A substation defect detection method, device, computer equipment and storage medium

Abstract

The application discloses a power transformation station defect detection method and device, computer equipment and a storage medium, wherein the method comprises the following steps: constructing a data set according to a device defect image of a power transformation station and a corresponding label; training an improved RT-DETR model by using the data set to obtain a trained improved RT-DETR model; obtaining a device defect image of a target power transformation station; inputting the device defect image into the trained improved RT-DETR model to identify a defect position of the target power transformation station; wherein a main network of the trained improved RT-DETR model is an RMNet network, the RMNet network is trained by using a residual structure in a training stage to improve training precision; and in an inference stage, a structure reparameterization is used to equivalently remove the residual structure to improve inference speed. The application can realize real-time online monitoring of device defects of a power transformation station and guarantee safe, stable and reliable operation of a power system.

Description

A method, apparatus, computer equipment, and storage medium for detecting defects in substations. Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method, apparatus, computer equipment, and storage medium for detecting defects in substations. Background Technology

[0002] Power substations are crucial components of a power system, serving as places for voltage and current transformation, receiving electrical energy, and distributing it. Over long-term operation, substation equipment can develop defects such as rust, oil leaks, and breakage due to inherent characteristics, poor manufacturing processes, improper maintenance during inspection and testing, external forces, and natural disasters. To ensure the stable operation of substation equipment, it is essential to promptly detect these defects and potential problems to prevent future occurrences.

[0003] Traditional substation defect detection typically relies on manual inspections, which is inefficient, costly, and susceptible to the subjective experience of inspectors, leading to insufficient accuracy. With the increasing demand for digital and intelligent substations, intelligent power inspections utilizing artificial intelligence technology to achieve intelligent, automated, and efficient operations are gradually replacing traditional manual methods and becoming the mainstream approach for substation inspections.

[0004] RT-DETR is a target detection model based on the Transformer architecture. However, RT-DETR requires a lot of computing resources during the training phase and is relatively slow during the inference phase, making it difficult to achieve real-time detection and identification of defects in substation equipment, which affects the safe, stable and reliable operation of the power system. Summary of the Invention

[0005] The purpose of this invention is to provide a method, device, computer equipment, and storage medium for detecting defects in substations, in order to solve the technical problem in related technologies that makes it difficult to detect defects in substation equipment in real time, thus affecting the safe and stable operation of the power system.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] Option 1, a substation defect detection method, includes the following steps:

[0008] A dataset was constructed based on images of equipment defects in substations and their corresponding labels.

[0009] The improved RT-DETR model is trained using the dataset to obtain the trained improved RT-DETR model;

[0010] Acquire images of equipment defects in the target substation;

[0011] The defect image of the equipment is input into the trained improved RT-DETR model to identify the defect location of the target substation;

[0012] The backbone network of the trained improved RT-DETR model is the RMNet network. During the training phase, the RMNet network is trained using a residual structure to improve training accuracy. During the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed.

[0013] Optionally, the prediction head of the trained improved RT-DETR model is a DETR Head, and further includes:

[0014] The DETR Head is used to output several prediction boxes;

[0015] A preset number of prediction boxes are selected from the plurality of prediction boxes according to preset conditions.

[0016] Optionally, the step of using structural reparameterization to equivalently remove the residual structure to improve inference speed includes:

[0017] The residual structure is converted into channels of a convolutional layer using structural reparameterization, so that the residual structure can be removed equivalently.

[0018] Optionally, the step of converting the residual structure into channels of a convolutional layer using structural reparameterization includes:

[0019] Increase the output channels of the Depthwise convolutional layer;

[0020] Increase the input channels of the Pointwise convolutional layer;

[0021] The residual structure is converted into the output channel and the input channel using reparameterization.

[0022] Optionally, the neck network of the trained improved RT-DETR model is a hybrid encoder, and further includes:

[0023] Global features of the S5 feature layer are extracted using the Transformer Block in the hybrid encoder.

[0024] The global features are fused with the S3 and S4 feature layers using PAFPN to obtain fused features.

[0025] Optionally, after the hybrid encoder extracts the global features of the S5 feature layer using the Transformer Block, it further includes:

[0026] The self-attention of each pixel in the global feature is obtained using the Transformer Block.

[0027] Optionally, after fusing the global features with the S3 and S4 feature layers using PAFPN to obtain the fused features, the method further includes:

[0028] Simultaneous prediction boxes are generated using IoU-aware Query Selection; the category confidence of the prediction boxes is greater than a preset confidence threshold, and the intersection-union ratio is greater than a preset intersection-union ratio threshold.

[0029] Option 2, a substation defect detection device, comprising:

[0030] The dataset construction module is used to build datasets based on equipment defect images and corresponding labels in substations.

[0031] The model training module is used to train the improved RT-DETR model using the dataset to obtain the trained improved RT-DETR model.

[0032] Image acquisition module, used to acquire images of equipment defects in the target substation;

[0033] The defect detection module is used to input the defect image of the equipment into the trained improved RT-DETR model to identify the defect location of the target substation;

[0034] The backbone network of the trained improved RT-DETR model is the RMNet network. During the training phase, the RMNet network is trained using a residual structure to improve training accuracy. During the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed.

[0035] Option 3, a computer device, including: a processor and memory;

[0036] The memory stores a computer program, and the processor executes the computer program to implement the steps of the substation defect detection method.

[0037] Option 4: A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the substation defect detection method.

[0038] Based on the above technical solution, the beneficial effects of this invention are:

[0039] An improved RT-DETR model was trained using a substation equipment defect dataset. The model's backbone network employs an RMNet network, incorporating reparameterization techniques. Residual structures are used during training to improve accuracy; in the inference phase, these structures are equivalently removed, converting to a single-path architecture to increase inference speed, achieving a dual improvement in both training accuracy and inference speed. A lightweight neck network (Hybrid Encoder) and a DETR Head are used, enhancing the identification rate of common substation defects and improving real-time detection. This avoids the excessively long non-maximum suppression time issue found in YOLO series detection methods, enabling real-time online monitoring of substation equipment defects and ensuring the safe, stable, and reliable operation of the power system. Attached Figure Description

[0040] Figure 1 is a flowchart illustrating an embodiment of the substation defect detection method of the present invention;

[0041] Figure 2 is a schematic diagram of the substation defect detection process in an embodiment of the substation defect detection method of the present invention.

[0042] Figure 3 is a schematic diagram of the backbone network training stage in an embodiment of the substation defect detection method of the present invention.

[0043] Figure 4 is a schematic diagram of the backbone network inference stage in an embodiment of the substation defect detection method of the present invention;

[0044] Figure 5 is a schematic diagram of the residual structure removal process of the backbone network in an embodiment of the substation defect detection method of the present invention;

[0045] Figure 6 is a structural schematic diagram of an embodiment of the substation defect detection device of the present invention. Detailed Implementation

[0046] Terminology Explanation:

[0047] RT-DETR (Real-Time Detection Transformer) is a real-time end-to-end detector based on the DETR architecture, achieving state-of-the-art (SOTA) performance in both speed and accuracy. Structurally, RT-DETR can be divided into three parts: the backbone network, the neck network, and the head network.

[0048] RT-DETR is a real-time object detection model that combines two classic object detection methods: Transformer and DETR (Detection Transformer). Transformer is a neural network architecture used for sequence modeling, originally developed for natural language processing but proven highly effective in computer vision. DETR is an end-to-end object detection model that transforms the object detection task into an object query problem, which is then solved using Transformer. RT-DETR adopts the structure of DETR but incorporates several optimizations to achieve real-time object detection.

[0049] PAFPN (Path Aggregation Network with Feature Pyramid Network): PAFPN is characterized by introducing a cascaded attention mechanism that fuses feature maps from different levels, thereby improving object detection accuracy. Simultaneously, PAFPN employs a Feature Pyramid Network (FPN) to extract features at different scales, thus better adapting to scale variations of various targets.

[0050] Structural re-parameterization refers to first constructing a series of structures (generally used for training) and then equivalently transforming their parameters into another set of parameters (generally used for inference or deployment), thus converting this series of structures into another series of structures. In real-world scenarios, training resources are generally relatively abundant, and the focus is usually on the overhead and performance during inference. Therefore, the desired training structure is larger and possesses a certain desirable property (higher accuracy or other useful properties, such as sparsity), while the resulting inference structure is smaller and retains this property (the same accuracy or other useful properties). In this way, the training structure corresponds to one set of parameters, and the desired inference structure corresponds to another set of parameters; as long as the parameters of the former can be equivalently transformed into the latter, the structure of the former can also be equivalently transformed into the latter. Structure A corresponds to a set of parameters X, and structure B corresponds to a set of parameters Y. If we can equivalently transform X into Y, then structure A can be equivalently transformed into B.

[0051] This invention provides a method, apparatus, computer equipment, and storage medium for detecting defects in substations, thereby addressing the technical problem in related technologies where real-time detection of defects in substation equipment is difficult, affecting the safe and stable operation of the power system.

[0052] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0053] Unless otherwise defined, 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. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0054] In recent years, the rapid development of deep learning technology has provided new solutions for power equipment defect detection. Among them, RT-DETR is a target detection model based on the Transformer architecture, which has excellent performance, but it still faces some challenges in handling substation defect detection. Specifically, RT-DETR requires a large amount of computational resources during the training phase, while the inference phase is relatively slow, making real-time detection a challenge.

[0055] To enable real-time detection and identification of substation defects, the substation defect detection method provided in this invention utilizes an improved RT-DETR model for real-time intelligent detection. The improved RT-DETR model employs structural reparameterization technology to replace the original backbone network with an RMNet network and uses a residual structure during the training phase, thereby improving training accuracy. During the inference phase, structural reparameterization technology is used to equivalently remove the residual branch structure, further improving inference speed. Simultaneously, the use of a lightweight hybrid encoder and DETR Head improves the identification rate of common substation defects and enhances the real-time performance of the detection.

[0056] Please refer to Figure 1. This invention provides an embodiment of a substation defect detection method, including the following steps:

[0057] S100: Construct a dataset based on equipment defect images and corresponding labels in substations;

[0058] S200: Use the dataset to train the improved RT-DETR model to obtain the trained improved RT-DETR model;

[0059] S300: Acquire images of equipment defects in the target substation;

[0060] S400: Input the equipment defect image into the trained improved RT-DETR model to identify the defect location of the target substation;

[0061] The backbone network of the trained improved RT-DETR model is the RMNet network. During the training phase, the RMNet network is trained using a residual structure to improve training accuracy. During the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed.

[0062] This invention utilizes images of equipment defects in substations to construct a dataset, and employs an improved RT-DETR model to perform real-time online detection of defects in substation equipment. The improved RT-DETR model uses an RMNet network as its backbone network, a Hybrid Encoder as its neck network, and a DETR Head as its detection head.

[0063] In this embodiment of the invention, a sufficient number of substation defect images are collected, and the defective parts are labeled to obtain defect labeling data, thus establishing a substation equipment defect image dataset (hereinafter referred to as the dataset). Then, an improved RT-DETR model is used to train the images in the dataset to obtain a trained improved RT-DETR model; the trained improved RT-DETR model is then used to identify equipment defect images of the target substation and detect the corresponding defect categories.

[0064] The improved RT-DETR model employs the RMNet network as its backbone. During training, the RMNet network uses a residual structure for training, which improves training accuracy. During inference, the RMNet network removes the residual structure through structural reparameterization, improving inference speed. The neck network of the improved RT-DETR model uses a HybridEncoder. First, a Transformer is used to extract global features from the S5 feature layer, fully learning the high-dimensional features in the network to obtain global features. Then, a path aggregation feature pyramid network (PAFPN) is used to fuse the global features with the S3 and S4 feature layers to obtain fused features, which are then output.

[0065] This invention generates more high-quality encoder features through IoU-aware Query Selection; the prediction head uses a DETR Head, which can output a series of preset number of prediction boxes; then, the top N prediction boxes are retained according to their scores, avoiding the problem of excessively long NMS time (non-maximum suppression time) in YOLO series detection; finally, the improved DT-DETR algorithm is deployed to the substation intelligent terminal, which can realize real-time online monitoring of substation defects.

[0066] In step S100, a dataset is constructed based on the equipment defect images of the substation and their corresponding labels.

[0067] In one embodiment, a certain number of substation equipment defect images are acquired, and the defective parts are labeled to obtain defect labeling data (defect labels), thus establishing a dataset of substation equipment defect images. The images in the dataset can originate from actual operating scenarios of substation power equipment, and may include, for example, common targets such as: abnormal meter readings, damaged meter casings, bird nests, abnormal door closures, pressure plate malfunctions, damaged cover plates, suspended objects, silicone discoloration, blurred meter dials, broken insulators, damaged meter dials, oil stains on the ground, not wearing safety helmets, not wearing work clothes, abnormal oil seal levels, damaged silicone cylinders, and smoking; the image data is divided into training, testing, and validation sets in an 8:1:1 ratio.

[0068] In step S200, the improved RT-DETR model is trained using the dataset to obtain the trained improved RT-DETR model.

[0069] In this embodiment of the invention, the improved RT-DETR model uses the RMNet network as its backbone network and introduces reparameterization technology into the backbone network. During the training phase, a residual structure is used for training to improve accuracy; during the inference phase, it is converted to a single-path architecture to improve inference speed. A lightweight Hybrid Encoder is combined for global feature extraction and multi-scale feature fusion. The detection head directly outputs a series of predicted boxes, and then retains the top N predicted boxes based on their scores. Unlike the YOLO Head, it does not need to predict the category and coordinates of each grid cell, nor does it require the non-maximum suppression step commonly used in traditional object detection to eliminate redundant detection results.

[0070] In one embodiment, an improved RT-DETR is used to train the images in the dataset. The backbone network of this model is an RMNet network, and a residual structure is used during the training phase to improve training accuracy.

[0071] Specifically, during the training phase, the RMNet network is a MobileNetV2 with residual structure. After the improved RT-DETR model is trained, the residual structure (residual branches) is converted into output or input channels in Depth-wise convolution or Point-wise convolution through the structural reparameterization method. The residual structure is equivalently removed by increasing the number of channels. The trained improved RT-DETR model is then converted into a MobileNetV1 structure without residual structure during the inference phase, thus possessing both the training accuracy of MobileNetV2 and the inference speed of MobileNetV1.

[0072] In step S300, an image of equipment defects in the target substation is acquired.

[0073] The target substation refers to the substation on which equipment defects will be identified. By acquiring images of equipment defects in the target substation, the specific location of defects in the equipment can be identified based on the images. Based on the location of the defects, it can be determined which type or categories of substation equipment have defects, thus enabling real-time online monitoring of substation equipment defects.

[0074] Since the installation location of each piece of equipment in a substation is fixed, once the specific defect location of the substation equipment is identified, that defect location is the defective part of the substation equipment.

[0075] In step S400, the equipment defect image is input into the trained improved RT-DETR model to identify the defect location of the target substation.

[0076] First, the captured images of substation equipment defects are resized and scaled to a size of 640×640 pixels. The 640×640 images are then input into a trained, improved RT-DETR model for defect detection, which identifies the location of the defects in the images.

[0077] Specifically, the input substation equipment defect image is processed by the backbone network RMNet of the improved RT-DETR model, which is trained. The RMNet network is used to extract features from the equipment defect image. The defect image is downsampled by 3x, 4x and 5x respectively to obtain the corresponding feature maps, namely feature maps of three sizes: S3 (80×80), S4 (40×40) and S5 (20×20), which correspond to 3x, 4x and 5x downsampling respectively.

[0078] The trained improved RT-DETR model uses a hybrid encoder in its neck network. First, it uses a Transformer Block to extract global features from the S5 feature layer, fully learning the high-dimensional features in the network. Then, it fuses the features with the S3 and S4 feature layers to obtain fused features.

[0079] In one embodiment, the neck network of the trained improved RT-DETR model is a hybrid encoder, and also includes:

[0080] Global features of the S5 feature layer are extracted using the Transformer Block in the hybrid encoder.

[0081] The global features are fused with the S3 and S4 feature layers using PAFPN to obtain fused features.

[0082] Specifically, a Hybrid Encoder can be used to perform multi-scale feature fusion on the S3, S4, and S5 feature maps. Since the S5 feature map contains high-dimensional semantic information, further feature extraction is performed on it by flattening the width and height of S5, and then global feature extraction is achieved through the self-attention mechanism of the Transformer Block, as shown in Equation (1):

[0083] Q = K = V = Flatten(S5); (1)

[0084] In the formula: Q, K, and V represent the query (query matrix), key (key matrix), and value (value matrix) in the Transformer Block, respectively. The three are calculated through self-attention and reshaped to obtain the F5 feature map, as shown in Equation (2):

[0085]

[0086] In the formula, Reshape means reshaping the flattened tensor into a feature map tensor, Atten means self-attention computation, T means the transpose of the tensor, and d means the dimension of each column of features in the key matrix.

[0087] Then, F5 is fused with S3 and S4 through PAFPN to obtain the output of the Hybrid Encoder, as shown in equation (3):

[0088] output=PAFPN(S3,S4,F5);(3)

[0089] Then, the output of the Hybrid Encoder is used to generate a prediction box (Query) with both high class confidence and high intersection-over-union (IoU) through IoU-aware Query Selection. The loss function formula is shown in Equation (4):

[0090]

[0091] In the formula, L(·) represents the total loss function, L box (·) represents the bounding box loss, L cls (·) represents the category loss, and IoU represents the intersection-union ratio of the predicted bounding box and the label bounding box. y represents the prediction and the label, respectively. Let represent the predicted class and coordinates, and c and b represent the true values ​​of the class and coordinates, respectively. By incorporating IoU into the class loss, the discrepancy between classification and regression can be aligned.

[0092] In the trained improved RT-DETR model, the prediction head can be a DETR Head. The DETR Head introduces the denoising idea of ​​DINO to improve the sample quality of bilateral matching, accelerate training convergence, and finally output a series of prediction boxes (Queries). The final detection result is obtained by selecting the top K of the scores.

[0093] In one embodiment, the backbone network of the trained improved RT-DETR model is an RMNet network. During the training phase, the RMNet network is trained using a residual structure to improve training accuracy. During the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed.

[0094] In this embodiment of the invention, the backbone network RMNet is shown in Figures 3 to 5. Figure 3 shows the structure of a normal ResNet. During the model inference stage, the residual structure of the ResNet can be removed by the RM (remove) operation, and it becomes a cylindrical structure (a cylindrical structure without residuals) as shown in Figure 4.

[0095] Because x + This is the output of the previous ResNet module after passing through the ReLU function, and all its values ​​are non-negative; first, increase the number of output channels of the convolutional layer Conv1 by an amount equal to x. + The number of channels is increased, and the convolution kernel is initialized with dirac. Therefore, the input of Conv1 after increasing the number of channels is the same as the original output. Add x to the base + ; due to x + Non-negative, its value remains unchanged after passing through the ReLU1 activation function. for The feature map after the ReLU1 activation function. Then, the number of input channels of the Conv2 convolutional layer is increased by x. + The number of channels is increased, and the convolution kernel is initialized with dirac. Therefore, the output of Conv2 after increasing the number of channels is changed from the original output. Become The output is the same as the original residual structure. That is, by increasing the number of output and input channels of Conv1 and Conv2, the residual structure can be equivalently removed, and the feature map y in the training phase is equal to the feature map z in the inference phase. Both are activated by the ReLU2 activation function y. + The final output and z + They are also equal; this is the principle of RM operation in RMNet.

[0096] Similarly, the RM principle can be applied to MobileNetV2 with residual structure. During the inference phase, the residual branches can be converted into output or input channels of Depth-wise convolution and Point-wise convolution through structural reparameterization technology, which can then be converted into MobileNetV1 structure. The RMNet network in this invention adopts MobileNetV2 with residual structure during the training phase and is converted into MobileNetV1 without residual structure during the inference phase.

[0097] In this embodiment of the invention, a normal ResNet model is used during the training phase. During the inference phase, additional channels are added to the convolutional layers, matching the number of channels in the input feature map. By initializing the expanded channels, all input channels are retained without changing their values ​​after the feature map is expanded through the convolutional layers. Since the input feature map values ​​are non-negative, they are also not changed after ReLU. The feature map operations on the residual part are merged with the retained feature map, and the final value is equivalent to the result before transformation.

[0098] It should be noted that structural reparameterization is the simplification of model structure by using parameter equivalence transformation. In this patent, it can be understood as equivalent removal of residual branches. In Figure 5, this is represented by converting residual branches into output or input channels in depth-wise convolution and point-wise convolution, thereby removing the residual branches of MobileNetV2 and converting it into the MobileNetV1 structure.

[0099] The improved RT-DETR model (object detection model) in this embodiment of the invention replaces the backbone network with an RMNet network. RMNet is a network that utilizes reparameterization technology. By introducing reparameterization technology into the backbone network, it achieves a dual improvement in training accuracy and inference speed.

[0100] This invention enables real-time intelligent detection of substation equipment defects based on an improved RT-DETR (Real-time Detection Transformer). First, a sufficient number of substation defect images are collected and the defect locations are labeled to establish a substation equipment defect image dataset. Then, the improved RT-DETR model is used to train and recognize the images. The improved RT-DETR model uses RMNet (reserving and merging network) for its backbone network. During the training phase, residual structures are used to improve training accuracy, and during the inference phase, structural reparameterization is used to remove residual branches equivalently, thereby improving inference speed. The neck network uses a hybrid encoder, which first uses Transformer Block to extract global features from the S5 feature layer, fully learning the high-dimensional features in the network, and then fuses them with the S3 and S4 feature layers. The prediction head uses DETR Head, which directly outputs a series of prediction boxes. Then, the top N prediction boxes are retained according to their scores, avoiding the problem of excessively long non-maximum suppression (NMS) time in YOLO (you only look once) series detection. Finally, the improved RT-DETR algorithm is deployed to the substation intelligent terminal to realize real-time online monitoring of substation defects.

[0101] The substation defect detection method provided in this invention utilizes a substation equipment defect dataset to train an improved RT-DETR model. The backbone network of this model employs an RMNet network, incorporating reparameterization techniques. During the training phase, a residual structure is used for training, improving training accuracy. In the inference phase, the residual structure is equivalently removed, converting to a single-path architecture, which improves inference speed, achieving a dual improvement in both training accuracy and inference speed. The use of a lightweight neck network (Hybrid Encoder) and a DETR Head enhances the identification rate of common substation defects, improves real-time detection, and avoids the excessively long non-maximum suppression time issue in YOLO series detection methods. This enables real-time online monitoring of substation equipment defects, ensuring the safe, stable, and reliable operation of the power system.

[0102] Please refer to Figure 6. This invention provides an embodiment of a substation defect detection device, comprising:

[0103] Dataset construction module 11 is used to construct a dataset based on equipment defect images and corresponding labels in a substation.

[0104] Model training module 22 is used to train the improved RT-DETR model using the dataset to obtain the trained improved RT-DETR model;

[0105] Image acquisition module 33 is used to acquire images of equipment defects in the target substation;

[0106] The defect detection module 44 is used to input the defect image of the equipment into the trained improved RT-DETR model to identify the defect location of the target substation.

[0107] The backbone network of the trained improved RT-DETR model is the RMNet network. During the training phase, the RMNet network is trained using a residual structure to improve training accuracy. During the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed.

[0108] The substation defect detection device provided in this invention utilizes a substation equipment defect dataset to train an improved RT-DETR model. The model's backbone network employs an RMNet network, incorporating reparameterization techniques. During training, a residual structure is used to improve training accuracy. In the inference phase, the residual structure is equivalently removed, converting to a single-path architecture, which improves inference speed, achieving a dual improvement in both training accuracy and inference speed. The use of a lightweight neck network (Hybrid Encoder) and a DETR Head enhances the identification rate of common substation defects, improves real-time detection, and avoids the excessively long non-maximum suppression time issue found in YOLO series detection methods. This enables real-time online monitoring of substation equipment defects, ensuring the safe, stable, and reliable operation of the power system.

[0109] In addition, the present invention also provides a computer device, including: a processor and a memory;

[0110] The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0111] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0113] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0116] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting defects in a substation, characterized in that, Includes the following steps: A dataset is constructed based on equipment defect images and corresponding labels of substations; an improved RT-DETR model is trained using the dataset to obtain a trained improved RT-DETR model; equipment defect images of the target substation are acquired; the equipment defect images are input into the trained improved RT-DETR model to identify the defect locations of the target substation; wherein, the backbone network of the trained improved RT-DETR model is an RMNet network, and during the training phase, the RMNet network is trained using a residual structure to improve training accuracy; during the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed; if the neck network of the trained improved RT-DETR model is a hybrid encoder, the model further includes: extracting data using the Transformer Block in the hybrid encoder. Global features of the feature layer; using PAFPN to combine the global features with 、 The feature layer performs feature fusion to obtain fused features.

2. The substation defect detection method according to claim 1, characterized in that, The prediction head of the trained improved RT-DETR model is called the DETR Head, and it also includes: outputting a number of prediction boxes using the DETR Head; and selecting a preset number of prediction boxes from the number of prediction boxes according to preset conditions.

3. The substation defect detection method according to claim 1, characterized in that, The method of using structural reparameterization to equivalently remove the residual structure to improve inference speed includes: using structural reparameterization to convert the residual structure into channels of a convolutional layer, so as to equivalently remove the residual structure.

4. The substation defect detection method according to claim 3, characterized in that, The method of converting the residual structure into channels of a convolutional layer using structural reparameterization includes: adding output channels of a depthwise convolutional layer; adding input channels of a pointwise convolutional layer; and converting the residual structure into the output channels and the input channels using reparameterization.

5. The substation defect detection method according to claim 4, characterized in that, The extraction using the Transformer Block in the hybrid encoder Following the global features of the feature layer, the process also includes: using the Transformer Block to obtain the self-attention of each pixel in the global features.

6. The substation defect detection method according to claim 4, characterized in that, The global features are obtained by utilizing PAFPN. 、 After the feature layer performs feature fusion to obtain fused features, it also includes: generating simultaneous prediction boxes using IoU-aware Query Selection; the category confidence of the prediction boxes is greater than a preset confidence threshold, and the intersection-union ratio is greater than a preset intersection-union ratio threshold.

7. A substation defect detection device, characterized in that, include: The dataset construction module is used to build datasets based on equipment defect images and corresponding labels in substations. The model training module is used to train an improved RT-DETR model using the dataset to obtain a trained improved RT-DETR model; the image acquisition module is used to acquire equipment defect images of the target substation. The defect detection module is used to input the defect image of the equipment into the trained improved RT-DETR model to identify the defect location of the target substation; wherein, the backbone network of the trained improved RT-DETR model is an RMNet network, and during the training phase, the RMNet network is trained using a residual structure to improve training accuracy; During the inference phase, structural reparameterization is used to equivalently remove the residual structure to improve inference speed; if the neck network of the trained improved RT-DETR model is a hybrid encoder, the device is further used to: extract using the Transformer Block in the hybrid encoder. Global features of the feature layer; using PAFPN to combine the global features with 、 The feature layer performs feature fusion to obtain fused features.

8. A computer device, characterized in that, include: A processor and a memory; wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the substation defect detection method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the substation defect detection method according to any one of claims 1 to 6.

Images