An insulator overheating detection method combining RB-YOLO and thermal imaging technology

Abstract

The application discloses an insulator overheating detection method combining RB-YOLO and thermal imaging technology, collects and labels infrared thermal imaging insulator pictures to obtain an insulator dataset, divides the insulator dataset into a training set, a verification set and a test set; inputs the pictures in the training set and the verification set into an RB-YOLO network model for training, obtains optimal weight data of the RB-YOLO network model through parameter adjustment and optimization, loads the optimal weight data to the RB-YOLO network model, and performs identification test on the test set to output an insulator overheating detection result. The RB-YOLO network is improved based on a YOLOv8 network, a new main stem end is introduced through introduction of a main stem network and structural reparameterization design, the feature expression capability is enhanced while the model detection precision and generalization capability are ensured, and the model memory and operation amount are greatly reduced. Compared with a traditional detection method, the insulator overheating detection method combining RB-YOLO and thermal imaging technology is further improved in precision and speed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of insulator detection, and particularly relates to an insulator overheating detection method combining RB-YOLO and thermal imaging technology. BACKGROUND

[0002] At present, it is an essential task to periodically detect insulators in an ultrahigh-voltage substation by using an infrared thermal imager. For about six hundred thermal imaging pictures taken each time, a manual picture-by-picture inspection method is currently adopted. However, the traditional manual inspection has problems such as low efficiency and high work intensity. Meanwhile, due to the large number of devices in a 500kV substation and the dense arrangement of the devices, there is a large background interference in insulator thermal imaging, which undoubtedly increases the difficulty of manual inspection and is prone to cause inaccurate judgment.

[0003] With the development of deep learning, target detection based on a convolutional neural network has surpassed many traditional methods and become the mainstream of current target detection. However, the current target detection needs a large amount of parameters and operations, and is limited in detection accuracy and speed. SUMMARY

[0004] The present application solves the technical problem of providing an insulator overheating detection method combining RB-YOLO and thermal imaging technology, which can improve the speed and accuracy of insulator overheating detection.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0006] An insulator overheating detection method combining RB-YOLO and thermal imaging technology, characterized in that it comprises the following steps:

[0007] Collecting and labeling infrared thermal imaging insulator pictures to obtain an insulator dataset, and dividing the insulator dataset into a training set, a validation set and a test set;

[0008] Inputting pictures in the training set and the validation set into an RB-YOLO network model for training, obtaining optimal weight data of the RB-YOLO network model through parameter optimization, and the RB-YOLO network model is a network model obtained by introducing a backbone network BoTNet and structural reparameterization technology into a YOLOv8 network to design a backbone end;

[0009] Loading the optimal weight data into the RB-YOLO network model, and performing identification testing on the test set to output an insulator overheating detection result.

[0010] The application has the beneficial effects that: infrared thermal imaging insulator pictures are collected and labeled to obtain an insulator dataset, the insulator dataset is divided into a training set, a verification set and a test set; the pictures in the training set and the verification set are input into an RB-YOLO network model for training, and optimal weight data of the RB-YOLO network model is obtained through parameter optimization; the optimal weight data is loaded into the RB-YOLO network model, and the test set is identified and tested, and a detection result of insulator overheating is output. The RB-YOLO network is improved based on a YOLOv8 network, a backbone network BoTNet is introduced, convolution is combined with an attention mechanism, the extraction capability of target features is improved, and a structural reparameterization technology is introduced, so that the model parameter amount is greatly reduced, a new backbone end is designed through the introduction of two module designs, the feature expression capability is enhanced while the model detection precision and generalization capability are ensured, and the model memory and operation amount are greatly reduced. The RB-YOLO network model is combined with infrared thermal imaging of insulators in an ultrahigh voltage substation, has smaller parameter amount and operation amount in insulator heating defect detection, and further improves the precision and speed compared with a traditional detection method. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 FIG. 1 is a flowchart of an insulator overheating detection method combining RB-YOLO and thermal imaging technology according to an embodiment of the application;

[0012] Figure 2 FIG. 2 is a model structure diagram of YOLOv8 in the prior art;

[0013] Figure 3 (a) is a network structure diagram of ResNet;

[0014] Figure 3 (b) is a network structure diagram of BoTNet;

[0015] Figure 4 (a) is a structure diagram of a single-head self-attention mechanism;

[0016] Figure 4 (b) is an MHSA structure diagram;

[0017] Figure 5 (a) is a detection process diagram of a RepVGG structure;

[0018] Figure 5 (b) is an inference process diagram of a RepVGG structure;

[0019] Figure 6 FIG. 7 is a flowchart of merging a large number of convolution layers and BN in a RepVGG module;

[0020] Figure 7A flowchart of an insulator overheating defect detection method combining the RB-YOLO and thermal imaging technology;

[0021] Figure 8 A bar chart of the number of insulators in the heat and normal categories, a schematic diagram of the shape set and number set of the label box, a schematic diagram of the position set of the label box center, and a schematic diagram of the length set and width set of the label box;

[0022] Figure 9 (a) is a verification process diagram of the loss function;

[0023] Figure 9 (b) is a training process diagram of the loss function;

[0024] Figure 10 It is an F1 score diagram of the RB-YOLO model;

[0025] Figure 11 It is a comparison diagram of the iteration loss function in the original YOLOv8 and the present embodiment;

[0026] Figure 12 It is a comparison diagram of the iteration evaluation index of the original YOLOv8 and the present embodiment. DETAILED DESCRIPTION

[0027] To explain the technical content, the purpose and effect of the present application in detail, the following will be described in conjunction with the drawings.

[0028] Please refer to Figure 1 The present embodiment provides an insulator overheating detection method combining the RB-YOLO and thermal imaging technology, characterized in that it comprises the following steps:

[0029] An infrared thermal imaging insulator picture is collected and labeled to obtain an insulator dataset, and the insulator dataset is divided into a training set, a verification set and a test set;

[0030] The pictures in the training set and the verification set are input into the RB-YOLO network model for training, and the optimal weight data of the RB-YOLO network model is obtained through parameter optimization, wherein the RB-YOLO network model is a network model obtained by introducing a backbone network BoTNet and a structural reparameterization technology into a YOLOv8 network to design a backbone end;

[0031] The optimal weight data is loaded into the RB-YOLO network model, and the test set is identified and tested, and the detection result of the insulator overheating is output.

[0032] From the above description, the beneficial effects of the present application are that the insulator dataset is obtained by collecting and labeling infrared thermal imaging insulator pictures, the insulator dataset is divided into a training set, a validation set and a test set; the pictures in the training set and the validation set are input into the RB-YOLO network model for training, and the optimal weight data of the RB-YOLO network model is obtained through parameter optimization; the optimal weight data is loaded into the RB-YOLO network model, and the test set is identified and tested, and the detection result of the overheating of the insulator is output. Among them, the RB-YOLO network is improved based on the YOLOv8 network, the main network BoTNet is introduced, the convolution and the attention mechanism are combined, and the extraction ability of the target feature is improved; and the structure reparameterization technology is introduced, which greatly reduces the model parameter amount, so that a new main end is designed through the introduction of two module designs, while ensuring the model detection precision and generalization ability, the feature expression ability is enhanced, and the model memory and operation amount are greatly reduced. The RB-YOLO network model is combined with the infrared thermal imaging of the insulator in the ultrahigh voltage substation, and has smaller parameter amount and operation amount in the insulator heating defect detection, and further improves the precision and speed compared with the traditional detection method.

[0033] Further, the introduction of the main network BoTNet and the structure reparameterization technology in the YOLOv8 network to design the main end comprises:

[0034] The structure reparameterization technology and the main network BoTNet are adopted in the main end of the YOLOv8 network to perform multiple convolutions on the input pictures to obtain feature maps of different scales.

[0035] From the above description, since the structure reparameterization technology includes training and inference, the convolution branch and the residual branch of the identity mapping are added during training, and each branch is combined during inference, so that the single-path structure can significantly improve the inference speed; and the BoTNet structure combines CNN based on Transformer, uses self-attention mechanism and convolution to enhance the model performance, and the main part of the BoTNet structure is global multi-head self-attention, which can effectively capture and summarize target information, enhance the closeness between global features, and improve the feature information recognition ability.

[0036] Further, the collection and labeling of the infrared thermal imaging insulator pictures to obtain the insulator dataset, and the division of the insulator dataset into a training set, a validation set and a test set comprise:

[0037] Collecting the infrared thermal imaging insulator pictures, the infrared thermal imaging insulator pictures include insulator pictures with temperatures within a preset normal range and insulator pictures with temperatures outside the preset normal range;

[0038] The insulator picture with the temperature in the preset normal range is labeled as a normal insulator, the insulator picture with the temperature out of the preset normal range is labeled as an overheated insulator, and the labeled insulator picture is stored into an insulator dataset;

[0039] The insulator pictures in the insulator dataset are divided into a training set, a verification set and a test set according to a ratio of 6:2:2.

[0040] From the above description, it can be known that the insulator pictures of infrared thermal imaging are collected and labeled, the labeled insulator pictures are stored into an insulator dataset, and the insulator pictures in the dataset are divided into a training set, a verification set and a test set, which facilitates subsequent detection of overheated insulators in combination with an RB-YOLO network.

[0041] Further, the collection and labeling of the insulator pictures of infrared thermal imaging to obtain the insulator dataset further include:

[0042] The insulator pictures of infrared thermal imaging are collected and labeled, and the insulator shapes in the labeled insulator pictures are expanded through mixed enhancement or mosaic enhancement to obtain expanded pictures;

[0043] The expanded pictures are labeled, and the labeled insulator pictures and the labeled expanded pictures are stored into the insulator dataset.

[0044] From the above description, it can be known that the mixed enhancement, mosaic and other enhancement technologies meet the richness of the insulator shapes, and the expanded pictures with rich insulator shapes can be obtained, and the number of pictures in the insulator dataset is increased.

[0045] Further, the pictures in the training set and the verification set are input into an RB-YOLO network model for training, and optimal weight data of the RB-YOLO network model is obtained through parameter optimization, including:

[0046] The input pictures are preprocessed;

[0047] The preprocessed pictures are sequentially processed through a structure reparameterization module, a backbone network module and an SPPF module, during which a plurality of feature maps are obtained and input into a Neck end of the RB-YOLO network model;

[0048] In the Neck end of the RB-YOLO network model, the feature maps are strengthened and input into a Head end of the RB-YOLO network model;

[0049] The strengthened feature maps are predicted in the Head end of the RB-YOLO network model, and the position, label information and confidence of a prediction box are generated;

[0050] The CIoU and the DFL_Loss are used as loss functions, a vector angle alpha is added to the loss functions, and parameters of the RB-YOLO network model are optimized by the loss functions to obtain optimal weight data of the RB-YOLO network model.

[0051] It can be known from the above description that the RB-YOLO technology combined with the infrared thermal imaging of the insulator in the EHV substation can reduce the parameter quantity and the operation quantity when the insulator heating defect is detected.

[0052] Further, in the Neck end of the RB-YOLO network model, the feature map is strengthened, including:

[0053] In the Neck end of the RB-YOLO network model, the feature map is strengthened based on the Cf2 module and the CBS module.

[0054] Further, the strengthened feature map is predicted in the Head end of the RB-YOLO network model, including:

[0055] In the Head end of the RB-YOLO network model, the strengthened feature map is predicted, and the mean value of the bounding box prediction loss and the mean value of the classification prediction loss of the strengthened feature map are calculated.

[0056] Further, the input picture is preprocessed, including:

[0057] For the input picture, the initial anchor box length value is automatically set, the initial anchor box is trained, and the comparison with the real anchor box is performed, and the best anchor box value is obtained by iteratively updating the network parameters;

[0058] The size of the input picture is converted to a preset standard size.

[0059] It can be known from the above description that the input picture is unified by setting the adaptive best anchor box value and the adaptive size adjustment, which is convenient for subsequent picture training.

[0060] Further, the output of the insulator overheating detection result further includes:

[0061] An evaluation index of the insulator overheating detection performance is calculated:

[0062]

[0063]

[0064]

[0065]

[0066] In the formula, P represents precision, R represents recall, TP represents the number of targets that are normal samples and whose prediction results are normal, FP represents the number of targets that are overheated samples but whose prediction results are normal, FN represents the number of targets that are normal samples but whose prediction results are overheated, FPS represents the number of frames transmitted per second, N represents the number of samples to be measured, T represents the time required for testing all samples, and mAP represents the average precision mean value.

[0067] The precision, the recall, the number of frames transmitted per second, and the average precision mean value are combined as evaluation indexes to evaluate the overheating detection performance of the insulator.

[0068] As can be seen from the above description, the precision, the recall, the number of frames transmitted per second, and the average precision mean value are combined as evaluation indexes to evaluate the overheating detection performance of the insulator, which facilitates further adjustment of parameters of the network model in the future.

[0069] The above-mentioned insulator overheating detection method combining RB-YOLO and thermal imaging technology is suitable for improving the speed and precision of insulator overheating detection, and the following specific embodiments are described:

[0070] Embodiment one

[0071] Please refer to Figure 1 An insulator overheating detection method combining RB-YOLO and thermal imaging technology, comprising the following steps:

[0072] S1, collecting and labeling infrared thermal imaging insulator pictures to obtain an insulator dataset, and dividing the insulator dataset into a training set, a validation set, and a test set.

[0073] S11, collecting infrared thermal imaging insulator pictures, wherein the infrared thermal imaging insulator pictures include insulator pictures with temperatures within a preset normal range and insulator pictures with temperatures outside the preset normal range.

[0074] S12, labeling the insulator pictures with temperatures within the preset normal range as normal insulators, labeling the insulator pictures with temperatures outside the preset normal range as overheated insulators, and storing the labeled insulator pictures to the insulator dataset.

[0075] In some embodiments, the infrared thermal imaging insulator pictures are collected and labeled, the insulator shapes in the labeled insulator pictures are expanded through mixed enhancement or mosaic enhancement to obtain expanded pictures.

[0076] The expanded pictures are labeled, and the labeled insulator pictures and the labeled expanded pictures are stored to the insulator dataset.

[0077] S13, dividing the insulator pictures in the insulator dataset into the training set, the validation set and the test set according to a ratio of 6:2:2.

[0078] S2, inputting the pictures in the training set and the validation set into an RB-YOLO network model for training, and obtaining optimal weight data of the RB-YOLO network model through parameter optimization, the RB-YOLO network model being a network model obtained by introducing a backbone network BoTNet and a structural reparameterization technology into a YOLOv8 network.

[0079] In the backbone end of the YOLOv8 network, the structural reparameterization technology and the backbone network BoTNet are used to perform multiple convolutions on the input pictures to obtain feature maps of different scales.

[0080] Specifically, the YOLOv8 network is the latest iteration in the YOLO detection model series, which is famous for its joint detection and segmentation function. Please refer to Figure 2 The YOLOv8 uses a convolutional neural network and can be divided into two main parts: the backbone end and the head end.

[0081] The backbone end is mainly used for feature extraction. The YOLOv8 uses a C2f module to replace the original C3 model, so as to obtain more rich gradient flow information while ensuring lightness. At the end of the backbone, an SPPF module is used, which serially passes three Maxpools and then cascades each layer, so as to ensure the accuracy of objects of various scales while ensuring lightness.

[0082] The head end is mainly used for feature fusion. The head end of the YOLOv8 predicts the bounding box through an anchor-free mechanism. This mechanism does not predict the distance of the item from the known anchor box, but explicitly estimates the center of the object, which reduces the number of box predictions and thus speeds up the non-maximum suppression (NMS) speed. A decoupled structure is used to independently process the objectness, classification and regression tasks. This design allows each branch to focus on its task and improves the overall accuracy of the model. In the output layer of the YOLOv8, a sigmoid function is used as the activation function of the feature score, which represents the probability that the bounding box contains an object. A softmax function is used to calculate the class probability, which represents the probability that the object belongs to each possible class.

[0083] In terms of loss function, the YOLOv8 uses CIoU and DFL Loss functions for bounding box loss, and uses binary cross-entropy for classification loss. This way improves the object detection performance, especially when dealing with smaller objects.

[0084] In this embodiment, the backbone network BoTNet is introduced, please refer to Figure 3 , the BoTNet structure is based on the Transformer combined with CNN, and uses self-attention mechanism and convolution to enhance the model performance, please refer to Figure 4 , the main part of which is global multi-head self-attention (MHSA), which is used to replace the original 3x3 spatial convolution in ResNet, which can effectively capture and summarize target information, enhance the closeness between global features, and improve the feature information recognition ability.

[0085] Among them, the calculation formula of single-head self-attention mechanism is:

[0086]

[0087] Q, K, V: matrix after linear transformation of input matrix

[0088] QK T : Calculate the similarity of Q and K

[0089] Scaling factor

[0090] Softmax: normalization function

[0091] MHSA is to calculate the output weighted feature matrix from the four single-head self-attention mechanisms and combine them into a total matrix column by column, and finally output through a layer of full connection. Reduce the model parameter amount while improving the target recognition ability.

[0092] In this embodiment, the structure reparameterization technology RepVGG is introduced, please refer to Figure 5 , RepVGG is a single-line simple architecture, which has higher parallelism than multi-line. In the training process, RepVGG parallelizes three branches: a main branch with a convolution kernel size of 3x3, a shortcut branch with a convolution kernel size of 1x1, and a shortcut branch only connected to BN. Because of the residual structure with multiple branches, it is equivalent to adding multiple gradient flow paths in the network. Training such a network is similar to training multiple networks, integrating multiple networks into one network. Similar to the idea of model fusion, therefore, it occupies less memory and has higher flexibility, thereby improving the operation speed.

[0093] The main core of RepVGG is to use the structure reparameterization method to divide the model into two parts, training and inference. In training, a 1x1 convolution branch and an identity mapping are added to the residual branch, and in inference, the branches are combined to adopt a single-path structure, thereby significantly improving the inference speed.

[0094] Please refer toFigure 6 RepVGG reparameterization converts the multi-branch structure in the training process into a 3x3 convolution with bias, and a large number of convolution layers are linearly operated with BN in the RepVGG module to achieve the purpose of merging, which can reduce the model parameter amount, ensure the network robustness and generalization, and improve the forward inference speed of the next step, and finally further lighten the entire detection model.

[0095] The specific merging operation process is as follows:

[0096] Convolution layer: Conv(x) = w(x) + b, wherein w(x) is a weight function, and b is a bias parameter;

[0097] BN layer: wherein μ is the mean, σ 2 is the variance, γ is the scale factor, β is the translation factor, and ε is a constant.

[0098] After merging:

[0099] The new weight is The new bias parameter is

[0100] Therefore, BN(x) = W new +B new From the above formula, it can be seen that the BN layer can be merged into the convolution layer, and the new convolution expression reduces the parameter amount, thereby improving the model operation speed.

[0101] Therefore, in combination with the RB-YOLO and the infrared thermal imaging technology, the network for detecting the overheat defect of the insulator is as shown in Figure 7 .

[0102] S21, pre-processing the input picture.

[0103] Specifically, for the input picture, the length value of the initial anchor frame is automatically set, the initial anchor frame is trained and compared with the real anchor frame, and the best anchor frame value is obtained by iteratively updating the network parameters; the size of the input picture is converted to a preset standard size.

[0104] S22, sequentially processing the pre-processed picture through the structure reparameterization module, the backbone network module and the SPPF module, obtaining a plurality of feature maps during the processing and inputting the plurality of feature maps into the Neck end of the RB-YOLO network model.

[0105] S23, strengthening the feature maps in the Neck end of the RB-YOLO network model, and inputting the feature maps into the Head end of the RB-YOLO network model.

[0106] Specifically, in the Neck end of the RB-YOLO network model, the feature map is strengthened based on a Cf2 module and a CBS module.

[0107] S24, the strengthened feature map is predicted in the Head end of the RB-YOLO network model, and the position, label information and confidence of the prediction box are generated.

[0108] Specifically, in the Head end of the RB-YOLO network model, the strengthened feature map is predicted, and the mean value of the bounding box prediction loss and the mean value of the classification prediction loss of the strengthened feature map are calculated respectively.

[0109] S25, CIoU and DFL_Loss are used as loss functions, a vector angle α is added to the loss functions, and the optimal weight data of the RB-YOLO network model is obtained by optimizing the parameters of the RB-YOLO network model through the loss functions.

[0110] Wherein, the loss function is composed of angle cost Λ, distance cost Δ, shape cost Ω and IoU cost IoU, and a vector angle α is added, and the specific expression is as follows:

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] In the formula, x represents the sine function value of the vector angle α; c h represents the height of the minimum circumscribed rectangle of the prediction box and the real box; σ represents the diagonal line of the circumscribed rectangle of the coordinate difference calculation of the prediction box and the real box; represents the horizontal axis coordinate value of the real box; represents the horizontal axis coordinate value of the prediction box; represents the vertical axis coordinate value of the real box; represents the vertical axis coordinate value of the prediction box; γ represents the weight of the time and distance value in Δ; ρ xrepresents the square of the difference between the horizontal coordinate values of the prediction box and the real box and the square of the minimum circumscribed box ratio; p y represents the square of the difference between the vertical coordinate values of the prediction box and the real box and the square of the minimum circumscribed box ratio; c w represents the width of the minimum circumscribed rectangle of the prediction box and the real box; the value of represents the degree of attention to control gt represents the width of the real box; h gt represents the height of the real box; w represents the width of the prediction box; w w represents the ratio of the absolute value of the difference between the width of the prediction box and the real box to the larger value; w y represents the ratio of the absolute value of the difference between the width of the prediction box and the real box to the larger value; h represents the height of the prediction box; L siou represents the regression loss function.

[0120] S3, load the optimal weight data to the RB-YOLO network model, and perform identification test on the test set to output the detection result of the insulator overheating.

[0121] Specifically, the trained model is used to perform identification test on the insulator thermal imaging test set, and the final effect of the model is verified.

[0122] In some embodiments, in order to accurately evaluate the actual performance of the overheating detection of the present embodiment, it is also necessary to calculate the evaluation indexes of the insulator overheating detection performance:

[0123]

[0124]

[0125]

[0126]

[0127] In the formula, P represents the precision, R represents the recall, TP represents the number of targets that are normal samples and have normal prediction results, FP represents the number of targets that are overheating samples but have normal prediction results, FN represents the number of targets that are normal samples but have overheating prediction results, FPS represents the number of frames transmitted per second, N represents the number of samples to be tested, T represents the time required to test all samples, and mAP represents the average precision mean;

[0128] The precision, the recall, the number of frames transmitted per second, and the average precision mean are combined as evaluation indexes to evaluate the performance of the insulator overheating detection.

[0129] The running environment is as follows: the CPU is Intel(R) Xeon(R) Platinum 8255C CPU@2.50GHz, the GPU is NVIDIA GeForce RTX 2080 Ti, the memory is 11G, the operating system is ubuntu20.04, the Cuda11.3 version library file is installed, the development language is Python, and the Pytorch framework is used. The parameter configuration is as follows: the SGD optimization algorithm is used for parameter training, the input image size is 640*640 pixels, the Batch size is 16; the maximum iteration number is 300; the momentum factor is 0.937; and the weight decay coefficient is 0.0005. The cosine annealing strategy is used to dynamically adjust the learning rate, and the initial learning rate is 0.01.

[0130] Please refer to Figure 8 , and the label information of the training set is collected in Figure 8 . The number of hot and normal insulator thermal images, the shape set and number set of the label frame, the position set of the label frame center, and the length set and width set of the label frame are insulated. The number ratio of the insulator thermal image of the heat defect and the normal is 2:1. It can be found from Figure 8 that the target position is mostly concentrated in the middle of the whole thermal image, and the identified target is smaller than the picture size, and the sample size is mostly concentrated in (0-0.05, 0-0.05).

[0131] Figure 9 is the loss function of the RB-YOLO training process, the loss value gradually decays and tends to be stable, and approaches to 0 after 300 iterations, and it can be seen that the proximity of the model prediction and the true result is high.

[0132] At the same time, the detection method of the embodiment is compared with four kinds of mainstream target detection technologies of traditional SSD, YOLOv3, YOLOv5 and original YOLOv8, and the running environment and configuration are the same, and the evaluation indexes are shown in Table 1. As can be seen from the table, the operation speed and parameter amount of the embodiment are better than the other four methods, greatly reducing the model memory and operation amount, and at the same time, the overall accuracy and recall rate are also improved, which are improved by 3.5% and 4.7% compared with the original YOLOv8 target detection technology, and the superiority of the embodiment in accuracy and speed can be seen.

[0133] Table 1 Comparison of various target detection algorithms

[0134] Detection technique Picture size P R FPS FLOPs(G) mAP@0.5 mAP@0.5:0.95 SSD 640×640 70.6％ 69.5％ 76.4 282 70.9％ 40.4％ YOLOv3 640×640 73.6％ 70.5％ 82.3 18.9 70.9％ 45.4％ YOLOv5s 640×640 75.3％ 74.2％ 87.1 16.3 78.6％ 50.4％ YOLOv8n 640×640 78.0％ 80.1％ 101.1 8.1 84.9％ 55.3％ The present embodiment 640×640 81.5％ 84.8％ 93.8 2.6 86.6％ 58.6％

[0135] Among them, mAP@0.5 represents the average precision when IoU is set to 0.5, and mAP@0.5:0.95 represents the average precision when IoU is from 0.5 to 0.95 with a step of 0.05.

[0136] Since precision and recall are a pair of contradictory performance metrics. Generally speaking, the higher the precision will bring low recall. Therefore, the F1 score is introduced as the harmonic mean of precision and recall. Please refer to Figure 10 , the F1 curve is a comprehensive analysis of the relationship between precision, recall and confidence. The model of the present embodiment can get a better F1 score between the confidence of 0.2-0.7, and the F1 image is relatively flat, which shows that while taking into account the higher confidence, it can also maintain a good precision and recall, which proves that the present embodiment is generally better.

[0137] Please refer to Figure 11 and Figure 12 , compare the iteration results of the present embodiment and the original YOLOv8 training results. box_loss represents the average of the bounding box prediction loss, cls_loss represents the average of the classification prediction loss, and dfl_loss represents the average loss of the confidence calculation network. The smaller the loss of the above three parameters, the better the effect of target recognition. From Figure 11 it can be seen that in the training iteration process, the present embodiment has lower curves than the YOLOv8 model in these three aspects. This means that the present embodiment can bring lower loss in the same number of iterations, and has better accuracy. From Figure 12 it can be seen that the comparison of the results of precision, recall, mAP@0.5 and mAP@0.5-0.95 also reflects that the present embodiment is more advantageous than the original YOLOv8 model in detection accuracy.

[0138] In summary, the insulator overheat detection method combining RB-YOLO and thermal imaging technology provided by the application collects and labels infrared thermal imaging insulator pictures to obtain an insulator dataset, divides the insulator dataset into a training set, a verification set and a test set, inputs the pictures in the training set and the verification set into an RB-YOLO network model for training, obtains optimal weight data of the RB-YOLO network model through parameter adjustment and optimization, loads the optimal weight data to the RB-YOLO network model, and performs identification test on the test set to output a detection result of insulator overheat. The RB-YOLO network is improved based on a YOLOv8 network, a backbone network BoTNet is introduced into a feature extraction network, convolution is combined with an attention mechanism to improve the extraction capability of target features, and a structure reparameterization technology is introduced to greatly reduce the model parameter amount, a new backbone network is designed with RepVGG and BoTNet, the feature expression capability is enhanced while the model detection precision and generalization capability are ensured, and the model memory and operation amount are greatly reduced. The RB-YOLO network model is combined with infrared thermal imaging of insulators in an ultrahigh voltage substation, has smaller parameter amount and operation amount in insulator heating defect detection, and further improves the precision and speed compared with a traditional detection method.

[0139] The above description is only an embodiment of the application, and does not limit the patent scope of the application, and any equivalent transformation or direct or indirect application in the related technical field based on the content of the specification and drawings is also included in the patent protection scope of the application.

Claims

1. An insulator overheating detection method combining RB-YOLO and thermal imaging technology, characterized in that, The method comprises the steps of: collecting and labeling infrared thermal imaging insulator pictures to obtain an insulator dataset, and dividing the insulator dataset into a training set, a verification set and a test set; inputting pictures in the training set and the verification set into an RB-YOLO network model for training, and obtaining optimal weight data of the RB-YOLO network model through parameter optimization, wherein the RB-YOLO network model is a network model obtained by introducing a backbone network BoTNet and a structure reparameterization technology into a YOLOv8 network to design a backbone end; loading the optimal weight data into the RB-YOLO network model, and performing identification test on the test set to output a detection result of insulator overheating; the introduction of the backbone network BoTNet and the structure reparameterization technology into the YOLOv8 network to design the backbone end comprises: using the structure reparameterization technology and the backbone network BoTNet in the backbone end of the YOLOv8 network to perform multiple convolutions on input pictures to obtain feature maps of different scales; inputting pictures into the RB-YOLO network model for training, and obtaining optimal weight data of the RB-YOLO network model through parameter optimization, comprising: preprocessing the input pictures; processing the preprocessed pictures in sequence through a structure reparameterization module, a backbone network module and an SPPF module, during which multiple feature maps are obtained and transmitted to a Neck end of the RB-YOLO network model; in the Neck end of the RB-YOLO network model, the feature maps are strengthened and transmitted to a Head end of the RB-YOLO network model; in the Head end of the RB-YOLO network model, the strengthened feature maps are predicted, and the position, label information and confidence of the prediction box are generated; CIoU and DFL_Loss are used as loss functions, a vector angle α is added to the loss functions, and the parameters of the RB-YOLO network model are optimized through the loss functions to obtain the optimal weight data of the RB-YOLO network model; the preprocessing of the input pictures comprises: for the input pictures, the length and number of the initial anchor boxes are automatically set, the initial anchor boxes are trained and compared with real anchor boxes, and the best anchor box value is obtained through iterative updating of network parameters; the size of the input pictures is converted to a preset standard size.

2. The insulator overheating detection method combining RB-YOLO and thermal imaging technology according to claim 1, characterized in that, the collection and labeling of infrared thermal imaging insulator pictures to obtain an insulator dataset, and the division of the insulator dataset into a training set, a verification set and a test set comprise: collecting infrared thermal imaging insulator pictures, wherein the infrared thermal imaging insulator pictures include insulator pictures with temperatures within a preset normal range and insulator pictures with temperatures outside the preset normal range; labeling the insulator pictures with temperatures within the preset normal range as normal insulators, and labeling the insulator pictures with temperatures outside the preset normal range as overheated insulators, and storing the labeled insulator pictures in the insulator dataset; The insulator pictures in the insulator dataset are divided into a training set, a validation set and a test set according to a ratio of 6:2:

2.

3. The insulator overheating detection method combining RB-YOLO and thermal imaging technology according to claim 1, characterized in that, The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises:

4. The insulator overheating detection method combining RB-YOLO and thermal imaging technology according to claim 1, characterized in that, The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises:

5. The insulator overheating detection method combining RB-YOLO and thermal imaging technology according to claim 1, characterized in that, The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises:

6. The insulator overheat detection method combining RB-YOLO and thermal imaging technology according to claim 1, characterized in that, The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: ； ； ； ； The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal imaging insulator pictures further comprises: The insulator dataset obtained by collecting and labeling the infrared thermal

Images