Insulator spontaneous explosion defect detection method based on lightweight adaptive YOLO model
By adopting the lightweight adaptive YOLO model in insulator defect detection, the problems of insufficient detection accuracy and limited computing resources are solved, and efficient and real-time insulator self-detonation defect detection is achieved.
Patent Information
- Application Number
- CN202411625574.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-14
- Publication Date
- 2025-02-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing insulator defect detection technologies have problems with insufficient accuracy when dealing with small object detection, especially in complex backgrounds and high-altitude images captured by drones, which are difficult to deploy in real-time on devices with limited computing resources.
The insulator self-destruction defect detection method based on the lightweight adaptive YOLO model is adopted. By designing a lightweight network architecture, achieving efficient feature fusion, optimizing gradient shunt and building a decoupled head network structure, the model complexity is reduced and detection accuracy is improved.
It significantly improves the accuracy and efficiency of insulator self-destruction defect detection, and can realize real-time detection on devices with limited computing resources, adapting to the requirements of real-time and resource limitations in actual applications.
Smart Images
Figure CN119445084A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electric power operation and maintenance, and in particular to a method for detecting insulator self-explosion defects based on a lightweight adaptive YOLO model. Background Art
[0002] Existing insulator defect detection technologies have insufficient accuracy when dealing with small target detection, especially in complex backgrounds and high-altitude images taken by drones. The detection accuracy of small targets such as insulator self-explosion defects is low. Although traditional target detection methods perform well in general scenarios, they are prone to missed detections and false detections when facing small targets with extremely small pixel ratios. In addition, existing detection models have complex structures, large number of parameters, and high computational overhead, making them difficult to deploy in real time on devices with limited computing resources (such as drones).
[0003] In order to solve these problems, the present invention aims to provide a lightweight and adaptive feature fusion detection method, which can not only improve the detection accuracy of small target insulator defects, but also reduce the computational complexity of the model, and adapt to the requirements of real-time and resource constraints in practical applications.
[0004] The above information is presented as background information only to assist with understanding the present disclosure and no determination or admission is made as to whether any of the above may be used as prior art with respect to the present disclosure. Summary of the invention
[0005] The present invention provides an insulator self-explosion defect detection method based on a lightweight adaptive YOLO model, which can not only improve the detection accuracy of small target insulator defects, but also reduce the calculation complexity of the model, and meet the requirements of real-time performance and resource limitation in practical applications.
[0006] To achieve the above object, the present invention provides an insulator self-explosion defect detection method based on a lightweight adaptive YOLO model, comprising:
[0007] Step S10: Design a lightweight network architecture: construct an efficient basic network unit, including customized convolution blocks and feature extraction modules, to achieve network lightweight and multi-scale feature capture;
[0008] Step S20: Implement feature fusion: Develop a feature enhancement and fusion mechanism, including designing a dedicated module to strengthen the feature information flow, using a self-learning method to achieve multi-scale feature fusion, and ensuring that the original feature information is preserved during the fusion process;
[0009] Step S30: Optimizing gradient splitting: introducing advanced convolution modules to reduce feature redundancy, and optimizing feature transfer by applying partial convolution algorithms at key layers to maintain the integrity of spatial information;
[0010] Step S40: constructing a head network: designing and implementing a decoupled head network structure, enhancing spatial perception capability and optimizing parameter efficiency through a convolution algorithm;
[0011] Step S50: training the model: using customized loss functions for model training, including binary cross entropy loss for classification tasks, and distributed focal loss and complete intersection-over-union loss for regression tasks;
[0012] Step S60: Experimental verification and optimization: The model is tested in a small target scene shot by a real drone and compared with existing methods to evaluate the performance of the model. The model is further optimized based on experimental data to improve detection accuracy and efficiency.
[0013] Step S70: Model application: using the optimized model to detect insulator self-explosion defects.
[0014] Optionally, the S10 includes:
[0015] S101: Define the basic convolution module, including k×k convolution layer, batch normalization and activation function;
[0016] S102: Construct a spatial pyramid pooling layer to extract feature representations at different scales.
[0017] Optionally, the S20 includes:
[0018] S201: Design C2f module and use bottleneck layer to enhance feature information flow;
[0019] S202: Using the Bi-AFPN-P2 network structure, the feature fusion of insulators and ISDs at different scales is achieved through self-learning;
[0020] S203: Introduce a horizontal flow feature path to ensure that the original feature information is not lost during the feature fusion process.
[0021] Optionally, the S202 includes:
[0022] Step S2021: extracting features from the input image to generate a multi-scale feature map;
[0023] Step S2022: input the generated feature map into the AFM, and adaptively adjust the weight and fusion strategy of the feature map;
[0024] Step S2023: bidirectionally fuse the adjusted feature map through the context fusion path of Bi-AFPN;
[0025] Step S2024: Output the fused feature map for subsequent defect detection and positioning.
[0026] Optionally, the S30 includes:
[0027] S301: Introduce the Faster-C2F module and use partial convolution to reduce feature redundancy;
[0028] S302: Implement PConv in the bottleneck layer of the C2f module to retain shallow spatial information.
[0029] Optionally, the S40 includes:
[0030] S401: Design DD-Head to achieve dual decoupling of tasks and structures;
[0031] S402: Introduce SAConv in the regression branch to enhance spatial perception capabilities;
[0032] S403: Apply PConv in the classification branch to reduce parameter and FLOPs cost.
[0033] Optionally, the S50 includes:
[0034] S501: Training for classification tasks using binary cross entropy loss;
[0035] S502: Apply distributed focal loss and complete intersection-over-union loss to train the regression task.
[0036] Optionally, the S60 includes:
[0037] S601: Test the model performance in the ISD small target scene photographed by the drone;
[0038] S602: Compare with the specified defect detection method to verify the detection accuracy and real-time performance of the model;
[0039] S603: Optimize the model according to the experimental results to ensure detection accuracy and efficiency.
[0040] Optionally, the S70 includes:
[0041] The trained model is deployed to the inspection system to perform real-time inspection on insulator images taken by drones and output defect information so that necessary maintenance can be performed based on the defect information.
[0042] Optionally, the defect information includes defect location and confidence level.
[0043] Compared with the prior art, the present invention provides an insulator self-explosion defect detection method based on a lightweight adaptive YOLO model, which has the following beneficial effects:
[0044] ① Network lightweighting and multi-scale feature capture: By designing a lightweight network architecture, including customized convolution blocks and feature extraction modules, the complexity and computational complexity of the model are effectively reduced, while capturing multi-scale features, which is beneficial for the detection of small targets such as insulator self-explosion defects (ISDs).
[0045] ② Efficient feature fusion mechanism: The C2f module and Bi-AFPN-P2 method are used to strengthen the feature information flow, and multi-scale feature fusion is achieved through self-learning method, ensuring that the original feature information is retained during the fusion process, thereby improving the detection capability of small targets.
[0046] ③ Optimize gradient diversion: The Faster-C2F module and partial convolution algorithm are introduced to reduce feature redundancy, optimize feature transfer, maintain the integrity of spatial information, and further improve the model's detection performance for small targets.
[0047] ④ Decoupled head network structure: The design of DD-Head achieves dual decoupling of tasks and structures, which not only enhances spatial perception capabilities, but also optimizes parameter efficiency, making the model more lightweight while ensuring accuracy.
[0048] ⑤ Customized loss function: By combining binary cross entropy loss, distributed focal loss and complete intersection-over-union loss, it provides a more effective training strategy for classification and regression tasks, which helps to improve the detection accuracy of the model.
[0049] The present invention has other features and advantages, which will be apparent from or will be described in detail in the accompanying drawings and the following detailed description incorporated herein, which together serve to explain the specific principles of the invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0051] Figure 1 It is an overall structural diagram of a lightweight network architecture provided by an embodiment of the present invention;
[0052] Figure 2 It is a visualization diagram of a backbone network feature graph provided by an embodiment of the present invention;
[0053] Figure 3is a schematic diagram of the fusion pathway of Bi-AFPN-P2 provided in an embodiment of the present invention;
[0054] Figure 4 Schematic diagrams of different convolution operations provided by embodiments of the present invention;
[0055] Figure 5 is a schematic diagram of the composition of the Faster-C2f module provided in an embodiment of the present invention;
[0056] Figure 6 is a structural diagram of a DD-Head provided by an embodiment of the present invention;
[0057] Figure 7 It is a comparison chart of test results of different defect detection methods provided by embodiments of the present invention in an ISD small target scene photographed by a drone.
[0058] Figure 8 It is a flow chart of a method for detecting insulator self-explosion defects based on a lightweight adaptive YOLO model provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to explain in detail the possible application scenarios, technical principles, specific schemes that can be implemented, and the purposes and effects that can be achieved, the following is a detailed description of the specific embodiments listed in conjunction with the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of the present application, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present application.
[0060] Reference to "embodiment" herein means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present application. The term "embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or association with other embodiments. In principle, in the present application, as long as there is no technical contradiction or conflict, the various technical features mentioned in the embodiments can be combined in any way to form a corresponding implementable technical solution.
[0061] Unless otherwise defined, the technical terms used in this document have the same meanings as those generally understood by those skilled in the art to which this application belongs; the use of relevant terms in this document is only for describing specific embodiments and is not intended to limit this application.
[0062] In the description of this application, the term "and / or" is an expression used to describe the logical relationship between objects, indicating that three relationships may exist, for example, A and / or B, which means: A exists, B exists, and A and B exist at the same time. In addition, the character " / " in this article generally indicates that the objects before and after are in an "or" logical relationship.
[0063] In the present application, terms such as “first” and “second” are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship of quantity, priority or sequence between these entities or operations.
[0064] Without further limitations, in this application, the words "include", "comprises", "has" or other similar expressions used in the sentences are intended to cover non-exclusive inclusion. These expressions do not exclude the presence of additional elements in the process, method or product including the elements, so that the process, method or product including a series of elements may include not only those limited elements, but also other elements not explicitly listed, or also include elements inherent to such process, method or product.
[0065] Similar to the understanding in the Examination Guidelines, in this application, expressions such as "greater than", "less than", "exceed" and the like are understood to exclude the number itself; expressions such as "above", "below", "within" and the like are understood to include the number itself. In addition, in the description of the embodiments of this application, "multiple" means more than two (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups", "multiple times", etc., unless otherwise clearly and specifically limited.
[0066] In the description of the embodiments of the present application, space-related expressions used, such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicate the orientation or position relationship based on the orientation or position relationship shown in the specific embodiments or drawings, and are only for the convenience of describing the specific embodiments of the present application or facilitating the reader's understanding, and do not indicate or imply that the referred device or component must have a specific position, a specific orientation, or be constructed or operated in a specific orientation, and therefore cannot be understood as a limitation on the embodiments of the present application.
[0067] Unless otherwise expressly specified or limited, in the description of the embodiments of the present application, the terms such as "install", "connect", "connect", "fix", "set", etc. used should be understood in a broad sense. For example, the "connection" can be a fixed connection, a detachable connection, or an integrated setting; it can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. For technicians in the technical field to which the present application belongs, the specific meanings of the above terms in the embodiments of the present application can be understood according to the specific circumstances.
[0068] Example 1
[0069] Please refer to the attached Figure 8 , is a flow chart of a method for detecting self-explosion defects of insulators based on a lightweight adaptive YOLO model provided by an embodiment of the present invention. The method is applicable to the scene of detecting self-explosion defects of insulators. The method is executed by a detection system, and the detection system can be implemented by software and / or hardware. The method specifically includes the following steps:
[0070] Step S10: Design a lightweight network architecture: construct an efficient basic network unit, including customized convolution blocks and feature extraction modules, to achieve network lightweight and multi-scale feature capture;
[0071] S101: Define the basic convolutional module (CBS), including k×k convolutional layers, batch normalization (BatchNorm) and activation function (SiLU);
[0072] S102: Construct a spatial pyramid pooling layer (SPPF) to extract feature representations at different scales;
[0073] The overall structure of the ISD small target detection system based on LA-YOLO is as follows Figure 1 As shown in (a). The basic convolution module (CBS) consists of a k×k convolution layer, batch normalization (BatchNorm) and an activation function (SiLU), as shown in Figure 1 The C2f module uses the bottleneck layer to enhance the feature information flow of gradient feedback, as shown in Figure 1 (c). The spatial pyramid pooling layer (SPPF) extracts feature representations of the same feature map at different scales to achieve the fusion of local and global features, as shown in Figure 1 (b). Figure 1 As shown in (d). On the training loss function, binary cross entropy (BCE) loss is applied in the classification task to calculate the difference between the predicted classification probability distribution and the true label. In the regression task, distributed focal loss (DFL) and complete intersection over union (CIoU) are used together to model the object bounding box as a general distribution to optimize the location of the object boundary. Faster-C2f, Bi-AFPN-P2 and DD-Head are introduced below.
[0074] Step S20: Implement feature fusion: Develop a feature enhancement and fusion mechanism, including designing a dedicated module to strengthen the feature information flow, using a self-learning method to achieve multi-scale feature fusion, and ensuring that the original feature information is preserved during the fusion process;
[0075] S201: Design C2f module and use bottleneck layer to enhance feature information flow;
[0076] In the acquired images, most insulators are medium targets, while insulator self-explosion defects (ISDs) are small targets, such as Figure 2 In the feature extraction process of the backbone network, the shallow feature map C3 contains richer local information about insulators and ISDs and has a higher feature resolution, as shown in (a). Figure 2 However, as the network deepens, as the number of downsampling or convolution increases, the global information of the insulators in the intermediate feature map C4 to the deep feature map C5 gradually becomes richer, but the information of the small target ISD is submerged, as shown in Figure 2 as shown in (c)-(d).
[0077] S202: Using the Bi-AFPN-P2 network structure, the feature fusion of insulators and ISDs at different scales is achieved through self-learning;
[0078] Furthermore, this patent proposes a Bi-AFPN-P2 method based on an adaptive fusion module (AFM). This method achieves the fusion of insulators and ISDs at different scales through self-learning, effectively solving the information conflict problem caused by feature differences. At the same time, the shallow feature map C2 is introduced into the fusion stage to enhance the feature information transmission of small target ISD.
[0079] The Bi-AFPN-P2 method based on adaptive fusion module (AFM) proposed in this paper aims to improve the accuracy and efficiency of insulator self-explosion defects (ISDs) detection. The method mainly involves the following structural modules:
[0080] (1) Adaptive Fusion Module (AFM):
[0081] AFM is a core component used to dynamically adjust and optimize the fusion process of feature maps of different scales.
[0082] This module adaptively selects the optimal fusion strategy by analyzing the scale, texture and context information of the input feature map.
[0083] AFM includes multiple sub-modules, such as the attention mechanism module and the feature map fusion layer, to enhance the expressiveness of small object features.
[0084] (2) Bidirectional Adaptive Feature Pyramid Network (Bi-AFPN):
[0085] Bi-AFPN is an improved feature pyramid network that can simultaneously perform top-down and bottom-up feature fusion.
[0086] The network structure improves the detection capability of small targets by introducing AFM, especially in the detection of small-sized targets such as insulator self-explosion defects.
[0087] At the P2 level, Bi-AFPN fuses high-level semantic information and low-level detail information through AFM to generate a feature map with rich semantics.
[0088] Specifically, the Bi-AFPN-P2 method based on the adaptive fusion module (AFM) includes the following steps and structures:
[0089] Specific implementation steps:
[0090] Step S2021: extract features from the input image and generate a multi-scale feature map.
[0091] Step S2022: Input the generated feature map into AFM and adaptively adjust the weight and fusion strategy of the feature map.
[0092] Step S2023: Bidirectionally fuse the adjusted feature map through the context fusion path of Bi-AFPN.
[0093] Step S2024: Output the fused feature map for subsequent defect detection and positioning.
[0094] S203: Introducing a horizontal flow feature path to ensure that the original feature information is not lost during the feature fusion process;
[0095] like Figure 3 As shown, in order to ensure that the original feature information is not lost during the feature fusion process, this embodiment adds a horizontal flow feature path from the input feature maps C3 and C4 to the terminal fusion feature maps P3 and P4, thereby improving the reusability of the features. At the same time, the top-down vertical feature path passes the semantic information of the deep feature map C5 to the shallow fusion feature map P2 through the intermediate layer fusion feature maps M4 and M3, enriching the classification information of the insulator and ISD in P2. The bottom-up vertical feature path passes the spatial information of the shallow feature map C2 to the deep fusion feature map P5, adding the positioning information of the insulator and the insulator self-explosion small-size defects in P5. These three together constitute a bidirectional feature pyramid feature fusion path, which makes full use of the feature information extracted from the backbone network and improves the fusion efficiency.
[0096] Step S30: Optimizing gradient splitting: introducing advanced convolution modules to reduce feature redundancy, and optimizing feature transfer by applying partial convolution algorithms at key layers to maintain the integrity of spatial information;
[0097] S301: Introduce the Faster-C2F module and use partial convolution (PConv) to reduce feature redundancy;
[0098] This embodiment uses a lightweight gradient splitting module (Faster-C2f). This module introduces the concept of partial convolution (PConv), which performs conventional convolution operations on only some feature channels to reduce the feature redundancy of insulators and ISDs. At the same time, shallow spatial information continues to be passed down from another part of the feature channels to improve the spatial information of ISD in the deep feature map.
[0099] S302: Implement PConv in the bottleneck layer of the C2f module to retain shallow spatial information;
[0100] In the bottleneck layer of the C2f module, the convolution (Conv) operation is used to perform convolution operations on all channels of the input features. Although this design greatly enriches the gradient information during training, it also performs redundant convolution operations on similar feature maps, such as Figure 4 (a) shows that partial convolution (PConv) only performs convolution operations on part of the input channel, such as Figure 4 (b) In the remaining channels, the shallow feature information is mapped into the deep network through PConv to preserve the spatial characteristics of insulators and insulator defects.
[0101] In terms of structural design, the Faster-C2f module uses Faster-Block as the gradient flow branch, which enables the model to learn more diverse and rich insulator and ISD feature information in continuous gradient splitting, such as Figure 5 As shown in (a). The IRB structure consists of 1×1 convolution (Conv) + batch normalization (BatchNorm) + activation function SiLU, which effectively deepens the network layer and reduces the computational complexity. At the same time, Faster-Block combines the advantages of 3×3 partial convolution (PConv) and IRB, and reuses input features through shortcut connections to avoid the gradient vanishing problem that may occur due to network deepening, such as Figure 5 PConv separates channel cp from the overall feature channel c and performs convolution operations on channel cp, while other channels ccp continue to pass spatial information downward, as shown in (b). Figure 5 as shown in (c).
[0102] Step S40: constructing a head network: designing and implementing a decoupled head network structure, enhancing spatial perception capability and optimizing parameter efficiency through a convolution algorithm;
[0103] S401: Design DD-Head to achieve dual decoupling of tasks and structures;
[0104] As the last part of object detection, the role of the head network is to predict the location and category of the object using the extracted and fused features.
[0105] DD-Head achieves dual decoupling in tasks and structures, and is also more lightweight. Specifically, SAConv with the ability to generate spatial coordinate information is introduced in the regression branch to enhance spatial perception. At the same time, PConv is introduced into the classification branch to reduce the parameters and FLOPs cost of redundant operations, such as Figure 6 shown.
[0106] S402: Introduce SAConv in the regression branch to enhance spatial perception capabilities;
[0107] The input of SAConv is a feature map with height h, width w, and number of channels c. First, two average pooling layers are used to encode and aggregate features for each channel in the horizontal and vertical directions to generate a pair of spatial feature maps, which are then concatenated together. Next, a spatially aware feature map with precise location information of insulators and ISDs (insulator self-explosion defects) is generated through a 1×1 convolution (Conv) + BatchNorm + activation function Hard Swish module, and is divided in the horizontal and vertical directions. The number of channels is adjusted through a 1×1 convolution, and spatially aware weights with horizontal and vertical directions are generated through a Sigmoid function. The coordinate information of the insulator and ISD is embedded in the original feature map. Finally, a 3×3 partial convolution (PConv) learns direction and position awareness information on some channels, while the remaining channels continue to pass on spatial features.
[0108] S403: Apply PConv in the classification branch to reduce the parameter and FLOPs cost;
[0109] In order to capture the spatial features with precise location information of insulators and ISDs, two global average pooling methods are used to aggregate spatial features along the horizontal and vertical directions of each channel, respectively, to generate spatial feature maps in two directions. Subsequently, the concatenated spatially-aware feature maps are encoded using a convolutional layer with batch normalization and activation function to generate spatially-aware feature maps with precise location information, and converted into spatially-aware feature maps containing both horizontal and vertical feature information. Next, two spatially-aware weights with horizontal and vertical spatial information are generated by 1×1 convolution and sigmoid function. Finally, the spatially-aware weights are embedded into the original feature map to obtain spatially-aware feature maps with long-range interactions between insulator and ISD spatial features.
[0110] Overall, DD-Head solves the problems of insufficient spatial perception capabilities and redundant parameters and floating-point operations (FLOPs) based on the dual decoupling of tasks and structures.
[0111] Step S50: training the model: using customized loss functions for model training, including binary cross entropy loss for classification tasks, and distributed focal loss and complete intersection-over-union loss for regression tasks;
[0112] S501: Training for classification tasks using binary cross entropy (BCE) loss;
[0113] S502: Apply distributed focal loss (DFL) and complete intersection over union (CIoU) for training the regression task;
[0114] Step S60: Experimental verification and optimization: The model is tested in a small target scene shot by a real drone and compared with existing methods to evaluate the performance of the model. The model is further optimized based on experimental data to improve detection accuracy and efficiency.
[0115] S601: Test the model performance in the ISD small target scene photographed by the drone;
[0116] S602: Compare with the specified defect detection method to verify the detection accuracy and real-time performance of the model;
[0117] S603: Optimize the model according to the experimental results to ensure detection accuracy and efficiency.
[0118] Figure 7 The test results of different defect detection methods in the ISD small target scene taken by a drone are shown. The red box represents the insulator and the light blue box represents the ISD. It can be seen that LA-YOLOs ( Figure 7 (i)) is not only better than the defect detection method ( Figure 7 (a)-(h) Figure 7 The (j) in the figure has a higher confidence level and can detect more insulators and ISD small targets. The detection results of LA-YOLOn are also better than those of other lightweight methods.
[0119] Step S70: Model application: using the optimized model to detect insulator self-explosion defects.
[0120] Specifically, the trained model is deployed to the inspection system to perform real-time inspection on insulator images taken by drones and output defect information so that necessary maintenance can be performed based on the defect information.
[0121] Furthermore, the defect information includes defect location and confidence level.
[0122] In summary, the insulator self-explosion defect detection method based on the lightweight adaptive YOLO model provided in this embodiment has the following advantages:
[0123] ① Network lightweighting and multi-scale feature capture: By designing a lightweight network architecture, including customized convolution blocks and feature extraction modules, the complexity and computational complexity of the model are effectively reduced, while capturing multi-scale features, which is beneficial for the detection of small targets such as insulator self-explosion defects (ISDs).
[0124] ② Efficient feature fusion mechanism: The C2f module and Bi-AFPN-P2 method are used to strengthen the feature information flow, and multi-scale feature fusion is achieved through self-learning method, ensuring that the original feature information is retained during the fusion process, thereby improving the detection capability of small targets.
[0125] ③ Optimize gradient diversion: The Faster-C2F module and partial convolution algorithm are introduced to reduce feature redundancy, optimize feature transfer, maintain the integrity of spatial information, and further improve the model's detection performance for small targets.
[0126] ④ Decoupled head network structure: The design of DD-Head achieves dual decoupling of tasks and structures, which not only enhances spatial perception capabilities, but also optimizes parameter efficiency, making the model more lightweight while ensuring accuracy.
[0127] ⑤ Customized loss function: By combining binary cross entropy loss, distributed focal loss and complete intersection-over-union loss, it provides a more effective training strategy for classification and regression tasks, which helps to improve the detection accuracy of the model.
[0128] In general, the method provided in this embodiment significantly improves the accuracy and efficiency of insulator self-explosion defect detection while keeping the model lightweight through strategies such as multi-scale feature fusion, optimized gradient diversion and decoupled head network, and has high practical value.
[0129] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of this application, this does not limit the scope of patent protection of this application. All technical solutions generated by replacing or modifying equivalent structures or equivalent processes based on the essential concept of this application using the contents recorded in the specification and drawings of this application, as well as directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are included in the scope of patent protection of this application.
Claims
1. A method for detecting insulator self-explosion defects based on a lightweight adaptive YOLO model, characterized in that: include: Step S10: Design a lightweight network architecture: construct an efficient basic network unit, including customized convolution blocks and feature extraction modules, to achieve network lightweight and multi-scale feature capture; Step S20: Implement feature fusion: Develop feature enhancement and fusion mechanisms, including designing dedicated modules to enhance feature information flow, using self-learning methods to achieve multi-scale feature fusion, and ensuring that the original feature information is preserved during the fusion process; Step S30: Optimizing gradient splitting: introducing advanced convolution modules to reduce feature redundancy, and optimizing feature transfer by applying partial convolution algorithms at key layers to maintain the integrity of spatial information; Step S40: constructing a head network: designing and implementing a decoupled head network structure, enhancing spatial perception capability and optimizing parameter efficiency through a convolution algorithm; Step S50: training the model: using customized loss functions for model training, including binary cross entropy loss for classification tasks, and distributed focal loss and complete intersection-over-union loss for regression tasks; Step S60: Experimental verification and optimization: The model is tested in a small target scene shot by a real drone and compared with existing methods to evaluate the performance of the model. The model is further optimized based on experimental data to improve detection accuracy and efficiency. Step S70: Model application: using the optimized model to detect insulator self-explosion defects.
2. The insulator self-explosion defect detection method based on the lightweight adaptive YOLO model according to claim 1 is characterized in that: The S10 includes: S101: Define the basic convolution module, including k×k convolution layer, batch normalization and activation function; S102: Construct a spatial pyramid pooling layer to extract feature representations at different scales.
3. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 2 is characterized in that: The S20 includes: S201: Design C2f module and use bottleneck layer to enhance feature information flow; S202: Using the Bi-AFPN-P2 network structure, the feature fusion of insulators and ISDs at different scales is achieved through self-learning; S203: Introduce a horizontal flow feature path to ensure that the original feature information is not lost during the feature fusion process.
4. The insulator self-explosion defect detection method based on the lightweight adaptive YOLO model according to claim 3 is characterized in that: The S202 includes: Step S2021: extracting features from the input image to generate a multi-scale feature map; Step S2022: input the generated feature map into the AFM, and adaptively adjust the weight and fusion strategy of the feature map; Step S2023: bidirectionally fuse the adjusted feature map through the context fusion path of Bi-AFPN; Step S2024: Output the fused feature map for subsequent defect detection and positioning.
5. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 3 is characterized in that: The S30 includes: S301: Introduce the Faster-C2F module and use partial convolution to reduce feature redundancy; S302: Implement PConv in the bottleneck layer of the C2f module to retain shallow spatial information.
6. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 5 is characterized in that: The S40 includes: S401: Design DD-Head to achieve dual decoupling of tasks and structures; S402: Introduce SAConv in the regression branch to enhance spatial perception capabilities; S403: Apply PConv in the classification branch to reduce parameter and FLOPs cost.
7. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 6 is characterized in that: The S50 includes: S501: Training for classification tasks using binary cross entropy loss; S502: Apply distributed focal loss and complete intersection-over-union loss to train the regression task.
8. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 7 is characterized in that: The S60 includes: S601: Test the model performance in the ISD small target scene photographed by the drone; S602: Compare with the specified defect detection method to verify the detection accuracy and real-time performance of the model; S603: Optimize the model according to the experimental results to ensure detection accuracy and efficiency.
9. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 7 is characterized in that: The S70 includes: The trained model is deployed to the inspection system to perform real-time inspection on insulator images taken by drones and output defect information so that necessary maintenance can be performed based on the defect information.
10. The insulator self-explosion defect detection method based on lightweight adaptive YOLO model according to claim 9 is characterized in that: The defect information includes defect location and confidence level.