Substation equipment infrared target detection method, device and system, and storage medium

Abstract

The invention discloses a substation equipment infrared target detection method, device and system, and a storage medium. The method comprises the following steps: S1, acquiring an infrared image of power equipment; s2, training an improved YOLOv9 network according to the infrared image of the power equipment; and S3, inputting a to-be-processed infrared image of the power equipment into the trained YOLOv9 network for target detection. According to the technical scheme, the problems that a traditional detection network is low in speed and large in size, the speed precision cannot be balanced, feature fusion is insufficient and the like are solved, the real-time detection requirement is met, meanwhile, the excellent detection precision is achieved, and the detection performance is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and device, a system and a storage medium for infrared target detection of substation equipment. Background Art

[0002] Maintaining the safety and stability of power equipment in substations is an important task in the operation and maintenance of power systems. It is necessary to monitor the operating status of power equipment in a timely manner to effectively prevent safety accidents caused by equipment defects and failures. Visible light images can clearly and intuitively present the morphological characteristics of power equipment, which is convenient for staff to quickly identify. However, visible light images have certain requirements for the environment. For example, light and weather factors will affect the imaging quality of visible light images. Infrared thermal imaging technology has the characteristics of all-weather monitoring. At night or in harsh environments, infrared images can provide better target detection and recognition capabilities and have good robustness. At the same time, infrared images are sensitive to heat source targets, can detect the temperature information of power equipment, and have a good monitoring effect on equipment heating failures. Therefore, target detection based on infrared images has become a hot topic in current power vision research, and how to accurately locate power equipment in infrared images has become a key issue in power intelligent inspection, which has great research significance.

[0003] Using traditional manual methods to detect target devices in infrared images is a time-consuming, labor-intensive and highly subjective method, which can easily cause fluctuations in detection accuracy. Faced with the huge growth in the number of current power equipment, existing detection methods can no longer meet the needs of the times, and there is an urgent need to improve detection efficiency, accuracy and real-time performance.

[0004] The rapid development of deep learning has made significant breakthroughs in computer vision tasks, and target detection is one of the important tasks of computer vision. Many methods apply neural network technology to target detection in infrared images of substation power equipment. However, there are still some shortcomings in the existing technology of infrared target detection of substation equipment.

[0005] First, although traditional target detection methods are relatively effective in detecting single scenes, they are often unable to cope with scenes with complex and diverse shapes, making it difficult to accurately locate and identify them, resulting in a significant reduction in detection accuracy.

[0006] Secondly, there are many types of power equipment in substations, and each type of equipment may have significant differences in shape, size, texture, etc., which puts higher requirements on the detection model. Traditional deep learning models have limitations in capturing these diverse features, especially when the defect features are unevenly distributed or the data is unbalanced. The model may not be able to fully learn these differences, which in turn affects the detection effect. In addition, in actual power inspection scenarios, infrared images may be affected by factors such as shooting angle and background noise, which further increases the difficulty of detection. Traditional methods are particularly inadequate under these complex conditions.

[0007] Finally, although deep learning models are theoretically capable of processing complex image tasks, they often face bottlenecks in actual deployment, such as high computing resource consumption and slow reasoning speed. Power inspection environments usually require real-time detection, but traditional deep learning models have complex network structures and large number of parameters, which makes it difficult for reasoning speed to meet real-time requirements.

[0008] How to improve the model's reasoning efficiency while maintaining high accuracy is an urgent problem to be solved in the field of infrared target detection for substation equipment using deep learning. Therefore, in the above context, more advanced target detection algorithms are applied to infrared target detection for substation equipment to solve the problems of existing methods and enable them to have excellent detection accuracy while meeting real-time detection requirements. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a method and device, system, and storage medium for infrared target detection of substation equipment, so as to solve the problems of slow speed and large size of traditional detection network, inability to balance speed and accuracy, and insufficient feature fusion, so as to meet the real-time detection requirements while having excellent detection accuracy and improve the detection performance.

[0010] To achieve the above object, the present invention adopts the following technical solution:

[0011] A method for detecting infrared targets of substation equipment, comprising:

[0012] Step S1, obtaining an infrared image of power equipment;

[0013] Step S2, training an improved YOLOv9 network according to infrared images of power equipment;

[0014] Step S3: input the infrared image of the power equipment to be processed into the trained YOLOv9 network for target detection.

[0015] Preferably, the improved YOLOv9 network adopts the single-stage detection network YOLOv9 as the basic detection model, introduces SPD convolution in the backbone network, integrates the CBAM attention module in the Neck network, and adopts Focal-IoU as the loss function of the model.

[0016] Preferably, in step S1, the infrared image of the power equipment is enhanced by rotation, cropping and Gaussian blur processing.

[0017] The present invention also provides an infrared target detection device for substation equipment, comprising:

[0018] An acquisition module, used for acquiring infrared images of power equipment;

[0019] A training module for training an improved YOLOv9 network based on infrared images of power equipment;

[0020] The detection module is used to input the infrared image of the power equipment to be processed into the trained YOLOv9 network for target detection.

[0021] Preferably, the improved YOLOv9 network adopts the single-stage detection network YOLOv9 as the basic detection model, introduces SPD convolution in the backbone network, integrates the CBAM attention module in the Neck network, and adopts Focal-IoU as the loss function of the model.

[0022] Preferably, the acquisition module performs image enhancement on the infrared image of the power equipment by means of rotation, cropping and Gaussian blur processing.

[0023] The present invention also provides an infrared target detection system for substation equipment, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes the infrared target detection method for substation equipment when executed by the processor.

[0024] The present invention also provides a storage medium, on which a computer program is stored, and the computer program executes the infrared target detection method for substation equipment when running.

[0025] The present invention uses a data enhancement method based on the infrared image dataset of substation power equipment to expand the dataset and obtain equipment images of different categories. The single-stage detection network YOLOv9 with excellent performance is used as the basic detection model, and the improved lightweight SPD convolution is used to replace the traditional convolution module to reduce the model volume, reduce the model complexity, and improve the running speed. In order to solve the problem of insufficient model feature fusion, the convolution block attention module is introduced into the neck network to highlight the feature information of the power equipment and improve the richness of feature fusion. The present invention applies the improved YOLOv9 network to the field of infrared target detection of substation equipment, realizes real-time detection, and improves the accuracy of infrared target detection of substation equipment, meeting the deployment needs of relevant departments of power inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] 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 embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0027] Figure 1 is a flow chart of an infrared target detection method for substation equipment according to an embodiment of the present invention;

[0028] Figure 2 Schematic diagram of the network structure of the SPD convolution module according to an embodiment of the present invention;

[0029] Figure 3 Schematic diagram of the structure of the convolutional block attention mechanism module of an embodiment of the present invention;

[0030] Figure 4 It is a schematic diagram of the overall structure of an embodiment of the present invention;

[0031] Figure 5 This is a diagram showing the effect of infrared target detection of substation equipment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Embodiment 1:

[0035] The embodiment of the present invention provides a method for infrared target detection of substation equipment, which solves the problems of slow speed and large volume of traditional detection network, inability to balance speed and accuracy, and insufficient feature fusion, and achieves the purpose of improving detection accuracy under the premise of meeting real-time detection. Figure 1 As shown, the following steps are included:

[0036] S1. Based on the infrared images of power equipment in substations, image enhancement methods (rotation, cropping, Gaussian blur) are used to expand data, build data sets, and obtain images of equipment of different categories;

[0037] S2. Train the improved YOLOv9 network, including:

[0038] S21. Select the YOLOv9 network as the basic framework of the detection network;

[0039] S22, the backbone network improves SPD convolution, reduces model complexity and increases running speed;

[0040] S23. Design a feature fusion method and add a convolutional block attention module to the Neck network to optimize the model's ability to extract key features of power equipment.

[0041] S24. Improve the basic model loss function and use Focal-IoU as the model loss function to enhance the model's discrimination ability and reduce the model's positioning error.

[0042] S3. Experiments were conducted on a dataset of infrared images of power equipment in a substation. The results showed that the embodiments of the present invention have a relatively small number of parameters while ensuring a relatively high detection accuracy.

[0043] Deep learning models require a large number of dataset image samples as support during training. Since it is difficult to collect infrared data of power equipment, most of the existing data are small sample datasets. Therefore, in step S1, the dataset is expanded using image enhancement methods based on the infrared image dataset of substation power equipment to obtain images of different categories of equipment, including:

[0044] The infrared images of power equipment taken by the power inspection infrared sensor are selected, and the images are enhanced by rotation, cropping and Gaussian blurring to expand the data volume and obtain equipment images of different categories.

[0045] In the embodiment of the present invention, considering that most detection network models have too many parameters, high requirements for hardware computing resources, slow detection speed, and cannot meet the needs of real-time detection; at the same time, the large model size also makes it unsuitable for operation platforms with relatively limited hardware resources, and traditional neural networks are lacking in detection accuracy, the embodiment of the present invention introduces a powerful detection network YOLOv9 to improve the accuracy and detection speed of infrared target detection of substation equipment. Among them, in the step S21, the YOLOv9 network is selected as the basic framework of the detection network, specifically including:

[0046] In order to verify the superiority of the embodiment of the present invention in infrared target detection of substation equipment, we compared several common single-stage and double-stage target detection algorithms on the current defect dataset, and finally selected the YOLOv9 model with the best effect as the benchmark.

[0047] In step S22, the backbone network improves the SPD convolution, reduces the model complexity, and improves the running speed, specifically including:

[0048] First, assume that the data input is: X = c 1 ×h×w,c 1 is the number of input channels, h and w represent the height and width of the input feature map.

[0049] Secondly, slice the input data feature map to obtain the sub-feature map f 0,0 , f 0,scale-1 , f scale-1,0 , f scale-1,scale-1 . Where scale = 2, that is, the original feature map is sliced ​​into four sub-feature maps, and the spatial dimension of each sub-feature map is 1 / 2 of the original one.

[0050] Secondly, the obtained sub-feature maps are concatenated in the channel dimension, and the feature map channels obtained are 4c 1 , and the spatial dimension is reduced to half of its original size.

[0051] Finally, the above feature map is convolved without stride (the stride is 1 and the convolution kernel size is 1×1×c 2 ), the strideless convolution can retain all the discriminative feature information as much as possible, and finally the size of (s,s,c 1 ) is output as a feature map of size (s / 2,s / 2,c 2 ) feature map, so that the feature information of the original feature map is retained as much as possible, and the size of the feature map is reduced, further reducing the number of model parameters and improving the efficiency of the model. The model structure is as follows Figure 2 shown.

[0052] In step S23, a feature fusion method is designed, and a convolutional block attention module is added to the Neck network to optimize the model's ability to extract key features of power equipment. Specifically, it includes:

[0053] The convolutional block attention module is mainly composed of two parts: the channel attention module and the spatial attention module. First, in the channel attention module, the module first compresses the spatial dimension of each channel through average pooling and maximum pooling to obtain descriptors F of different spaces. c avg With F c max , then the two descriptors are passed into the shared network, and the channel attention map M is generated through the network c , and finally the channel attention map is obtained. The channel attention can be expressed as:

[0054]

[0055] Among them, σ represents the sigmoid function, W 0 ∈RC / r×C,W 1 ∈RC×C / r. W 0 With W 1 are the two shared weights in the MLP network.

[0056] Secondly, in the spatial attention module, the module first aggregates the channel information of the feature map through average pooling and maximum pooling operations to obtain two 2D feature maps: and That is, the average pooling feature and the maximum pooling feature of each channel, and then the feature map is processed by the convolution layer to generate a spatial attention feature map. The spatial attention can be expressed as:

[0057]

[0058] Where σ represents the sigmoid function, f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0059] Finally, the two attention modules are arranged in sequence to obtain the adjusted convolutional block attention module.

[0060] In step S24, the basic model loss function is improved, and Focal-IoU is used as the model loss function, which strengthens the model's discrimination ability and reduces the model's positioning error. Specifically, it includes:

[0061] First, calculate the most common target detection evaluation standard IoU, which is defined as follows:

[0062]

[0063] Among them, B and B gt Represent the predicted box and the true box respectively.

[0064] Then, based on IoU, CIoU is calculated, which is defined as follows:

[0065]

[0066] Among them, b gt and b represent the center points of the predicted box and the real box respectively, ρ is the Euclidean distance, and c is the distance between b and b gt The minimum diagonal distance between the enclosing boxes. gt With h gt is the width and height of the real box, w and h are the width and height of the predicted box.

[0067] In order to focus on different samples and thus different application scenarios, a linear interval mapping method is used to reconstruct IoU. The reconstructed IoU expression is as follows:

[0068]

[0069] Among them, IoU focaler is the reconstructed IoU, where IoU is the initial value of CIoU calculated by (4), [d,u]∈[0,1], IoU focaler The values ​​of d and u in the formula can be adjusted to adapt to different regression samples. The loss is defined as follows:

[0070] L Focaler-IoU =1-IoU focaler (8)

[0071] Applying formula (9) to the IoU-based bounding box regression loss function, the final loss function expression is as follows: (9)

[0072] L Focaler-CIoU =L CIoU +IoU-IoU Focaler

[0073] Further, if Figure 4 As shown, the processing flow of the improved YOLOv9 network is as follows:

[0074] First, the raw image data is obtained as input. These images come from infrared images of power equipment obtained during power inspections at substations of the Southern Power Grid. Before being input into the model, these images undergo a series of preprocessing steps, including resizing the images to fit the model’s input requirements, normalization to ensure that the data is at a uniform scale, and data augmentation operations such as rotation, flipping, or cropping to enhance the model’s generalization capabilities.

[0075] The preprocessed image is then fed into the backbone network of the model. In the improved YOLOv9 network, the backbone network uses convolutional layers, SPD convolutional layers, and ResNSCPELAN4 modules connected layer by layer, which are designed to extract key features in the image at a lower computational cost. These feature maps contain key information of the target object in the image, such as shape, texture, and position. At the same time, the extracted features are input into the CBAM attention module so that the present invention can focus on key areas and features in the image.

[0076] Next, the feature map is fed into the feature pyramid network for multi-scale fusion. The feature pyramid network fuses feature maps of different scales through operations such as upsampling, downsampling, and lateral connections to generate multi-scale feature maps with rich contextual information. These feature maps help the model detect target objects at different scales, especially for the detection of small targets and occluded targets.

[0077] After being processed by the feature pyramid network, the multi-scale feature map is sent to the detection head for target detection. The detection head usually contains a series of convolutional layers and classification / regression layers to predict the category, location, and confidence of the target object. In the improved YOLOv9 network, the detection head FCOS head and other structures are used to improve the detection accuracy and speed. At the same time, the improved YOLOv9 network also introduces programmable gradient information (PGI) technology to generate reliable gradients through auxiliary reversible branches to optimize the representation ability of deep features.

[0078] Finally, the target detection results output by the detection head will go through post-processing steps, including non-maximum suppression (NMS), to remove redundant bounding boxes and retain the best detection results. NMS compares the confidence and overlap of adjacent bounding boxes and selects the bounding box with the highest confidence and the smallest overlap as the final detection result. These results contain information such as the category, location (bounding box coordinates) and confidence of the target object, which can be used for subsequent decision-making or visualization.

[0079] In step S3, experiments were conducted on infrared image datasets of power equipment in substations. The results show that the embodiments of the present invention have a relatively small number of parameters while ensuring a relatively high detection accuracy. Specifically, the embodiments include:

[0080] A trained model usually requires some evaluation indicators to judge the performance. The embodiment of the present invention adopts multiple evaluation indicators, including mean average precision (mAP) and inference speed (FPS).

[0081] mAP is the most commonly used evaluation index in target detection, which measures the average detection accuracy under different thresholds. This article uses mAP@0.5 to measure the accuracy of detection. FPS is the number of image frames processed per second. The higher the value, the better the real-time performance of the model. The embodiment of the present invention uses this index to evaluate the inference speed of the model.

[0082] The performance is verified on an infrared image dataset of power equipment in a substation, and the verification results show that the embodiments of the present invention can effectively improve the comprehensive performance of infrared target detection of substation equipment.

[0083] The detection effect of the embodiment of the present invention is as follows Figure 5 As shown. Firstly, in terms of expanding the data set, the present invention proposes to expand the data by using image enhancement methods (rotation, cropping, Gaussian blur, etc.) based on the initial data. Secondly, based on the YOLOv9 model, the improved lightweight SPD convolution is used to replace the traditional convolution module to reduce the model volume, reduce the model complexity, and improve the running speed. In view of the problem of insufficient model feature fusion, the convolution block attention module is introduced into the neck network to highlight the feature information of the power equipment and improve the richness of feature fusion. The present invention effectively improves the detection performance of the YOLOv9 network and achieves the best trade-off between speed and accuracy.

[0084] Embodiment 2:

[0085] The present invention also provides an infrared target detection device for substation equipment, comprising:

[0086] An acquisition module, used for acquiring infrared images of power equipment;

[0087] A training module for training an improved YOLOv9 network based on infrared images of power equipment;

[0088] The detection module is used to input the infrared image of the power equipment to be processed into the trained YOLOv9 network for target detection.

[0089] As an implementation method of an embodiment of the present invention, the improved YOLOv9 network adopts the single-stage detection network YOLOv9 as the basic detection model, introduces SPD convolution in the backbone network, integrates the CBAM attention module in the Neck network, and adopts Focal-IoU as the loss function of the model.

[0090] As an implementation method of the embodiment of the present invention, the acquisition module performs image enhancement on the infrared image of the power equipment by using rotation, cropping and Gaussian blur processing.

[0091] Embodiment 3:

[0092] The present invention also provides an infrared target detection system for substation equipment, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes the infrared target detection method for substation equipment when executed by the processor.

[0093] Embodiment 3:

[0094] The present invention also provides a storage medium, on which a computer program is stored, and the computer program executes the infrared target detection method for substation equipment when running.

[0095] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for infrared target detection of substation equipment, characterized in that: include: Step S1, obtaining an infrared image of power equipment; Step S2, training an improved YOLOv9 network according to infrared images of power equipment; Step S3: input the infrared image of the power equipment to be processed into the trained YOLOv9 network for target detection.

2. The infrared target detection method for substation equipment according to claim 1, characterized in that: The improved YOLOv9 network uses the single-stage detection network YOLOv9 as the basic detection model, introduces SPD convolution in the backbone network, integrates the CBAM attention module into the Neck network, and uses Focal-IoU as the loss function of the model.

3. The infrared target detection method for substation equipment according to claim 2, characterized in that: In step S1, the infrared image of the power equipment is enhanced by rotation, cropping and Gaussian blur processing.

4. An infrared target detection device for substation equipment, characterized in that: include: An acquisition module, used for acquiring infrared images of power equipment; A training module for training an improved YOLOv9 network based on infrared images of power equipment; The detection module is used to input the infrared image of the power equipment to be processed into the trained YOLOv9 network for target detection.

5. The infrared target detection device for substation equipment according to claim 4, characterized in that: The improved YOLOv9 network uses the single-stage detection network YOLOv9 as the basic detection model, introduces SPD convolution in the backbone network, integrates the CBAM attention module into the Neck network, and uses Focal-IoU as the loss function of the model.

6. The infrared target detection device for substation equipment according to claim 5, characterized in that: The acquisition module uses rotation, cropping and Gaussian blur processing to enhance the infrared image of power equipment.

7. A substation equipment infrared target detection system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the infrared target detection method for substation equipment according to any one of claims 1 to 3 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, and the computer program executes the infrared target detection method for substation equipment according to any one of claims 1 to 3 when running.

Images