Defect detection method and device based on improved YOLOv11 under power transformation scene and computer readable storage medium

Abstract

This invention belongs to the field of target detection technology in substation scenarios, and more specifically, relates to an improved YOLOv11 defect detection method, device, and computer-readable storage medium based on substation scenarios. The method includes: constructing a defect detection dataset for substation scenarios; constructing an improved YOLOv11 algorithm model: proposing a SACELU activation function to replace the original activation function, and simultaneously proposing an HRF module to replace the standard c3k2 structure, effectively improving the model's spatial semantic understanding of equipment defects through a multi-level receptive field fusion mechanism; training a defect detection model on the improved YOLOv11 model, and deploying the trained model and algorithm for defect detection in substation scenarios. This invention solves the problem that a relatively small receptive field makes it difficult to effectively capture the semantic relationship between equipment defects and the surrounding environment, leading to misjudgments of highly similar defect types.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology in substation scenarios, and more specifically, relates to a defect detection method, device and computer-readable storage medium based on an improved YOLOv11 for substation scenarios. Background Technology

[0002] With the rapid development and widespread application of intelligent inspection technology in the power sector, the performance bottleneck of traditional defect detection algorithms in complex substation scenarios is becoming increasingly apparent. Current substation equipment operating environments are characterized by high equipment density, complex structures, diverse defect characteristics, and strong mutual interference. Defect identification in such complex scenarios often requires comprehensive analysis of the global context information of the target object.

[0003] Chinese patent document CN116030360A discloses a lightweight multi-scale remote sensing image rotation target detection method and system. Based on the YOLOv4 model, it incorporates shallow feature information into the feature fusion network. To alleviate feature loss caused by excessive network layers, a weighted bidirectional feature pyramid is used as the feature fusion network. To improve the receptive field of features, a multi-scale dilated convolution module is added after feature fusion to enhance detection accuracy. For remote sensing image rotation target detection, the angle parameter is represented by the long side representation method, and the angle parameter is transformed from a regression problem to a classification problem using a Gaussian function to smooth the label, reducing the loss caused by the periodicity of the angle. A lightweight module is designed using depthwise separable convolution and a global attention module to replace traditional convolution in the detection model.

[0004] In summary, while the YOLO series of algorithms currently perform well in object detection, they still have limitations when dealing with certain challenges. Their receptive field is relatively small, making it difficult to effectively capture long-distance dependencies and global contextual information in images. They also struggle to effectively capture the semantic relationships between device defects and their surrounding environment, leading to misjudgments of highly similar defect types. In particular, the accuracy of defect identification needs improvement when considering the overall state of the device. Summary of the Invention

[0005] This invention aims to overcome at least one of the shortcomings of the existing technologies and provide an improved YOLOv11 defect detection method based on substation scenarios. While the YOLO series algorithms perform well in target detection, they still have limitations when handling certain challenges. Their receptive field is relatively small, making it difficult to effectively capture long-distance dependencies and global contextual information in images. Furthermore, they struggle to effectively capture the semantic relationships between equipment defects and their surrounding environment, leading to misjudgments of highly similar defect types. In particular, the accuracy of defect identification needs improvement when considering the overall state of the equipment.

[0006] The detailed technical solution of this invention is as follows: An improved YOLOv11 defect detection method based on substation scenarios, the method comprising: S1. Collect substation defect data to construct a defect dataset, and perform preprocessing and annotation; S2. Establish a defect detection model based on the YOLOv11 algorithm, train the defect detection model using a defect dataset, use the Adam optimizer during training, train for Z rounds, and use an image size of 640*640. S3. Deploy the defect detection model and perform defect detection to obtain the defect detection results; S4. The obtained defect detection results are post-processed to improve the accuracy of defect detection. Then, the processed defect detection results are visualized and superimposed on the original image for display.

[0007] Furthermore, the preprocessing includes: data cleaning and data augmentation; Data cleaning: Perform preliminary screening on the collected raw data to remove images of poor quality or containing noise; Data augmentation: Various data augmentation methods are used to process the labeled data. These augmentation methods include rotation, translation, cropping, mirroring, noise addition, brightness adjustment, and contrast adjustment.

[0008] Furthermore, the defect detection model is an improved YOLOv11 model, specifically including: S21. Replace the activation function of the YOLOv11 model with the SACELU activation function. Furthermore, the SACELU activation function constructs a spatial dynamic adjustment mechanism through depthwise separable convolution, adaptively capturing local context features and significantly enhancing the model's region perception capability. At the same time, through the negative interval exponential smoothing strategy, it can maintain stable gradient propagation. Its negative interval adopts an exponential smoothing strategy, which effectively alleviates the gradient truncation phenomenon while preserving gradient continuity, thereby comprehensively improving the nonlinear modeling performance of the network. The SACELU (Spatially Adaptive CELU) activation function has the following expression: (1); In formula (1), T(x) is the spatial conditional feature generated by depthwise separable convolution, x is the input value of the activation function, the specific value is determined according to the location, α is the learnable spatial conditional weight coefficient, and β is the value that controls the smoothness of the negative region.

[0009] S22. Replace the c3k2 module of the YOLOv11 model with the HRF module. The HRF module adopts a cascaded structure of depthwise convolution-point convolution-dilated convolution, which forms a feature extraction path through two stacks and introduces cross-layer residual connections. Among them, depthwise convolution performs lightweight channel-wise spatial feature extraction of input features; pointwise convolution achieves cross-channel information fusion through a 1×1 kernel, enhancing feature expression capabilities; dilated convolution adopts an interval sampling strategy to effectively expand the receptive field and capture a wide range of contextual features without increasing the number of parameters; and residual connections preserve the original feature distribution through identity mapping.

[0010] Furthermore, S3 specifically includes: S31. Use the ONNX toolkit to convert the trained pt format defect detection model and its weights into an ONNX format model. S32. Use the ATC model conversion tool in the Ascend toolkit to convert the ONNX format model into the OM format model supported by the Ascend310 chip. S33. Implement the model inference service based on the ais_bench and aclruntime packages, and perform defect detection in substation scenarios on the Ascend310 chip.

[0011] Furthermore, the post-processing includes bounding box parsing, confidence filtering, and non-maximum suppression (NMS) steps to improve the accuracy of defect detection.

[0012] In another aspect of the invention, an apparatus is provided for an improved YOLOv11 defect detection method based on a substation scenario, the apparatus comprising: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the improved YOLOv11 defect detection method based on the substation scenario as described above.

[0013] In another aspect of the invention, a computer-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform the improved YOLOv11 defect detection method based on a substation scenario as described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The improved YOLOv11 defect detection method, device and computer-readable storage medium based on substation scenario provided by the present invention address the technical difficulties in the field of power equipment defect detection, such as the strong dynamic changes of defect types and the dependence of the discrimination process on context semantic features. It proposes the SACELU activation function, which adds spatial context information to the activation function and effectively alleviates the gradient truncation phenomenon while preserving gradient continuity. From the perspective of activation function, it improves the network's overall ability to perceive context information.

[0015] (2) The present invention provides an improved YOLOv11 defect detection method, device, and computer-readable storage medium based on substation scenarios. It proposes an HRF module, which employs a cascaded structure of depthwise convolution, pointwise convolution, and dilated convolution. This structure forms a feature extraction path through two stacking operations and introduces cross-layer residual connections. This significantly reduces computational costs, enhances feature representation capabilities, and effectively expands the receptive field without increasing the number of parameters, capturing a wide range of contextual features. The residual connections preserve the original feature distribution through identity mapping, alleviating the gradient vanishing problem. It can simultaneously process local and global information, effectively capture and fuse image contextual information, and significantly improve the model's ability to utilize contextual information, overcoming the deficiency of insufficient contextual information utilization in traditional detection methods. Attached Figure Description

[0016] Figure 1 This is a flowchart of the improved YOLOv11 defect detection method based on substation scenarios described in this invention.

[0017] Figure 2 This is a schematic diagram of the HRF module in Embodiment 1 of the present invention.

[0018] Figure 3 This is a schematic diagram of the improved model structure based on YOLOv11 in Embodiment 1 of the present invention.

[0019] Figure 4 This is a schematic diagram of the ground oil pollution defect detection results in Embodiment 1 of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0024] Embodiment 1 Refer Figure 1 , this embodiment provides a defect detection method for improving YOLOv11 based on a substation scenario, and the method includes: S1. Collect substation defect data, and the specific operations for constructing a substation defect data set are as follows: Data collection: Technicians collect images in the actual operation scenario of the substation. During collection, ensure that images at different positions and angles within the substation are covered to ensure the diversity and comprehensiveness of the data. Use high-quality photographic equipment to ensure image clarity and detail display, and capture different types of defects, such as equipment corrosion, wear, poor contact, etc. Data cleaning: Conduct a preliminary screening of the collected original data and剔除 images with poor quality or containing noise. During the data cleaning process, artificial inspection of the images will be carried out to remove those images that cannot accurately identify defects due to insufficient light, out-of-focus, severe occlusion, or other factors. In addition, duplicate images will also be deleted.

[0025] Data augmentation: To increase the diversity of the data set, improve the robustness and generalization ability of the model, various data augmentation methods are used to process the labeled data. The augmentation methods include: rotation, translation, cropping, mirror flipping, noise addition, brightness, and contrast adjustment.

[0026] Data annotation: After data cleaning, manual defect annotation is carried out. Professional personnel accurately annotate the defects in each image, and the annotation content includes the location, type, and possible severity of the defects.

[0027] S2. Construct a defect detection model based on the YOLOv11 algorithm, and use the substation defect data set to train the constructed defect detection model; during the training process, use the Adam optimizer, the number of training rounds is Z rounds, and the image size is 640*640.

[0028] Preferably, the defect detection model is an improved YOLOv11 model, specifically including: S21. Replace the activation function of the YOLOv11 model with the SACELU activation function; Specifically, the activation function is used to map the input signal into a nonlinear output. The original activation functions in the backbone and neck are replaced with the SACELU activation function. The SACELU activation function constructs a spatial dynamic adjustment mechanism through depthwise separable convolution, adaptively capturing local contextual features and significantly enhancing the model's region perception ability. At the same time, through the negative interval exponential smoothing strategy, the gradient propagation is kept stable. The negative interval adopts an exponential smoothing strategy, which effectively alleviates the gradient truncation phenomenon while preserving the gradient continuity, thereby comprehensively improving the nonlinear modeling performance of the network.

[0029] The SACELU (Spatially Adaptive CELU) activation function has the following expression: (1); In formula (1), T(x) is the spatial conditional feature generated by depthwise separable convolution, x is the input value of the activation function, the specific value is determined according to the location, α is the learnable spatial conditional weight coefficient, and β is the value that controls the smoothness of the negative region.

[0030] S22. Replace the c3k2 module in the YOLOv11 model with the HRF module (Hybrid-Receptive FusionModule), as follows: Figure 2 As shown, the HRF module adopts a cascaded structure of depthwise convolution-point convolution-dilated convolution, forming a feature extraction path through two stacks, and introducing cross-layer residual connections; Specifically, depthwise convolution performs lightweight, channel-wise spatial feature extraction from the input features, significantly reducing computational costs while preserving local details; pointwise convolution achieves cross-channel information fusion through a 1×1 kernel, enhancing feature representation capabilities; dilated convolution employs an interval sampling strategy, effectively expanding the receptive field and capturing a wide range of contextual features without increasing the number of parameters. Residual connections preserve the original feature distribution through identity mapping, alleviating the gradient vanishing problem; the HRF module can process both local and global information simultaneously, effectively capturing and fusing contextual information from the image. This multi-layered feature extraction capability enables the model to effectively identify various defects in substation scenarios.

[0031] It possesses multi-level feature extraction capabilities, high computational efficiency, adaptability to complex substation environments, and powerful contextual feature extraction capabilities. The improved YOLOv11 model structure diagram is shown below. Figure 3 As shown, the c3k2 modules in both the backbone and neck networks of the YOLOv11 model are replaced with HRF modules.

[0032] SACELU's negative interval exponential smoothing strategy ensures stable gradient propagation, forming a dual protection mechanism with the residual connections of the HRF module to jointly alleviate the gradient vanishing problem and ensure the training stability of deep networks.

[0033] S3. Deploy the defect detection model and perform defect detection on the image to be detected to obtain the defect detection results; First, use appropriate tools to convert the trained model weights into a universal format to adapt to different hardware environments. Then, use model conversion tools supported by the AI ​​inference chip to convert the model format into a version compatible with the target chip to optimize the computing architecture and improve inference performance. Finally, use an inference framework or tool to complete the inference deployment of the model.

[0034] Deploying the defect detection model includes loading the model, performing input data preprocessing, inference computation, and parsing the output results. Finally, the optimized algorithm, namely the improved YOLOv11 model and its weight file, is fully deployed to the AI ​​inference chip to ensure that its defect detection task in the substation scenario can run stably and efficiently. The improved YOLOv11 model and its weight file are versions compatible with the target chip.

[0035] Furthermore, S3 specifically includes: S31. Use the ONNX toolkit to convert the trained pt format defect detection model and its weights into an ONNX format model. S32. Use the ATC model conversion tool in the Ascend toolkit to convert the ONNX format model into the OM format model supported by the Ascend310 chip. S33. Implement the model inference service based on the ais_bench and aclruntime packages, and perform defect detection in substation scenarios on the Ascend310 chip.

[0036] When deploying the defect detection model, the onnx toolkit, Ascend toolkit, ais_bench and aclruntime packages are used to jointly implement the AL deployment of the model of this invention, which is more suitable for the target detection field in substation scenarios.

[0037] S4. The obtained defect detection results need to be post-processed to improve the accuracy of defect detection. Then, the processed defect detection results are visualized and superimposed on the original image for display.

[0038] The post-processing includes bounding box parsing, confidence filtering, and non-maximum suppression (NMS) steps to improve the accuracy of defect detection.

[0039] The defect detection process of this invention is as follows: (1) Input the image to be tested: First, input the substation scene image to be tested into the AI ​​inference chip. The input image needs to be properly preprocessed, such as size adjustment and normalization, to meet the input requirements of the model and ensure the accuracy and stability of the inference process.

[0040] (2) Terminal inference: On the AI ​​inference chip, the deployed model is loaded using inference tools (such as ais_bench or aclruntime) and inference computation is performed. During the inference process, the model will perform feature extraction and target detection on the input image, identify possible defect areas, and generate corresponding detection results.

[0041] (3) The obtained detection results need to be post-processed, including bounding box parsing, confidence screening, non-maximum suppression and other steps, in order to improve the accuracy of defect detection.

[0042] (4) Finally, the processed defect detection results are visualized and superimposed on the original image to intuitively display the detected defect areas and provide corresponding output data for subsequent analysis and decision-making.

[0043] Ground oil stain defect detection results as follows Figure 4 As shown in the figure, the red box represents the detection box for oil stain defects on the ground, and the green box represents the detection box for oil stain defects on the surface of the component. Figure 4 The image in (a) shows the detection results of the original YOLOv11 model, which can be seen that it can only identify relatively obvious oil stain defects on the ground. Figure 4 (b) shows the detection results of the YOLOv11 model with only the improved activation function. It can be seen that one ground oil stain defect and one component surface oil stain defect were identified, with confidence levels of 0.895 and 0.432, respectively. Figure 4 The middle (c) diagram shows the detection results of the YOLOv11 model using only the HRF module. It can be seen that one ground oil stain defect and one component surface oil stain defect were identified, with confidence levels of 0.885 and 0.456 respectively. Figure 4 Figure (d) shows the detection results of the method used in this invention. It can be seen that one ground oil stain defect and one component surface oil stain defect were identified, with confidence levels of 0.915 and 0.521, respectively.

[0044] The original model was significantly inferior to the improved model in both the accuracy of the detection box location and the confidence level of the defect category, and it also missed detecting oil stains on relatively hidden component surfaces. In contrast, the improved model, thanks to a more effective local and contextual information awareness mechanism, showed significant improvements in both the accuracy and stability of defect detection.

[0045] The SACELU activation function uses a spatial dynamic adjustment mechanism built through depthwise separable convolution to adaptively enhance the response intensity of key regions, while the depthwise convolution in the HRF module also extracts local detail features in a lightweight manner. The combination of the two significantly improves the model's sensitivity to local structures.

[0046] SACELU's negative interval exponential smoothing strategy ensures stable gradient propagation, forming a dual protection mechanism with the residual connections of the HRF module to jointly alleviate the gradient vanishing problem and ensure the training stability of deep networks.

[0047] Therefore, the original model could not accurately distinguish whether the oil stain was located on the ground or on the surface of the component. However, the present invention, by replacing the custom SACELU activation function, is more suitable for the recognition in the field to which the present invention pertains. At the same time, by replacing the c3k2 module of the YOLOv11 model with the HRF module and introducing cross-layer residual connections, the improved model shows a significant improvement in oil stain location recognition due to the introduction of a more effective local context information perception mechanism.

[0048] Example 2 This embodiment provides an apparatus for implementing an improved YOLOv11 defect detection method based on a substation scenario. The apparatus includes: At least one processor; and The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the improved YOLOv11 defect detection method based on the substation scenario as described above.

[0049] In this embodiment, electronic devices include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile computing devices, smartphones, tablet computers, cellular phones, personal digital assistants (PDAs), handheld devices, messaging devices, wearable computing devices, consumer electronic devices, etc.

[0050] Example 3 This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed, cause the machine to perform the improved YOLOv11 defect detection method based on a substation scenario as described above.

[0051] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0052] In this case, the program code itself, which can be read from the readable medium, can perform the functions of any of the above embodiments, and therefore the computer-readable code and the readable storage medium storing the computer-readable code constitute a part of this specification.

[0053] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0054] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A defect detection method based on improved YOLOv11 of power transformation scene, characterized in that, The method includes: S1. Collect substation defect data to construct a defect dataset, and perform preprocessing and annotation; S2. Establish a defect detection model based on the YOLOv11 algorithm, train the defect detection model using a defect dataset, use the Adam optimizer during training, train for Z rounds, and use an image size of 640*640. The defect detection model is an improved YOLOv11 model, specifically including: S21. Replace the activation function of the YOLOv11 model with the SACELU activation function. The SACELU activation function constructs a spatial dynamic adjustment mechanism through depthwise separable convolution to adaptively capture local context features. At the same time, it maintains stable gradient propagation through an exponential smoothing strategy in the negative interval. The SACELU activation function is as follows: （1）； In formula (1), T(x) is the spatial conditional feature generated by depthwise separable convolution, x is the input value of the activation function, the specific value is determined according to the location, α is the learnable spatial conditional weight coefficient, and β is the value that controls the smoothness of the negative region. S22. Replace the c3k2 module of the YOLOv11 model with the HRF module. The HRF module adopts a cascaded structure of depthwise convolution-point convolution-dilated convolution, which forms a feature extraction path through two stacks and introduces cross-layer residual connections. Among them, depthwise convolution performs lightweight channel-wise spatial feature extraction of input features; pointwise convolution achieves cross-channel information fusion through a 1×1 kernel, enhancing feature expression capabilities; dilated convolution adopts an interval sampling strategy to effectively expand the receptive field and capture a wide range of contextual features without increasing the number of parameters; and residual connections preserve the original feature distribution through identity mapping. S3. Deploy the defect detection model and perform defect detection to obtain the defect detection results; S4. The obtained defect detection results are post-processed to improve the accuracy of defect detection. Then, the processed defect detection results are visualized and superimposed on the original image for display.

2. The method of claim 1, wherein the method is based on an improved YOLOv11 for defect detection in a power transformation scene. The preprocessing includes: data cleaning and data augmentation; Data cleaning: Perform preliminary screening on the collected raw data to remove images of poor quality or containing noise; Data augmentation: Various data augmentation methods are used to process the labeled data. Augmentation methods include rotation, translation, cropping, mirroring, noise addition, brightness and contrast adjustment.

3. The method of claim 2, wherein the method is based on an improved YOLOv11 for defect detection in a power transformation scene. S3 specifically includes: S31. Use the ONNX toolkit to convert the trained pt format defect detection model and its weights into an ONNX format model. S32. Use the ATC model conversion tool in the Ascend toolkit to convert the ONNX format model into the OM format model supported by the Ascend310 chip. S33. Implement the model inference service based on the ais_bench and aclruntime packages, and perform defect detection in substation scenarios on the Ascend310 chip.

4. The method of claim 2, wherein the method is based on an improved YOLOv11 for defect detection in a variable electric field scenario. The post-processing includes bounding box parsing, confidence filtering, and nonmaximum suppression.

5. A device for implementing the defect detection method based on improved YOLOv11 of variable field scenario according to any one of claims 1-4. The device includes: At least one processor; and, The memory stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the improved YOLOv11 defect detection method based on the substation scenario as described above.

6. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 5. When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 4.

Images