A substation intelligent monitoring system and a monitoring method
By improving the C2f module and neck network structure of the YOLOv8 model, the detection capability of small target defects in substation equipment is enhanced, the problem of insufficient detection accuracy of existing models in complex backgrounds is solved, and efficient equipment condition monitoring is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHUZHOU SUBURBAN POWER SUPPLY CO OF STATE GRID ANHUI ELECTRIC POWER CO LTD
- Filing Date
- 2026-03-26
- Publication Date
- 2026-06-26
AI Technical Summary
Existing YOLO models lack sufficient accuracy in detecting small target defects in substation equipment inspection, and their ability to extract and utilize key features is limited in complex backgrounds, resulting in high rates of missed and false detections.
The YOLOv8 model is improved by replacing the C2f module with the C2f_DC module and inserting multiple shallow information perception branches into the neck network, including a multi-scale shallow feature fusion module and an information supplementation module. An attention mechanism is used to fuse shallow texture features and deep semantic features, and the Focal-CIoU loss function is used to optimize bounding box regression.
It significantly improves the detection accuracy and robustness of small target defects in substation equipment, reduces the rate of missed detections and false detections, realizes automatic and real-time monitoring of equipment status, and improves the reliability and real-time performance of monitoring.
Smart Images
Figure CN122290044A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of substation monitoring, specifically to an intelligent substation monitoring system and monitoring method. Background Technology
[0002] As a crucial hub connecting power generation, transmission, and distribution in a power system, substations undertake critical tasks such as voltage transformation, energy distribution, and system protection. Their operational status directly impacts the safety and stability of the entire power grid. A typical substation contains a wide variety of equipment, including transformers, circuit breakers, switchgear, voltage transformers, surge arresters, disconnect switches, and terminal boxes. These devices operate under high voltage, high load, and complex environments for extended periods. During routine maintenance, assessing the health of equipment and identifying potential risks involves checking for clear instrument displays, damaged meters, deformed casings, cracked insulators, and any hanging or suspended debris. Therefore, accurate and timely inspection of substation equipment and its defects is a fundamental requirement for ensuring the safe and stable operation of the power grid.
[0003] Patent CN116977321A discloses an intelligent method for identifying defects in substation equipment, utilizing the YOLO algorithm framework to identify and classify appearance defects in equipment images. This method includes an image input module, a backbone network module, a neck module, and a detection head module, and combines feature point extraction, feature point matching, and difference degree discrimination to achieve automatic identification of equipment condition defects. This method can automatically identify various appearance defects and abnormal conditions of substation equipment, providing technical support for equipment monitoring.
[0004] However, while existing technologies, such as the method disclosed in patent CN116977321A, have achieved automatic identification of equipment defects to a certain extent, they still have several limitations in practical applications. First, the appearance defects of substation equipment are diverse in form and scale, especially subtle texture changes (such as minor cracks in the dial or fine lines on the insulator) and small targets (such as small, suspended foreign objects). Under complex on-site background interference, the existing models lack the ability to extract shallow detail features, resulting in high false negative and false positive rates for these critical defects. Second, in the feature fusion stage of the YOLO algorithm framework used in this method, while extracting high-level semantic features, shallow features containing rich detail information (such as edges and textures) are easily diluted or lost when transferred to deeper layers, affecting the accuracy of detection.
[0005] Therefore, the existing YOLO series (such as YOLOv5 and YOLOv8) still lacks sufficient detection accuracy and robustness when dealing with complex substation scenarios, making it difficult to meet the requirements of condition monitoring for high accuracy and reliability. Summary of the Invention
[0006] The purpose of this invention is to address the problem mentioned in the background art that the existing methods have insufficient detection accuracy for small target defects in substation equipment, and that the model has limited ability to extract and utilize key features under complex background interference, resulting in high rates of missed detection and false detection. Therefore, this invention proposes an intelligent monitoring system and monitoring method for substations.
[0007] A first aspect of this invention provides a method for intelligent monitoring of substations, the method comprising: Acquire image data of the target substation; The image data is preprocessed to obtain the target image set; The target image is input into a pre-trained state detection model, which outputs the state detection result of the device in the target image; the target image is any image in the target image set. Based on the status detection results, an alarm message is generated and pushed to the monitoring platform; The state detection model is an improvement upon the YOLOv8 model, specifically including: Replace the C2f modules in the backbone and neck networks with the improved C2f_DC modules; Multiple pre-defined shallow information perception branches are inserted into the neck network. Each shallow information perception branch includes a multi-scale shallow feature fusion module and an information supplementation module. The multi-scale shallow feature fusion module is used to weightedly fuse shallow features of different scales output by the backbone network to obtain shallow texture features. The information supplementation module is used to fuse shallow texture features with deep semantic features at the corresponding scale of the neck network using an attention mechanism to supplement the detection head information.
[0008] Optionally, the improvements of the improved C2f_DC module compared to the original C2f module include: Replace the Bottleneck module in the C2f module with the improved Bottleneck_DC module; the operation process of the Bottleneck_DC module includes: ; Where X is the input of the Bottleneck_DC module; CBS represents the convolutional layer Conv, the batch normalization layer BatchNorm, and the activation layer SiLU; DConv represents the deformable convolutional layer; the kernel size of each convolutional layer is 3×3; shortcut is a Boolean parameter, which performs residual connections when it is true; Y is the output of the Bottleneck_DC module.
[0009] Optionally, the computation process of the multi-scale shallow feature fusion module includes: ; in, , , These are shallow features at different scales output by the backbone network; GAP represents global average pooling; concat represents the concatenation operation; MLP represents a multilayer perceptron. It is the channel weight vector; split indicates the splitting operation; , , They represent , , The corresponding weights; This indicates a convolutional layer with a kernel size of 1×1, used to standardize the number of channels; , , These are shallow features after weight calibration; resize indicates size adjustment, used to unify spatial resolution; This is the output of the multi-scale shallow feature fusion module.
[0010] Optionally, the calculation process of the information supplementation module includes: ; in, and These are shallow texture features and deep semantic features, respectively; concat means concatenation; This indicates a convolutional layer with a kernel size of 3×3, compressing the number of channels; flatten indicates a flattening operation, which flattens a three-dimensional tensor into a two-dimensional matrix; and In this context, the superscript T indicates transpose; D is the feature dimension of Q; Softmax is a normalization function; and recover represents the recovery operation, restoring a two-dimensional matrix to a three-dimensional tensor. This indicates a convolutional layer with a kernel size of 1×1, with the number of channels adjusted. This is the output of the information supplementation module.
[0011] Optionally, during model training, the Focal-CIoU loss function is used as the bounding box regression loss for the detector head; the formula for calculating the Focal-CIoU loss function is as follows: ; in, This is the Focal-CIoU loss we are looking for; It is the CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; It is a hyperparameter greater than 0.
[0012] A second aspect of this invention provides a substation intelligent monitoring system, the system comprising: The data acquisition module is used to acquire image data of the target substation; The preprocessing module is used to preprocess the image data to obtain the target image set; The intelligent recognition module is used to input the target image into a pre-trained state detection model and output the state detection result of the device in the target image; the target image is any image in the target image set; the state detection model is an improvement based on the YOLOv8 model, and the specific improvements include: Replace the C2f modules in the backbone and neck networks with the improved C2f_DC modules; Multiple pre-defined shallow information perception branches are inserted into the neck network. Each shallow information perception branch includes a multi-scale shallow feature fusion module and an information supplementation module. The multi-scale shallow feature fusion module is used to weightedly fuse shallow features of different scales output by the backbone network to obtain shallow texture features. The information supplementation module is used to fuse shallow texture features with deep semantic features at the corresponding scale of the neck network using an attention mechanism to supplement the detection head information; The alarm module is used to generate alarm information based on the status detection results and push the alarm information to the monitoring platform.
[0013] Optionally, the improvements of the improved C2f_DC module compared to the original C2f module include: Replace the Bottleneck module in the C2f module with the improved Bottleneck_DC module; the operation process of the Bottleneck_DC module includes: ; Where X is the input of the Bottleneck_DC module; CBS represents the convolutional layer Conv, the batch normalization layer BatchNorm, and the activation layer SiLU; DConv represents the deformable convolutional layer; the kernel size of each convolutional layer is 3×3; shortcut is a Boolean parameter, which performs residual connections when it is true; Y is the output of the Bottleneck_DC module.
[0014] Optionally, the computation process of the multi-scale shallow feature fusion module includes: ; in, , , These are shallow features at different scales output by the backbone network; GAP represents global average pooling; concat represents the concatenation operation; MLP represents a multilayer perceptron. It is the channel weight vector; split indicates the splitting operation; , , They represent , , The corresponding weights; This indicates a convolutional layer with a kernel size of 1×1, used to standardize the number of channels; , , These are shallow features after weight calibration; resize indicates size adjustment, used to unify spatial resolution; This is the output of the multi-scale shallow feature fusion module.
[0015] Optionally, the calculation process of the information supplementation module includes: ; in, and These are shallow texture features and deep semantic features, respectively; concat means concatenation; This indicates a convolutional layer with a kernel size of 3×3, compressing the number of channels; flatten indicates a flattening operation, which flattens a three-dimensional tensor into a two-dimensional matrix; and In this context, the superscript T indicates transpose; D is the feature dimension of Q; Softmax is a normalization function; and recover represents the recovery operation, restoring a two-dimensional matrix to a three-dimensional tensor. This indicates a convolutional layer with a kernel size of 1×1, with the number of channels adjusted. This is the output of the information supplementation module.
[0016] Optionally, during model training, the Focal-CIoU loss function is used as the bounding box regression loss for the detector head; the formula for calculating the Focal-CIoU loss function is as follows: ; in, This is the Focal-CIoU loss we are looking for; It is the CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; It is a hyperparameter greater than 0.
[0017] The beneficial effects of this invention are: 1. Based on the improved YOLOv8 condition detection model, the system can automatically identify and diagnose defects in substation monitoring images without relying on manual inspections or additional sensing equipment. The system enables continuous monitoring of large-scale equipment and can convert detection results into alarm information in real time and push it to the monitoring platform, significantly improving the real-time performance, coverage, and efficiency of monitoring, while significantly reducing manual intervention and inspection costs.
[0018] 2. By improving YOLOv8, shallow texture features are directly added to the detection head, providing key detail support for the detection of small targets (such as bird nests and suspended foreign objects) and subtle defects (such as dial scratches and insulator cracks), significantly reducing the rate of missed detections and false judgments, and improving the stability and reliability of detection results. Attached Figure Description
[0019] Figure 1 A flowchart of a substation intelligent monitoring method provided in an embodiment of the present invention; Figure 2 A diagram of the original network architecture of YOLOv8 provided in an embodiment of the present invention; Figure 3 This is a network architecture diagram of a state detection model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of C2f_DC provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a multi-scale shallow feature fusion module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an information supplementation module provided in an embodiment of the present invention; Figure 7 This is an architecture diagram of a substation intelligent monitoring system provided in an embodiment of the present invention. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.
[0021] This invention provides an intelligent monitoring method for substations. See also... Figure 1 , Figure 1 A flowchart illustrating a substation intelligent monitoring method provided in an embodiment of the present invention. The method includes the following steps: S101, acquire image data of the target substation.
[0022] S102, preprocess the image data to obtain the target image set.
[0023] S103: Input the target image into the pre-trained state detection model and output the state detection result of the device in the target image.
[0024] S104 generates alarm information based on the status detection results and pushes the alarm information to the monitoring platform.
[0025] The target image is any image in the target image set. The state detection model is an improvement on the YOLOv8 model, which includes the introduction of a C2f_DC module based on deformable convolution and the addition of a shallow information perception branch to supplement the detection head information.
[0026] The substation intelligent monitoring method provided by this invention improves the YOLOv8 model, effectively enhancing the detection capability of small target defects in substation equipment. It also improves robustness in complex backgrounds, realizes automatic and real-time monitoring of equipment status, significantly reduces the rate of missed detections and false detections, and thus improves the reliability of automated monitoring.
[0027] In one implementation, image acquisition can be performed using fixed camera equipment, mobile robots, or drones. Preprocessing includes filtering and noise reduction, image cropping, and normalization to ensure that the input size matches the model.
[0028] In one embodiment, see Figure 2 , Figure 2 This diagram illustrates the original network architecture of YOLOv8 as provided in this embodiment of the invention. The YOLOv8 model includes a backbone network, a neck network, and a head network. The backbone network is used for shallow feature extraction, the neck network is used for feature fusion and deep feature mining, and the head network is used for target recognition and classification, obtaining classification results, location, and confidence scores. The modules CBS, C2f, SPPF, Upsample, Concat, and Detect shown in the diagram are inherent modules of the original YOLO model and will not be elaborated upon here.
[0029] See Figure 3 , Figure 3 The network architecture diagram of a state detection model provided in this embodiment of the invention includes the following improvements compared to the original YOLOv8 model: 1. The C2f module is improved by using deformable convolution to obtain the C2f_DC module.
[0030] See Figure 4 , Figure 4 This is a schematic diagram of a C2f_DC structure provided in an embodiment of the present invention. Figure 4 As shown in (a), the C2f module includes the standard convolutional module CBS and n bottleneck structure modules. Figure 4 As shown in (b), each Bottleneck module consists of two CBS modules. Inserting a deformable convolutional layer DConv (preserving feature scale invariance) between the two CBS modules of the Bottleneck module yields an improved bottleneck module, denoted as Bottleneck_DC. See also... Figure 4 In section (c), the operation process of the Bottleneck_DC module includes: ; Where X is the input of the Bottleneck_DC module; CBS represents the convolutional layer Conv, the batch normalization layer BatchNorm, and the activation layer SiLU; DConv represents the deformable convolutional layer; the kernel size of each convolutional layer is 3×3; shortcut is a Boolean parameter, which performs residual connections when it is true; Y is the output of the Bottleneck_DC module.
[0031] Replace the Bottleneck module in C2f with the improved Bottleneck_DC module to obtain the C2f_DC module.
[0032] In one implementation, the C2f modules in the backbone network and neck network are replaced with improved C2f_DC modules.
[0033] Substation equipment defects (such as insulator cracks, circuit breaker mechanical misalignment, and bushing contamination layers) are mostly irregular in shape. Traditional standard convolution relies on a fixed sampling grid, which makes it difficult to fully capture deformation features. C2f-DC, through the dynamic sampling mechanism of DConv, can learn the offset of sampling points, allowing the convolution receptive field to adaptively fit the defect shape. This avoids the problem of insufficient feature extraction caused by defect deformation, improving the accuracy of equipment condition detection and its robustness in complex scenarios.
[0034] 2. A shallow information perception branch is proposed to fuse the shallow features output by the backbone network and the deep semantic features of the neck network. Specifically, a shallow information perception branch includes a multi-scale shallow feature fusion module (FFM) and an information supplementation module (ISM).
[0035] The multi-scale shallow feature fusion module is used to weightedly fuse shallow features of different scales output by the backbone network to obtain shallow texture features. See also Figure 5 , Figure 5 This is a schematic diagram of a multi-scale shallow feature fusion module provided in an embodiment of the present invention. In the figure, ( ), ( ), ( ) represents shallow features at different scales output by the backbone network; GAP represents global average pooling; concat represents concatenation operation; MLP represents multilayer perceptron (using two fully connected layers); split represents segmentation operation; Mul represents weighted multiplication; 1×1Conv represents a convolutional layer with a kernel size of 1×1, used to unify the number of channels; resize represents size adjustment, used to unify spatial resolution; Add represents element-wise addition.
[0036] The computational process of the multi-scale shallow feature fusion module includes: ; in, It is the channel weight vector ( ); , , They represent , , The corresponding weights; This indicates a convolutional layer with a kernel size of 1×1, used to standardize the number of channels; , , These are shallow features after weight calibration; for FFM_1, , , For FFM_2, , , For FFM_3, , , ; This is the output of the multi-scale shallow feature fusion module, and the output size corresponds to the fused deep features. To maintain consistency, for FFM_1, For FFM_2, For FFM_3, .
[0037] Resizing includes identity mapping, upsampling, and downsampling. Different branches have different adjustment methods. resolution resolution The resolution. FFM_1, for Perform identity mapping, for and Upsampling is performed to make its spatial resolution consistent with... Keep it consistent. FFM_2, for Perform identity mapping, for Perform downsampling, for Upsampling is performed to make its spatial resolution consistent with... Keep it consistent. FFM_3, Perform identity mapping, for , Perform downsampling to make its spatial resolution consistent with Maintain consistency.
[0038] The multi-scale shallow feature fusion module effectively preserves and enhances local texture and structural information by adaptively calibrating the weights of shallow features at different scales, thereby enhancing the expressive power of shallow features and making them more discriminative in capturing details and subtle changes.
[0039] The information supplementation module uses an attention mechanism to fuse shallow texture features with deep semantic features at the corresponding scale of the neck network, supplementing the detection head information. See also... Figure 6 , Figure 6 This is a schematic diagram of an information supplementation module provided in an embodiment of the present invention. In the figure, and These represent shallow texture features and deep semantic features, respectively; `concat` indicates channel concatenation; `3×3Conv` represents a convolutional layer with a kernel size of 3×3, used to compress the number of channels; `reshape` indicates dimensionality reshaping, including flattening operations from 3D to 2D and restoring operations from 2D to 3D; `Attention` indicates the calculation of attention weights, calculated using the formula: Mul represents weighted multiplication; Add represents element-wise addition; 1×1Conv represents a convolutional layer with a kernel size of 1×1, used to restore the number of channels.
[0040] The information supplementation module's operation process includes: ; in, This indicates a convolutional layer with a kernel size of 3×3, compressing the number of channels; `flatten` indicates a flattening operation, flattening a three-dimensional tensor into a two-dimensional matrix; Q, , , , It is the feature sequence generated during the operation (when the input is...) and All dimensions are hour Q, , , , ); and The superscript T in the text indicates transpose; D is the feature dimension of Q (e.g., ...). If D = C / 2), then Softmax is a normalization function; A is the attention weight; recover represents the recovery operation, restoring the two-dimensional matrix to a three-dimensional tensor; This indicates a convolutional layer with a kernel size of 1×1, with the number of channels adjusted (the number of channels is restored to C). This is the output of the information supplementation module, which serves as the input to the detection head. Its size is consistent with the size of the original model's detection head features.
[0041] The information supplementation module effectively fuses shallow texture features with deep semantic features through an attention mechanism, enabling the model to adaptively select and emphasize the most useful information for object detection. This supplements the key features required by the detection head, thereby improving the feature discrimination ability and detection accuracy. In one implementation, three shallow information perception branches are inserted into the neck network to supplement the input features of different detection heads and alleviate information dilution.
[0042] In existing YOLO-based detection models, as network depth increases, information loss easily occurs when features from the backbone network are transmitted to the detection head, leading to insufficient utilization of shallow texture details (such as minor damage to the device surface and edges of small foreign objects). This embodiment addresses this by directly extracting multi-scale shallow features from the backbone through its shallow information perception branch. , , After weight allocation and size unification, the data is fused, skipping the indirect transmission process between traditional FPN layers and reducing the loss of texture details caused by upsampling / downsampling. Shallow features are directly supplemented to the detection head, providing key detail support for the detection of small targets (such as bird nests and suspended foreign objects) and subtle defects (such as dial scratches and insulator cracks), thus reducing the false negative rate.
[0043] In one embodiment, during model training, the Focal-CIoU loss function is used instead of the original CIoU loss as the bounding box regression loss for the detector head. The formula for calculating the Focal-CIoU loss function is: ; in, This is the Focal-CIoU loss we are looking for; It is the CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; It is a hyperparameter greater than 0, with a value range of [0.5, 2]. It can be set to 1 first and then adjusted during training.
[0044] The Focal-CloU loss function assigns higher weights to targets that are difficult to predict or have little overlap in bounding box regression. By weighting the CloU loss, it enhances the model's sensitivity to small or difficult-to-detect targets while maintaining the regression accuracy for large targets, thereby improving the overall detection performance and robustness.
[0045] In one embodiment, YOLOv8n is used as the baseline model, retaining its original parameter settings, and a C2f_DC module and a shallow information perception branch are inserted. The parameter settings of the new module are shown in Table 1.
[0046] Table 1 In one implementation, the training process of the state detection model includes: Dataset Construction: Collect and label image data of substation equipment, including defect categories such as external equipment damage, equipment deformation, condensation, meter damage, insulator damage, oil leakage, breather damage, abnormal door closure, foreign objects, damaged or missing cover plates, failure to wear safety helmets, failure to wear work clothes, smoking, silicone discoloration, and pressure plate status. Preprocess the image data, such as normalization, resizing (input size 640×640), and enhancement (rotation, flipping, etc.), to generate training and validation sets.
[0047] Model training: Cross-entropy loss was used as the classification loss, and Focal-CloU loss was used as the bounding box regression loss. For training strategies, the batch size was set to 32, the initial learning rate was 0.001, and a warm-up + cosine decay learning rate strategy was adopted, with 5 warm-up rounds. The Adam optimizer was used for iterative updates. The experimental environment was an Ubuntu 22.04 system, equipped with an RTX 4090 graphics card and the PyTorch 2.1.0 framework.
[0048] In the substation scenario, compared to the original YOLOv8n model, the state detection model proposed in this invention improves mAP@0.5 from 92.45% to 95.84%. Although the model parameters and weights have increased, the inference time still meets the requirements of practical applications.
[0049] This invention provides an intelligent monitoring system for substations. See also... Figure 7 , Figure 7 This is an architecture diagram of a substation intelligent monitoring system provided in an embodiment of the present invention. The system includes: The data acquisition module is used to acquire image data of the target substation.
[0050] The preprocessing module is used to preprocess the image data to obtain the target image set.
[0051] The intelligent recognition module is used to input the target image into the pre-trained state detection model and output the state detection results of the device in the target image.
[0052] The alarm module is used to generate alarm information based on the status detection results and push the alarm information to the monitoring platform.
[0053] The substation intelligent monitoring system provided by this invention improves the YOLOv8 model, effectively enhancing the detection capability of small target defects in substation equipment. It also improves robustness in complex environments, realizes automatic and real-time monitoring of equipment status, significantly reduces the rate of missed detections and false detections, and thus improves the reliability of automated monitoring.
[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the claims of the present invention.
Claims
1. A method for intelligent monitoring of substations, characterized in that, The method includes: Acquire image data of the target substation; The image data is preprocessed to obtain the target image set; The target image is input into a pre-trained state detection model, which outputs the state detection result of the device in the target image; the target image is any image in the target image set. Based on the status detection results, an alarm message is generated and pushed to the monitoring platform; The state detection model is an improvement upon the YOLOv8 model, specifically including: Replace the C2f modules in the backbone and neck networks with the improved C2f_DC modules; Multiple pre-defined shallow information perception branches are inserted into the neck network. Each shallow information perception branch includes a multi-scale shallow feature fusion module and an information supplementation module. The multi-scale shallow feature fusion module is used to weightedly fuse shallow features of different scales output by the backbone network to obtain shallow texture features. The information supplementation module is used to fuse shallow texture features with deep semantic features at the corresponding scale of the neck network using an attention mechanism to supplement the detection head information.
2. The intelligent monitoring method for substations according to claim 1, characterized in that, Compared to the original C2f module, the improved C2f_DC module has the following specific improvements: Replace the Bottleneck module in the C2f module with the improved Bottleneck_DC module; the operation process of the Bottleneck_DC module includes: ； Where X is the input of the Bottleneck_DC module; CBS represents the convolutional layer Conv, the batch normalization layer BatchNorm, and the activation layer SiLU; DConv represents the deformable convolutional layer; the kernel size of each convolutional layer is 3×3; shortcut is a Boolean parameter, which performs residual connections when it is true; Y is the output of the Bottleneck_DC module.
3. The intelligent monitoring method for substations according to claim 1, characterized in that, The computation process of the multi-scale shallow feature fusion module includes: ； in, , , These are shallow features at different scales output by the backbone network; GAP represents global average pooling; concat represents the concatenation operation; MLP represents a multilayer perceptron. It is the channel weight vector; split indicates the splitting operation; , , They represent , , The corresponding weights; This indicates a convolutional layer with a kernel size of 1×1, used to standardize the number of channels; , , These are shallow features after weight calibration; resize indicates size adjustment, used to unify spatial resolution; This is the output of the multi-scale shallow feature fusion module.
4. The intelligent monitoring method for substations according to claim 3, characterized in that, The operation process of the information supplementation module includes: ； in, and These are shallow texture features and deep semantic features, respectively; concat means concatenation; This indicates a convolutional layer with a kernel size of 3×3, compressing the number of channels; flatten indicates a flattening operation, which flattens a three-dimensional tensor into a two-dimensional matrix; and In this context, the superscript T indicates transpose; D is the feature dimension of Q; Softmax is a normalization function; and recover represents the recovery operation, restoring a two-dimensional matrix to a three-dimensional tensor. This indicates a convolutional layer with a kernel size of 1×1, with the number of channels adjusted. This is the output of the information supplementation module.
5. The intelligent monitoring method for substations according to claim 1, characterized in that, During model training, the Focal-CIoU loss function is used as the bounding box regression loss for the detector head; the formula for calculating the Focal-CIoU loss function is as follows: ； in, This is the Focal-CIoU loss we are looking for; It is the CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; It is a hyperparameter greater than 0.
6. A substation intelligent monitoring system, characterized in that, The system includes: The data acquisition module is used to acquire image data of the target substation; The preprocessing module is used to preprocess the image data to obtain the target image set; The intelligent recognition module is used to input the target image into a pre-trained state detection model and output the state detection result of the device in the target image; the target image is any image in the target image set; the state detection model is an improvement based on the YOLOv8 model, and the specific improvements include: Replace the C2f modules in the backbone and neck networks with the improved C2f_DC modules; Multiple pre-defined shallow information perception branches are inserted into the neck network. Each shallow information perception branch includes a multi-scale shallow feature fusion module and an information supplementation module. The multi-scale shallow feature fusion module is used to weightedly fuse shallow features of different scales output by the backbone network to obtain shallow texture features. The information supplementation module is used to fuse shallow texture features with deep semantic features at the corresponding scale of the neck network using an attention mechanism to supplement the detection head information; The alarm module is used to generate alarm information based on the status detection results and push the alarm information to the monitoring platform.
7. The intelligent monitoring system for substations according to claim 6, characterized in that, Compared to the original C2f module, the improved C2f_DC module has the following specific improvements: Replace the Bottleneck module in the C2f module with the improved Bottleneck_DC module; the operation process of the Bottleneck_DC module includes: ； Where X is the input of the Bottleneck_DC module; CBS represents the convolutional layer Conv, the batch normalization layer BatchNorm, and the activation layer SiLU; DConv represents the deformable convolutional layer; the kernel size of each convolutional layer is 3×3; shortcut is a Boolean parameter, which performs residual connections when it is true; Y is the output of the Bottleneck_DC module.
8. The intelligent monitoring system for substations according to claim 6, characterized in that, The computation process of the multi-scale shallow feature fusion module includes: ； in, , , These are shallow features at different scales output by the backbone network; GAP represents global average pooling; concat represents the concatenation operation; MLP represents a multilayer perceptron. It is the channel weight vector; split indicates the splitting operation; , , They represent , , The corresponding weights; This indicates a convolutional layer with a kernel size of 1×1, used to standardize the number of channels; , , These are shallow features after weight calibration; resize indicates size adjustment, used to unify spatial resolution; This is the output of the multi-scale shallow feature fusion module.
9. A substation intelligent monitoring system according to claim 8, characterized in that, The operation process of the information supplementation module includes: ； in, and These are shallow texture features and deep semantic features, respectively; concat means concatenation; This indicates a convolutional layer with a kernel size of 3×3, compressing the number of channels; flatten indicates a flattening operation, which flattens a three-dimensional tensor into a two-dimensional matrix; and In this context, the superscript T indicates transpose; D is the feature dimension of Q; Softmax is a normalization function; and recover represents the recovery operation, restoring a two-dimensional matrix to a three-dimensional tensor. This indicates a convolutional layer with a kernel size of 1×1, with the number of channels adjusted. This is the output of the information supplementation module.
10. A substation intelligent monitoring system according to claim 6, characterized in that, During model training, the Focal-CIoU loss function is used as the bounding box regression loss for the detector head; the formula for calculating the Focal-CIoU loss function is as follows: ； in, This is the Focal-CIoU loss we are looking for; It is the CIoU loss; IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box; It is a hyperparameter greater than 0.