute abnormal heating points even in chaotic field environments, significantly improving the accuracy and reliability of detection. For specific insulator scenarios, a complete solution from data to deployment has been developed, providing strong technical support for the intelligent and preventative maintenance of power lines. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the implementation of the neural network-based insulator abnormal heating point detection method provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the defect detection model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the Fused-MBConv module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the MBConv module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the SE module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of DAttention provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the CSPPC module provided in an embodiment of the present invention; Figure 8This is a schematic diagram of the structure of the insulator abnormal heating point detection device based on neural network provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0016] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] See Figure 1 The flowchart illustrating the implementation of the neural network-based insulator abnormal heating point detection method provided in this embodiment of the invention is described in detail below: Step 101: Obtain infrared images of the insulators of the transmission line.

[0018] In this embodiment, infrared images of transmission line insulators can be acquired using equipment such as an infrared thermal imager mounted on a drone. These images reflect the temperature distribution on the surface of the insulator; abnormal heating areas typically appear as localized high-temperature points, which are important indicators of the presence of defects.

[0019] To eliminate temperature characterization differences caused by non-defect factors and ensure the model focuses on real heating anomalies, a standardized preprocessing procedure can be added to the infrared images input into the model. First, emissivity correction is performed: the emissivity of different components of insulators (especially composite insulators), such as skirts, sheaths, and fittings, varies (typically between 0.85 and 0.95). By using preset component material information or reading it from image metadata, the corresponding emissivity value (ε) is retrieved, converting the radiance value received by the detector into a true surface temperature value, eliminating temperature measurement errors caused by material differences. Next, ambient temperature compensation is performed: fluctuations in ambient temperature (T_env) affect the contrast between the target and the background. T_env is estimated from environmental sensor data accompanying the image or by analyzing known constant-temperature reference objects in the image (such as distant mountains or sky background), and compensation calculations are performed on the entire image to ensure a consistent temperature baseline for images taken at different times and locations. Finally, distance normalization is performed: based on the intrinsic parameters of the infrared thermal imager (focal length, pixel size) and the recorded shooting distance (D), the actual spatial size represented by each pixel in the image is normalized. For example, the same heat source may occupy only a few pixels in an image when it is far away, but tens of pixels when it is close. By normalizing the distance, the area of ​​the heat source can be converted into an approximate actual physical size (such as square centimeters), laying the foundation for subsequent area-based defect determination.

[0020] This series of preprocessing steps essentially transforms the original infrared "image" into a "thermal feature map" that reflects the true thermal state of the target and is physically comparable, greatly enhancing the model's generalization ability and the physical significance of the interpretation under different shooting conditions.

[0021] Step 102: Input the infrared image into the trained defect detection model to identify abnormal heating point defects in the insulator in the infrared image; wherein, the defect detection model is built based on the YOLOV8 framework and includes a backbone network, a neck network and a detection head, the backbone network being an EfficientNetV2 backbone network with a deformable attention mechanism.

[0022] In this embodiment, after acquiring the image, it is input into a pre-trained defect detection model for intelligent analysis. The core of this model lies in its architecture design; it is built upon the current advanced target detection framework YOLOv8 and incorporates key improvements tailored to the specific needs of insulator defect detection. For example... Figure 2 As shown, the model follows the component division of YOLOv8, including a backbone network for feature extraction, a neck network for multi-scale feature fusion, and a detection head for final classification and localization. The most significant improvement lies in the backbone network, which uses EfficientNetV2 as its core to significantly reduce model complexity and computational cost while maintaining powerful feature extraction capabilities. Crucially, this backbone network innovatively integrates a Deformable Attention (DAttention) mechanism. This design allows the network to not only extract global features but also dynamically focus computational resources on regions in the image more likely to contain key structures such as insulator skirts. This enables the network to accurately capture minute abnormal heat points even under complex background interference (such as trees, sky, and transmission towers), achieving a balance between high accuracy and high efficiency, and meeting the application requirements of real-time inspection of UAV edge devices.

[0023] This invention achieves a balance between lightweight design and high performance at the model level by combining a computationally efficient backbone network with an attention mechanism. This allows complex deep learning models to be deployed on edge computing devices such as drones, enabling real-time, online intelligent identification of insulator defects and significantly improving the automation level and efficiency of inspections. Furthermore, by introducing a dynamically focusing attention mechanism, the model can adaptively suppress complex background noise, concentrating computational resources on analyzing the characteristics of the insulator itself and its key components. This allows for precise location of minute abnormal heat points even in chaotic field environments, significantly improving the accuracy and reliability of detection. For specific insulator scenarios, a complete solution from data to deployment has been developed, providing strong technical support for the intelligent and preventative maintenance of power lines.

[0024] In one possible implementation, an infrared image is input into a trained defect detection model to identify abnormal heating points in the insulator within the infrared image, including: The infrared image is input into the backbone network to obtain the multi-scale feature map of the infrared image; The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image; The fused feature map is input into the detection head to obtain the abnormal heating point defect of the insulator in the infrared image.

[0025] In this embodiment, when an infrared image is input into the model, it is first processed by the backbone network. The backbone network extracts feature information with different levels of abstraction and spatial resolution from the input image through a series of hierarchical convolution and attention operations. This information is organized into multi-scale feature maps. For example, shallow feature maps have higher resolution and contain rich details such as edges and textures, which is beneficial for locating small targets or fine structures; deep feature maps have lower resolution but stronger semantic information, which helps in understanding the category and overall shape of the target. These multi-scale feature maps are then fed into the neck network. The core task of the neck network is to perform feature pyramid fusion. Through operations such as upsampling, downsampling, and concatenation, it effectively integrates the feature maps of different scales output by the backbone network, ensuring that each scale of feature simultaneously contains rich semantic information and accurate spatial information, thereby generating enhanced fused feature maps. Finally, these fused feature maps are fed into the detection head. The detection head typically consists of a series of convolutional layers, responsible for predicting target probability, regressing bounding box coordinates, and classifying defect categories for each preset anchor point (or grid) on the fused feature map. The final output includes the location (bounding box) of the insulator's infrared heating point defect and the category confidence score. This process ensures end-to-end, multi-scale feature optimization processing from the original image to the final detection result.

[0026] In one possible implementation, the backbone network includes three Fused-MBConv modules, a first MBConv module, a second MBConv module, a third MBConv module, a DAttention module, and an SPPF module connected in sequence. Infrared images are input into the backbone network to obtain multi-scale feature maps of the infrared images, including: The infrared image is subjected to multi-scale feature extraction using three Fused-MBConv modules to obtain the initial feature map of the infrared image; The first MBConv module extracts spatial features and dynamically weights the initial feature map to obtain and output the first feature map. The second feature map is obtained and output by performing spatial feature extraction and dynamic weighting on the first feature map through the second MBConv module. The second feature map is obtained and output by performing spatial feature extraction and dynamic weighting, deformable attention weighting, and multi-scale fusion on the third MBConv module, DAttention module, and SPPF module.

[0027] In this embodiment, to ensure the scalability of the algorithm for edge segment applications, the EfficientNetV2 high-efficiency convolutional network is introduced, significantly reducing the algorithm's complexity and computational load while maintaining network performance. Furthermore, to improve the accuracy of the algorithm in identifying composite insulators, the DAttention deformable attention mechanism is chosen, allowing the network to focus on key feature points. Since the DAttention deformable attention mechanism only focuses on certain regions, its computational cost is relatively low, maintaining the network's lightweight nature.

[0028] Depend on Figure 2 As can be seen, the backbone network begins with three sequentially connected Fused-MBConv modules. The Fused-MBConv module merges the traditional 3x3 depthwise convolution from a depthwise separable convolution with the previous 1x1 pointwise convolution into a standard 3x3 convolution. This "fusion" design reduces memory access overhead and improves computational efficiency in the early stages of the model (when the input resolution is high). The infrared image passes through these three modules sequentially, completing initial downsampling and basic feature extraction, outputting an initial feature map. Subsequently, the network connects to the first MBConv module. The MBConv module is based on depthwise separable convolution and the SE (Squeeze-and-Excitation) attention mechanism. While extracting spatial features, it adaptively calibrates the feature responses along the channel dimensions through the SE module, weighting important feature channels and suppressing secondary channels, thus obtaining a dynamically weighted first feature map. This map is output to the neck network for subsequent fusion. Next, the second MBConv module further refines and weights the channels of the first feature map, outputting a second feature map with lower spatial resolution, which is also output to the neck network. Next, the second feature map flows through the third MBConv module for deep feature extraction, and then enters the core DAttention module. The DAttention module allows the network to dynamically adjust the spatial position of its attention focus on the feature map based on the input content, rather than being limited to regular grid points, thus adapting more flexibly to the deformation of the insulator target and complex backgrounds. Finally, the features pass through a SPPF (Spatial Pyramid Pooling Fast) module, which performs parallel pooling using multiple max-pooling kernels of different sizes and concatenates the results. This efficiently aggregates multi-scale contextual information on a fixed-size feature map, outputting a third feature map rich in multi-scale semantics. These three outputs (the first, second, and third feature maps) constitute the multi-scale feature pyramid provided by the backbone network.

[0029] The MBConv module integrates depthwise separable convolution technology, and its structure is as follows: Figure 3As shown, the number of input channels is first adjusted using a 1x1 convolution, then a depthwise separable convolution is applied to extract spatial features, and finally, a 1x1 convolution is used again to restore the number of channels. Furthermore, MBConv integrates a Squeeze-and-Excitation (SE) module, which dynamically weights the feature maps output from all channels. The SE module learns the importance of each channel, explicitly constructs the interdependencies between feature channels, and weights the features accordingly to enhance key features and suppress less important ones.

[0030] like Figure 4 As shown, similar to MBConv, Fused-MBConv also includes SE modules and 1x1 convolutional layers, but its core difference lies in replacing depthwise separable convolutions with standard 3x3 convolutions. This design aims to reduce computational costs and improve the overall performance of the network. The structure of the SE module is as follows: Figure 5 As shown.

[0031] In one possible implementation, the input features are subjected to deformable attention weighting via the DAttention module, including: The query tags are obtained by linearly projecting the input features; Input the query tag into the offset subnetwork to obtain the position offset; The positional offset is superimposed on the input features and bilinear interpolation is performed to obtain the relative positional deviation between the sampled features and the deformed points. By performing a linear projection on the sampled features, deformed keys and values ​​are obtained; Based on the relative positional deviation between the query marker and the deformation point, attention weighting is applied to the deformable keys and values ​​to obtain input features that undergo deformable attention weighting.

[0032] In this embodiment, the background of the insulator images captured on-site is often quite complex, which affects the algorithm's recognition of the insulator skirt features. Therefore, to effectively improve the algorithm's efficiency and performance, an attention mechanism is used to change the network's focus, making it concentrate on the insulator skirt. Simultaneously, while maintaining good performance, the computational load on the algorithm is minimized; therefore, this paper chooses to use the DAttention deformable attention mechanism.

[0033] DAttention is a deep learning model (Transformer) for vision based on a self-attention mechanism, employing an innovative deformable attention mechanism. Unlike the self-attention mechanism used in traditional Transformers, which requires processing all pixels of the entire image and increases computational burden, DAttention focuses on key regions in the image, effectively reducing computational cost while maintaining model performance. Furthermore, DAttention's deformable attention mechanism dynamically selects sampling points instead of applying fixed processing to the entire image. This dynamic selection strategy allows the model to more precisely focus on regions crucial to the current task. Figure 6 This is a diagram of the DAttention network structure.

[0034] The core idea behind the internal working mechanism of the DAttention module is to generate irregular spatial sampling points that are related to the input content. For example... Figure 6 As shown in the figure, x is the input feature map, q is the query label to which the feature map is linearly projected, used to obtain the offset of each reference point, and θoffset is a lightweight offset subnetwork used to generate the offset. p, v˜, and k˜ are the keys and values ​​projected from the sampled features based on the deformation points. Relative positional biases are also calculated using the deformation points, enhancing the multi-head attention of the output transformed features. Specifically, for a feature (denoted as X) input to the DAttention module, it is first mapped to query tokens through a linear projection layer. These tokens encode the information requirements at various locations in the feature map. Subsequently, these tokens are fed into a lightweight offset sub-network (typically composed of a few fully connected or convolutional layers). This sub-network learns and outputs a set of positional offsets (Δp), the number of which is related to the spatial location of the query token and the preset number of sampling points. These offsets indicate the deformation sampling positions relative to the original regular grid points. Using the obtained offsets Δp, bilinear interpolation sampling is performed on the original input feature X to obtain a set of sampled features corresponding to the deformation positions. Simultaneously, based on these deformation points, the relative positional bias information between them can be calculated. This information serves as a bias term in the attention calculation to preserve important positional priors. Next, the sampled features are linearly projected again to generate "deformed keys" and "deformed values." Finally, attention-weighted calculations are performed: the deformed keys (K) are scored with the query label (Q), and the calculated relative positional deviation is added. The scores are then normalized using the Softmax function to obtain attention weights. These weights are used to perform a weighted summation of the deformed values ​​(V), ultimately outputting features weighted by deformable attention. This process allows the network to adaptively focus computational resources on feature regions related to the insulator skirt structure, improving the detection accuracy of large insulator skirt targets and significantly enhancing the ability to perceive key local features.

[0035] In one possible implementation, the C2f module in the neck network is replaced with a CSP module, which is a CSP module improved by partial convolution. The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image, including: The third feature map is upsampled and fused with the second processed feature map to obtain the first intermediate feature map; After the first intermediate feature map is convolved by the first CSPCC module, it is upsampled and fused with the first feature map to obtain the second intermediate feature map. After the second fused feature map is convolved by the second CSPCC module, the first fused feature map is obtained. After the first fused feature map is convolved by the first CBS module, it is fused with the first intermediate feature map to obtain the third intermediate feature map. The third intermediate feature map is convolved by the third CSPCC module to obtain the second fused feature map; After the second fused feature map is convolved by the second CBS module, it is fused with the third feature map to obtain the fourth intermediate feature map. The fourth intermediate feature map is convolved by the fourth CSPCC module to obtain the third fused feature map; Output the first fused feature map, the second fused feature map, and the third fused feature map.

[0036] In this embodiment, in order to effectively reduce unnecessary computation in the feature fusion part and process redundant information in the feature map, a lightweight convolutional structure CSPPC is used to improve the C2f module. This structure reduces the complexity of network computation and maintains the performance level of the model.

[0037] Specifically, the neck network is responsible for performing bidirectional (top-down and bottom-up) feature pyramid fusion. Its data processing flow is as follows: First, the deepest third feature map output from the backbone network is upsampled (e.g., using nearest neighbor interpolation or transposed convolution) to match its spatial dimensions with the second feature map. Then, the two are concatenated or added together to obtain the first intermediate feature map. Next, the first intermediate feature map undergoes lightweight convolution processing through a first CSPCC module to fuse features and reduce redundancy. After processing, it is upsampled again and fused with the first feature map output from the backbone network to obtain the second intermediate feature map. The second intermediate feature map is processed by a second CSPCC module, outputting the first fused feature map (P1), which has the highest resolution and is suitable for detecting small-sized heat spots. Subsequently, to pass the fine location information from the shallow layers to the deeper layers, the first fused feature map (P1) is first downsampled through a CBS module (usually composed of Conv+BatchNorm+SiLU), and then fused with the previous first intermediate feature map to obtain the third intermediate feature map. After further processing by the third CSPCC module, the output is the second fused feature map (P2), used to detect medium-sized targets. Similarly, the second fused feature map (P2) is downsampled by another CBS module and fused with the third feature map directly output by the backbone network to obtain the fourth intermediate feature map. Finally, after processing by the fourth CSPCC module, the third fused feature map (P3) is output, used to detect large-sized targets. Ultimately, the neck network outputs these three fused feature maps of different scales, P1, P2, and P3, to the detection head, achieving robust detection of multi-scale insulator defects.

[0038] In one possible implementation, the CSPCC module includes a ConvModule layer, a SPlit layer, multiple PConv layers, a Concat layer, and a Conv layer connected in sequence.

[0039] In this embodiment, considering the application of the algorithm of this invention at the edge of a drone, there are certain requirements for network speed and lightweight design. Therefore, it is necessary to make a more lightweight improvement to the original YOLOv8 structure. In this embodiment, a lightweight convolutional structure composed of a combination of CSP structure and partial convolution (PartialConv) structure is used to replace the original C2f module of YOLOv8, reducing the computational complexity.

[0040] The PartialConv architecture is a lightweight design aimed at reducing the computational cost of convolutional neural networks. It effectively reduces unnecessary computation by utilizing redundant information in feature maps. Because many similar or repetitive data points often exist across multiple channels of the network, these data are repeatedly processed during propagation without actually adding any new useful features, instead leading to an unnecessary increase in computational resources and memory access. Unlike traditional convolution, PartialConv performs regular convolution for spatial feature extraction only on a small subset of the input channels, leaving the other channels unchanged. This computational strategy reduces floating-point operations (FLOPS), thus simplifying the computation process. By optimizing this process, the PartialConv architecture helps improve the efficiency and performance of the network.

[0041] The CSPCC module is a product of combining the CSP (Cross Stage Partial) structure with the idea of ​​Partial Convolution (PConv). For example... Figure 7 The CSPCC module consists of the following layers in sequence: The first layer is a standard ConvModule layer (containing convolution, batch normalization, and activation functions) used for initial transformation of the input features. Next is the Split layer, which roughly divides the input feature map into two parts along the channel dimension. One part flows through multiple cascaded PConv layers for depth processing. The core idea of ​​the PConv layer is: instead of performing spatial convolution on all input channels, only a subset of channels (e.g., the first quarter or half) are selected for regular convolution operations to extract spatial features, while the remaining channels remain unchanged and pass directly through. This significantly reduces floating-point operations and memory access. Multiple cascaded PConv layers can progressively extract features while maintaining lightweight architecture. The feature portion processed by the PConv stream, along with the unprocessed feature portion from the Split layer, is reassembled along the channel dimension in the Concat layer. Finally, a Conv layer (usually a ConvModule) fuses the concatenated features and adjusts the number of channels, outputting the final result.

[0042] The CSPPC architecture combines the advantages of both CSP and PartialConv architectures, employing PartialConv for convolution operations in each sub-layer of the CSP architecture. This architecture promotes gradient flow and enriches feature combinations through the split-process-merge approach of CSP, while significantly reducing the computational complexity of modules through PConv, making the model more suitable for deployment on edge devices with limited computing power.

[0043] In one possible implementation, before inputting the infrared image into the trained defect detection model to identify the abnormal heating point defect in the insulator in the infrared image, the following steps are also included: Multiple infrared images of insulators with abnormal heating points were acquired, and the locations of the abnormal heating points were marked to form the original dataset; The original dataset was divided into a training set and a test set, and the infrared images of insulators in the training set were augmented using methods such as Mosaic, max pooling, photometric distortion, and geometric distortion. The initial defect detection model is trained using a loss function and an augmented training set to obtain a trained defect detection model.

[0044] In this embodiment, the model's performance depends on high-quality training data. First, a large number of infrared images of insulators containing both normal and abnormal heating points need to be collected from actual inspection scenarios (such as drone aerial photography) or laboratory environments to form an initial image library. Then, annotation tools (such as LabelImg) are used to accurately label the abnormal heating point regions in these images, i.e., drawing bounding boxes and labeling their categories (such as "umbrella skirt heating," "iron cap heating," etc.), forming a structured initial dataset. To enhance the model's generalization ability and robustness, the initial dataset needs to be randomly divided into training and testing sets at a certain ratio (e.g., 8:2). Effective data augmentation strategies are implemented on the training set, such as: Mosaic data augmentation stitches four training images into one, simulating complex scenes with multiple targets and scales; max pooling can simulate downsampling, increasing the model's robustness to scale changes; photometric distortion simulates different weather and lighting conditions by adjusting the image's brightness, contrast, saturation, and hue; geometric distortion, such as random rotation, translation, scaling, and cropping, can simulate changes in the shooting angle. These enhancements significantly expand the diversity and scale of the training data without introducing new data. The expanded training set is used to train the defect detection model initialized according to the structure described in claims 1-6. During training, the error between the model's predictions and the ground truth labels is calculated using a loss function (typically including classification loss, bounding box regression loss, and object confidence loss), and all model weight parameters are iteratively updated using a backpropagation algorithm and an optimizer (such as SGD or Adam). The training process continues for multiple rounds (e.g., 400 rounds), and the model performance (e.g., mean accuracy mAP) is periodically evaluated on independent test sets. When the loss function value stabilizes and the evaluation metric reaches a satisfactory level, the model is considered to have converged, and training is complete. Finally, the resulting model weight file is saved and can be used for subsequent offline or online insulator defect detection tasks.

[0045] To address the issue of complex background interference, this solution innovatively introduces a negative sample attention guidance strategy targeting typical interference objects. Specifically, during training data labeling, not only are the insulator heating points (positive samples) labeled, but also frequently occurring interference objects in the image, such as perched birds, strong sunlight reflections on metal components, and vegetation leaves partially obscuring the insulator, are labeled as "interference objects" and given special category labels. During the forward propagation of model training, when the feature map passes through the DAttention module, in addition to calculating the regular attention loss designed to make the model focus on the insulator region, a new contrastive attention loss term is added. The design goal of this term is: for regions labeled as "interference objects," the attention weights obtained after DAttention calculation for their corresponding query tokens should be minimized. The loss function can be formally expressed as encouraging the attention weight distribution of interference object regions to tend towards uniformity or below a certain threshold, specifically:

[0046] in, To compare attention loss, the losses for each component are calculated as follows: (1) Distractor-induced attentional inhibition loss :

[0047] For the negative sample (interference) region mask, at the scale corresponding to the feature map, the additionally labeled typical interference (such as birds, reflective, shading vegetation) regions are mapped to 1. This represents the total number of pixels with a value of 1 in the mask. Let i be the attention weight vector at a specific position i. This is a constant. This loss term encourages the DAttention module to assign lower attention weights to the interference region. When the average attention weight of the interference region approaches 0, the loss approaches 0; conversely, the loss increases.

[0048] (2) Loss of enhanced target attention :

[0049] This is the mask for the positive sample (target) region. This loss term encourages the DAttention module to assign higher attention weights to the insulator target region. The loss approaches 0 when the average attention weight of the target region approaches 1.

[0050] (3) Loss of consistency of attention distribution :

[0051] in, This represents the attention weight distribution in the background region; Uniform represents a uniform distribution. It is the attention weight distribution of the target region. It is a Gaussian distribution with the mean at the center of the target.

[0052] This term, through KL divergence constraints, encourages the attention weight distribution of the background region to tend to be uniform, avoiding over-attention to specific background regions, while encouraging the attention weight distribution of the target region to tend to be a Gaussian distribution with the target center as the focus, which is consistent with the shape prior of the insulator target.

[0053] By integrating the basic detection loss and the newly introduced contrastive attention loss, the improved total loss function is obtained:

[0054] in, Based on the detection loss, λ1=0.05, λ2=0.5, λ3=1.0, To compare attention loss, α=1.0, β=0.5, γ=0.3. The attention loss weight coefficient is set to 0.1 at the beginning of training and gradually increased to 0.5 as training progresses.

[0055] Through this contrastive learning-based guidance, the DAttention mechanism is trained not only to actively seek out key features of insulators but also to actively ignore misleading background interference features. This is equivalent to encoding the prior experience of inspection experts into the model's attention mechanism, making it more robust and focused in complex field environments.

[0056] Specifically, the complete process of training a defect detection model includes the following steps: (1) Collect infrared images of insulators with abnormal heating points by drone aerial photography or laboratory photography.

[0057] (2) Manual or semi-automatic annotation is performed using data annotation software to form the original dataset.

[0058] (3) The dataset is randomly divided into training and test sets in an 8:2 ratio, and the training data is augmented by using Mosaic, max pooling, photometric distortion and geometric distortion methods.

[0059] (4) The prepared training set is passed through the feature extraction part, feature fusion part and detection head of the built target detection network in sequence. During the feature extraction process, the feature maps of the formed image are fused at different scales. Finally, the loss function is obtained from the anchor boxes formed by clustering and the labeled detection boxes, and the error is backpropagated to correct the network parameters.

[0060] (5) Input the test set into the network to test the accuracy of the network model in order to determine whether the convergence condition has been met.

[0061] (6) After completing steps (4) and (5), the network weights obtained from the first training are fed back into the deep network for repeated training and verification for 400 rounds. Training is stopped after the loss function and average accuracy converge.

[0062] (7) Extract the weight file obtained from the training and deploy it at the edge. The model can then automatically identify abnormal heating point defects of insulators in the UAV inspection video.

[0063] In practical applications, the model's output not only includes the defect location bounding box but also allows for preliminary defect severity analysis, directly supporting operational and maintenance decisions. This function is implemented in the post-processing stage of the detection head. The system first extracts the highest temperature value (T_max) within the predicted heating area and the reference temperature value (T_ref) of the adjacent normal skirt or similar environmental background of the insulator from the preprocessed standardized thermal feature map, calculating the relative temperature rise ΔT = T_max - T_ref. Simultaneously, combining distance normalization information, it calculates the approximate actual area (S) of the heating area. Subsequently, the system incorporates a configurable criterion rule base, with thresholds set according to power industry standards. The judgment logic includes: when ΔT < 10K and S < 2cm², it is classified as a "general defect" and further attention is recommended; when 10K ≤ ΔT < 20K or 2cm² ≤ S < 5cm², it is classified as a "major defect" and planned maintenance is recommended as soon as possible; when ΔT ≥ 20K or S ≥ 5cm², or when the temperature rise and area do not reach the threshold but exhibit a specific pattern such as continuous heating of multiple umbrella skirts, it is classified as an "emergency defect," triggering a real-time alarm and recommending emergency handling. Finally, the model output includes the defect location, category, confidence level, calculated ΔT and S, and the severity level.

[0064] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0065] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0066] Figure 8 A schematic diagram of the insulator abnormal heating point detection device based on neural network provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 8 As shown, the neural network-based insulator abnormal heating point detection device 8 includes: Acquisition module 81 is used to acquire infrared images of insulators of transmission lines; The detection module 82 is used to input the infrared image into the trained defect detection model to identify the infrared heating point defects of the insulator in the infrared image; wherein, the defect detection model is built based on the YOLOV8 framework and includes a backbone network, a neck network and a detection head, wherein the backbone network is an EfficientNetV2 backbone network with a deformable attention mechanism.

[0067] In one possible implementation, the detection module 82 is specifically used for: The infrared image is input into the backbone network to obtain the multi-scale feature map of the infrared image; The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image; The fused feature map is input into the detection head to obtain the abnormal heating point defect of the insulator in the infrared image.

[0068] In one possible implementation, the backbone network includes three Fused-MBConv modules, a first MBConv module, a second MBConv module, a third MBConv module, a DAttention module, and an SPPF module connected in sequence. Detection module 82 is specifically used for: The infrared image is subjected to multi-scale feature extraction using three Fused-MBConv modules to obtain the initial feature map of the infrared image; The first MBConv module extracts spatial features and dynamically weights the initial feature map to obtain and output the first feature map. The second feature map is obtained and output by performing spatial feature extraction and dynamic weighting on the first feature map through the second MBConv module. The second feature map is obtained and output by performing spatial feature extraction and dynamic weighting, deformable attention weighting, and multi-scale fusion on the third MBConv module, DAttention module, and SPPF module.

[0069] In one possible implementation, the detection module 82 is specifically used for: The query tags are obtained by linearly projecting the input features; Input the query tag into the offset subnetwork to obtain the position offset; The positional offset is superimposed on the input features and bilinear interpolation is performed to obtain the relative positional deviation between the sampled features and the deformed points. By performing a linear projection on the sampled features, deformed keys and values ​​are obtained; Based on the relative positional deviation between the query marker and the deformation point, attention weighting is applied to the deformable keys and values ​​to obtain input features that undergo deformable attention weighting.

[0070] In one possible implementation, the C2f module in the neck network is replaced with a CSP module, which is a CSP module improved by partial convolution. Detection module 82 is specifically used for: The third feature map is upsampled and fused with the second processed feature map to obtain the first intermediate feature map; After the first intermediate feature map is convolved by the first CSPCC module, it is upsampled and fused with the first feature map to obtain the second intermediate feature map. After the second fused feature map is convolved by the second CSPCC module, the first fused feature map is obtained. After the first fused feature map is convolved by the first CBS module, it is fused with the first intermediate feature map to obtain the third intermediate feature map. The third intermediate feature map is convolved by the third CSPCC module to obtain the second fused feature map; After the second fused feature map is convolved by the second CBS module, it is fused with the third feature map to obtain the fourth intermediate feature map. The fourth intermediate feature map is convolved by the fourth CSPCC module to obtain the third fused feature map; Output the first fused feature map, the second fused feature map, and the third fused feature map.

[0071] In one possible implementation, the CSPCC module includes a ConvModule layer, a SPlit layer, multiple PConv layers, a Concat layer, and a Conv layer connected in sequence.

[0072] In one possible implementation, the detection module 82 is further used for: Before inputting the infrared images into the trained defect detection model to identify the abnormal heating point defects in the insulators in the infrared images, multiple infrared images of insulators with abnormal heating points are acquired and the locations of the abnormal heating points are marked to form the original dataset. The original dataset was divided into a training set and a test set, and the infrared images of insulators in the training set were augmented using methods such as Mosaic, max pooling, photometric distortion, and geometric distortion. The initial defect detection model is trained using a loss function and an augmented training set to obtain a trained defect detection model.

[0073] This invention achieves a balance between lightweight design and high performance at the model level by combining a computationally efficient backbone network with an attention mechanism. This allows complex deep learning models to be deployed on edge computing devices such as drones, enabling real-time, online intelligent identification of insulator defects and significantly improving the automation level and efficiency of inspections. Furthermore, by introducing a dynamically focusing attention mechanism, the model can adaptively suppress complex background noise, concentrating computational resources on analyzing the characteristics of the insulator itself and its key components. This allows for precise location of minute abnormal heat points even in chaotic field environments, significantly improving the accuracy and reliability of detection. For specific insulator scenarios, a complete solution from data to deployment has been developed, providing strong technical support for the intelligent and preventative maintenance of power lines.

[0074] Figure 9 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 9 As shown, the electronic device 9 of this embodiment includes a processor 90 and a memory 91. The memory 91 stores a computer program 92. When the processor 90 executes the computer program 92, it implements the steps in the various method embodiments described above. Alternatively, when the processor 90 executes the computer program 92, it implements the functions of each module / unit in the various device embodiments described above.

[0075] For example, computer program 92 may be divided into one or more modules / units, which are stored in memory 91 and executed by processor 90 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 92 in electronic device 9.

[0076] Electronic device 9 may include, but is not limited to, processor 90 and memory 91. Those skilled in the art will understand that... Figure 9 This is merely an example of electronic device 9 and does not constitute a limitation on electronic device 9. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 9 may also include input / output devices, network access devices, buses, etc.

[0077] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0078] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0079] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting abnormal heating points in insulators based on neural networks, characterized in that, include: Acquire infrared images of insulators of power transmission lines; The infrared image is input into a trained defect detection model to identify abnormal heating points in the insulator in the infrared image. The defect detection model is built on the YOLOV8 framework and includes a backbone network, a neck network, and a detection head. The backbone network is an EfficientNetV2 backbone network with a deformable attention mechanism.

2. The method for detecting abnormal heating points in insulators based on neural networks according to claim 1, characterized in that, The step of inputting the infrared image into a trained defect detection model to identify abnormal heating points in the insulator within the infrared image includes: The infrared image is input into the backbone network to obtain a multi-scale feature map of the infrared image; The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image; The fused feature map is input into the detection head to obtain the abnormal heating point defect of the insulator in the infrared image.

3. The method for detecting abnormal heating points in insulators based on neural networks according to claim 2, characterized in that, The backbone network includes three Fused-MBConv modules, a first MBConv module, a second MBConv module, a third MBConv module, a DAttention module, and an SPPF module connected in sequence. The step of inputting the infrared image into the backbone network to obtain the multi-scale feature map of the infrared image includes: The infrared image is subjected to multi-scale feature extraction using the three Fused-MBConv modules to obtain the initial feature map of the infrared image; The initial feature map is spatially feature-extracted and dynamically weighted by the first MBConv module to obtain and output the first feature map. The second feature map is obtained by performing spatial feature extraction and dynamic weighting on the first feature map through the second MBConv module, and then output. The second feature map is subjected to spatial feature extraction and dynamic weighting, deformable attention weighting, and multi-scale fusion through the third MBConv module, the DAttention module, and the SPPF module to obtain the third feature map and output it.

4. The method for detecting abnormal heating points in insulators based on neural networks according to claim 3, characterized in that, The input features are subjected to deformable attention weighting through the DAttention module, including: The query tags are obtained by linearly projecting the input features; Input the query tag into the offset sub-network to obtain the position offset; The positional offset is superimposed on the input feature and bilinear interpolation is performed to obtain the relative positional deviation between the sampled feature and the deformation point. Linear projection is performed on the sampled features to obtain deformed keys and values; Based on the relative positional deviation between the query marker and the deformation point, attention weighting is applied to the keys and values ​​of the deformation to obtain input features that undergo deformable attention weighting.

5. The method for detecting abnormal heating points in insulators based on neural networks according to claim 3, characterized in that, The C2f module in the neck network is replaced with a CSPC module, which is a CSP module improved by partial convolution. The step of inputting the multi-scale feature map into the neck network to obtain the fused feature map of the infrared image includes: The third feature map is upsampled and fused with the second processed feature map to obtain the first intermediate feature map; After the first intermediate feature map is convolved by the first CSPCC module, it is upsampled and fused with the first feature map to obtain the second intermediate feature map. After the second fused feature map is convolved by the second CSPCC module, the first fused feature map is obtained. After the first fused feature map is convolved by the first CBS module, it is fused with the first intermediate feature map to obtain the third intermediate feature map. The third intermediate feature map is convolved by the third CSPCC module to obtain the second fused feature map. After the second fused feature map is convolved by the second CBS module, it is fused with the third feature map to obtain the fourth intermediate feature map. The fourth intermediate feature map is convolved by the fourth CSPCC module to obtain the third fused feature map. Output the first fused feature map, the second fused feature map, and the third fused feature map.

6. The method for detecting abnormal heating points in insulators based on neural networks according to claim 5, characterized in that, The CSPCC module includes a ConvModule layer, a SPlit layer, multiple PConv layers, a Concat layer, and a Conv layer connected in sequence.

7. The method for detecting abnormal heating points in insulators based on neural networks according to any one of claims 1 to 6, characterized in that, Before inputting the infrared image into the trained defect detection model to identify the abnormal heating point defect in the insulator in the infrared image, the method further includes: Multiple infrared images of insulators with abnormal heating points were acquired, and the locations of the abnormal heating points were marked to form the original dataset; The original dataset is divided into a training set and a test set, and the infrared images of insulators in the training set are augmented using methods such as Mosaic, max pooling, photometric distortion, and geometric distortion. The initial defect detection model is trained using a loss function and an augmented training set to obtain a trained defect detection model.

8. A neural network-based insulator abnormal heating point detection device, characterized in that, include: The acquisition module is used to acquire infrared images of insulators of transmission lines; The detection module is used to input the infrared image into a trained defect detection model to identify abnormal heating points of insulators in the infrared image; wherein, the defect detection model is built based on the YOLOV8 framework and includes a backbone network, a neck network and a detection head, wherein the backbone network is an EfficientNetV2 backbone network with a deformable attention mechanism.

9. The insulator abnormal heating point detection device based on neural network according to claim 8, characterized in that, The detection module is specifically used for: The infrared image is input into the backbone network to obtain a multi-scale feature map of the infrared image; The multi-scale feature map is input into the neck network to obtain the fused feature map of the infrared image; The fused feature map is input into the detection head to obtain the abnormal heating point defect of the insulator in the infrared image.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

Images