A deep learning-based detection method and device for open / close state of an isolating switch

Abstract

The application provides a kind of based on deep learning's disconnecting switch opening and closing state detection method and device, including obtaining disconnecting switch image;Utilize the disconnecting switch detection network that is trained in advance to disconnecting switch image detection, obtain the detection result of disconnecting switch opening and closing state;Wherein, disconnecting switch detection network utilizes first feature extraction and processing module and second feature extraction and processing module and carries out feature extraction;First feature extraction and processing module are used to adopt multi-branch convolution structure to the disconnecting switch image of input preliminary feature extraction, to obtain preliminary feature in the basis of maintaining image spatial resolution;Second feature extraction and processing module are used to adopt depth separable convolution layer and dynamic convolution layer to preliminary feature depth feature mining, to obtain the key feature of disconnecting switch image.The detection network of the application can improve detection precision by multi-branch convolution and depth feature mining, and meet real-time detection demand.

Description

Technical Field

[0001] This invention belongs to the field of power system equipment condition detection technology, specifically relating to a method and device for detecting the opening and closing status of disconnecting switches based on deep learning. Background Technology

[0002] In power systems, the stable operation of various devices is crucial for ensuring the reliability of power supply. Disconnect switches, as one of the key pieces of equipment, play a vital role in achieving electrical isolation in power lines, providing safety assurance for equipment maintenance and fault handling. Their operational status directly affects the overall stability and security of the power system.

[0003] With the continuous expansion of power system scale and the increasing demand for intelligence, higher requirements are placed on the accuracy, efficiency, and real-time performance of disconnector switch opening and closing status detection. Traditional methods relying on manual inspection or simple sensor monitoring are gradually becoming insufficient to meet these needs. Visual inspection methods based on sensors and image processing have emerged, and in recent years, detection methods based on deep learning have become a hot topic in research and application.

[0004] Current sensor-based detection methods suffer from high installation and maintenance costs and are easily affected by environmental factors, leading to inaccurate results. While image processing-based visual detection methods reduce costs, they have stringent requirements for image quality. Traditional image processing techniques struggle to capture subtle state changes in disconnect switches, resulting in limited detection accuracy. Deep learning-based detection methods, in particular, generally rely on single-scale convolutional operations, failing to simultaneously capture both local details and global structural features of the disconnect switch. Furthermore, they lack key feature enhancement mechanisms, resulting in poor performance in complex scenarios. Moreover, their complex network structures and high computational costs make them unsuitable for real-time detection requirements. Summary of the Invention

[0005] In view of this, the present invention provides a method and apparatus for detecting the open / closed state of a disconnector switch based on deep learning, aiming to solve the problems of low detection accuracy, poor environmental adaptability, high computational complexity and poor real-time performance in at least one traditional method for detecting the open / closed state of a disconnector switch.

[0006] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for detecting the open / closed state of an isolating switch based on deep learning, comprising the following steps:

[0008] Acquire images of disconnect switches;

[0009] A pre-trained disconnector detection network is used to detect disconnector images to obtain the detection results of the disconnector's open and closed state.

[0010] The disconnector detection network utilizes a first feature extraction and processing module and a second feature extraction and processing module for feature extraction. The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnector image using a multi-branch convolutional structure to obtain preliminary features while maintaining the spatial resolution of the image. The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers to obtain the key features of the disconnector image.

[0011] Furthermore, the disconnector detection network includes:

[0012] The system consists of a backbone feature extraction network, a multi-scale feature processing module, an attention enhancement module, and a state detection module.

[0013] The backbone feature extraction network includes a first feature extraction and processing module and a second feature extraction and processing module;

[0014] The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnect switch image. Through different convolutional layers in the multi-branch convolutional structure, features are extracted from different angles, and then the results of each branch are stitched together to obtain a preliminary feature map while maintaining the spatial resolution of the image.

[0015] The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers, thereby obtaining the key feature map of the disconnector image.

[0016] The multi-scale feature processing module is used to extract and fuse key features at multiple scales to obtain multi-scale feature maps.

[0017] The attention enhancement module is used to calculate and adjust the importance weights in multi-scale feature maps based on the EA attention mechanism to obtain feature maps that highlight key features;

[0018] The status detection module is used to determine the final open / closed status detection result of the disconnector switch based on the feature map that highlights key features.

[0019] Furthermore, the first feature extraction and processing module is the MSCA-Conv module, and the image processing procedure of the MSCA-Conv module includes:

[0020] By performing convolution, dilated convolution, and depthwise separable convolution operations on the input isolating switch feature map using a set of Conv modules, the second, third, and fourth isolating switch feature maps are obtained.

[0021] The feature maps of the second, third, and fourth disconnectors are standardized using a BN layer.

[0022] pass The layer performs a stitching operation on the standardized disconnector feature map to obtain the fifth disconnector feature map;

[0023] The feature map of the sixth disconnector is obtained by performing a convolution operation on the feature map of the fifth disconnector through a Conv module;

[0024] The feature map of the sixth disconnector is processed by the global average pooling module to obtain the feature map of the seventh disconnector.

[0025] Attention weights are generated on the feature map of the seventh disconnector through a fully connected layer, and then normalized using the Sigmoid function to obtain the feature map of the eighth disconnector.

[0026] The eighth disconnector feature map is multiplied by the input disconnector feature map to obtain the ninth disconnector feature map, which is then used as the final output of the MSCA-Conv module.

[0027] Furthermore, the transformation formulas for the feature diagrams of the fifth and ninth disconnecting switches are as follows:

[0028]

[0029]

[0030] In the formula, This represents the feature diagram of the fifth disconnector switch. This indicates a splicing operation. Indicates standardized processing; , and These represent convolution operations with a kernel size of 1 and a stride of 1, dilated convolution operations with a kernel size of 3 and a stride of 1, and depthwise separable convolution operations with a kernel size of 5 and a stride of 1, respectively. The feature map representing the input isolating switch; This represents the feature diagram of the ninth disconnector.

[0031] Furthermore, the second feature extraction and processing module is the DASC-Conv module. The image processing procedure of the DASC-Conv module includes:

[0032] The first Conv module performs a convolution operation on the input isolating switch feature map to obtain the tenth isolating switch feature map;

[0033] The eleventh disconnector feature map is obtained by performing a convolution operation on the tenth disconnector feature map using the second Conv module.

[0034] The feature map of the eleventh disconnector is processed by the max pooling module to obtain the feature map of the twelfth disconnector.

[0035] The feature map of the twelfth disconnector is processed by bilinear interpolation module to obtain the feature map of the thirteenth disconnector.

[0036] The Swish function module activates the thirteenth disconnector feature map and multiplies it with the tenth disconnector feature map to obtain the fourteenth disconnector feature map, which is then used as the final output of the DASC-Conv module.

[0037] Furthermore, the transformation formulas for the thirteenth disconnector feature diagram and the fourteenth disconnector feature diagram are as follows:

[0038]

[0039]

[0040] In the formula, This represents the characteristic diagram of the thirteenth disconnector. This represents the bilinear interpolation operation. This indicates a max pooling operation. This represents a convolution operation with a kernel size of 3 and a stride of 1. This represents a convolution operation with a kernel size of 1 and a stride of 1. The feature map representing the input isolating switch; express Activation function This represents the characteristic diagram of the tenth disconnector. This represents the characteristic diagram of the fourteenth disconnector switch.

[0041] Furthermore, the multi-scale feature processing module is the MSFE-Conv module, and the image processing procedure of the MSFE-Conv module includes:

[0042] The input isolating switch feature maps are convolved using two Conv modules of different sizes to obtain the fifteenth isolating switch feature map and the sixteenth isolating switch feature map.

[0043] The feature images of the fifteenth and sixteenth disconnect switches are spliced ​​together using the channel splicing module to obtain the feature image of the seventeenth disconnect switch.

[0044] The feature map of the seventeenth disconnector is obtained by performing global average pooling on the feature map of the seventeenth disconnector through the global average pooling module.

[0045] The 18th disconnector feature map is activated by the GELU activation function to obtain the 19th disconnector feature map;

[0046] The feature map of the twentieth disconnector is obtained by performing a convolution operation on the feature map of the nineteenth disconnector through a Conv module.

[0047] The 20th disconnector feature map and the input disconnector feature map are multiplied by a dot product to obtain the 21st disconnector feature map, which is then used as the final output of the MSFE-Conv module.

[0048] Furthermore, the transformation formulas for the seventeenth disconnector feature diagram, the twentieth disconnector feature diagram, and the twenty-first disconnector feature diagram are as follows:

[0049]

[0050]

[0051]

[0052] In the formula, This diagram shows the characteristics of the seventeenth disconnector switch. This indicates a channel splicing operation. This represents a convolution operation with a kernel size of 1 and a stride of 1. This represents a convolution operation with a kernel size of 3 and a stride of 1. The feature map representing the input isolating switch; This represents the feature diagram of the twentieth disconnector. express Activation function This indicates a global average pooling operation. This represents the characteristic diagram of the 21st disconnector switch.

[0053] Furthermore, the state detection module is called the Detect module. Let the feature map highlighting key features be the twenty-second disconnector feature map. Based on the feature map highlighting key features, the final disconnector open / closed state detection result is determined, including:

[0054] The 22nd disconnector feature map is generated by averaging the feature values ​​of each channel along the spatial dimension using the global average pooling module.

[0055] The feature map of the 23rd disconnector switch is passed through a fully connected layer to map the channel dimension to the number of target categories, and the feature vector is output.

[0056] The eigenvectors are normalized to a probability distribution using the Softmax function. ,in ,in, This represents the probability that the disconnector switch is in the open state. This represents the probability that the disconnector switch is in the closed state.

[0057] Secondly, the present invention provides a deep learning-based device for detecting the open / closed state of an isolation switch, comprising:

[0058] The image acquisition module is used to acquire images of the disconnecting switch;

[0059] The detection module is used to detect disconnector images using a pre-trained disconnector detection network to obtain the detection results of the disconnector's open and closed state.

[0060] The disconnector detection network utilizes a first feature extraction and processing module and a second feature extraction and processing module for feature extraction. The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnector image using a multi-branch convolutional structure to obtain preliminary features while maintaining the spatial resolution of the image. The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers to obtain the key features of the disconnector image.

[0061] In summary, this invention provides a method and apparatus for detecting the open / closed state of a disconnector switch based on deep learning. The method includes acquiring an image of the disconnector switch; detecting the disconnector switch image using a pre-trained disconnector switch detection network to obtain the detection result of the disconnector switch's open / closed state; wherein the disconnector switch detection network uses a first feature extraction and processing module and a second feature extraction and processing module to extract features; the first feature extraction and processing module uses a multi-branch convolutional structure to perform preliminary feature extraction on the input disconnector switch image to obtain preliminary features while maintaining image spatial resolution; the second feature extraction and processing module uses depthwise separable convolutional layers and dynamic convolutional layers to perform deep feature mining on the preliminary features, thereby obtaining the key features of the disconnector switch image. This invention, based on image processing, avoids the high installation and maintenance costs and environmental interference problems of traditional sensor detection. Compared with traditional visual detection, its detection network, through multi-branch convolution and deep feature mining, can better capture subtle state changes, improve detection accuracy, and take into account both local and global features, enhance key features, simplify the network structure, reduce computational overhead, and meet real-time detection requirements. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart of a deep learning-based method for detecting the open / closed state of a disconnector switch, provided as an embodiment of the present invention;

[0064] Figure 2 A technical roadmap for a deep learning-based method for detecting the open / closed state of a disconnector switch, provided for embodiments of the present invention;

[0065] Figure 3 This is a diagram of the MDMENet network structure provided in an embodiment of the present invention;

[0066] Figure 4 This is a framework diagram of the MSCA-Conv module provided in an embodiment of the present invention;

[0067] Figure 5 This is a structural diagram of the DASC-Conv module provided in an embodiment of the present invention;

[0068] Figure 6 This is a structural diagram of the MSFE-Conv module provided in an embodiment of the present invention;

[0069] Figure 7 This is a structural diagram of the Detect module provided in an embodiment of the present invention;

[0070] Figure 8 A block diagram of a deep learning-based disconnector switch opening / closing state detection device provided in an embodiment of the present invention;

[0071] Figure 9 This is a block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0072] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0073] The following is a brief introduction to some of the technical terms involved in this invention:

[0074] (1) Disconnecting switch: A type of switch mainly used for "isolating power supply, switching operation, and connecting and disconnecting small current circuits". It plays an important role in the power system, and the accurate detection of its opening and closing status is crucial for the safe and stable operation of the power system.

[0075] (2) Deep Learning (DL): is a branch of machine learning. It is based on artificial neural networks to build models with multiple layers and automatically extracts features and patterns from data through learning from a large amount of data.

[0076] (3) Neural Network (NN): A computational model composed of a large number of neurons that simulates the workings of the human nervous system. It learns the features and patterns of data through the connections and weight adjustments between neurons. In deep learning, neural networks typically have multiple layers, including input layers, hidden layers, and output layers.

[0077] (4) Convolutional Neural Network (CNN): A type of neural network specifically designed to process data with a grid structure (such as images). It uses components such as convolutional layers and pooling layers to automatically extract features from images.

[0078] Please see Figure 1 This invention provides a method for detecting the open / closed state of a disconnector switch based on deep learning, comprising the following steps:

[0079] S11: Obtain the image of the disconnect switch.

[0080] It is understandable that images of the disconnect switch can be acquired using certain image acquisition devices (such as high-definition cameras, drones, etc.) and used as input data for subsequent testing.

[0081] S12: Use a pre-trained disconnector detection network to detect disconnector images and obtain the detection results of the disconnector's open / closed state;

[0082] The disconnector detection network utilizes a first feature extraction and processing module and a second feature extraction and processing module for feature extraction. The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnector image using a multi-branch convolutional structure to obtain preliminary features while maintaining the spatial resolution of the image. The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers to obtain the key features of the disconnector image.

[0083] It should be noted that the first feature extraction and processing module employs a multi-branch convolutional structure to perform preliminary feature extraction on the input isolating switch image from different angles, obtaining preliminary features while maintaining image spatial resolution. The multi-branch convolutional structure means that multiple different convolutional kernels or convolutional methods are used to process the image in parallel, capturing richer feature information from the image. The second feature extraction and processing module uses depthwise separable convolutional layers and dynamic convolutional layers to perform deep feature mining on the preliminary features obtained from the first module, thereby obtaining the key features of the isolating switch image. Depthwise separable convolutional layers reduce computational cost, while dynamic convolutional layers can adaptively adjust the convolutional kernels according to the input, enhancing the ability to extract key features.

[0084] The detection method provided in this embodiment is based on the fundamental principles of Convolutional Neural Networks (CNNs) in deep learning, extracting image features through convolution operations. The multi-branch convolutional structure of the first feature extraction and processing module uses different convolutional kernels to convolve the image from multiple scales and directions, obtaining preliminary feature representations. These features contain local details and certain global structural information of the image. The depthwise separable convolutional layer and dynamic convolutional layer of the second feature extraction and processing module further process the preliminary features. The depthwise separable convolutional layer reduces computational complexity, while the dynamic convolutional layer dynamically adjusts the convolutional kernel weights according to the image features, thereby more accurately extracting key features. These key features are crucial for determining the open / closed state of the disconnect switch. The features processed by the two feature extraction and processing modules are input into the subsequent detection network structure (although not detailed in this embodiment, it generally includes structures such as a classifier). Through the analysis and processing of the key features, the detection result of the open / closed state of the disconnect switch is output.

[0085] This method differs from traditional deep learning-based detection methods that commonly rely on single-scale convolution operations. It employs a multi-branch convolutional structure to extract image features simultaneously from multiple angles and scales, taking into account both local details and global structural features of the isolator, thus improving the comprehensiveness and accuracy of feature extraction. Furthermore, in the second feature extraction and processing module, depthwise separable convolutional layers reduce computational cost, making the model more efficient in image processing. Meanwhile, dynamic convolutional layers adaptively adjust kernel weights based on input features, enhancing the model's adaptability to different image features and its ability to extract key features.

[0086] In one embodiment, the disconnector detection network includes:

[0087] The system consists of a backbone feature extraction network, a multi-scale feature processing module, an attention enhancement module, and a state detection module.

[0088] The backbone feature extraction network includes a first feature extraction and processing module and a second feature extraction and processing module;

[0089] The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnect switch image. Through different convolutional layers in the multi-branch convolutional structure, features are extracted from different angles, and then the results of each branch are stitched together to obtain a preliminary feature map while maintaining the spatial resolution of the image.

[0090] The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers, thereby obtaining the key feature map of the disconnector image.

[0091] The multi-scale feature processing module is used to extract and fuse key features at multiple scales to obtain multi-scale feature maps.

[0092] The attention enhancement module is used to calculate and adjust the importance weights in multi-scale feature maps based on the EA attention mechanism to obtain feature maps that highlight key features;

[0093] The status detection module is used to determine the final open / closed status detection result of the disconnector switch based on the feature map that highlights key features.

[0094] In this embodiment, the disconnector detection network consists of a backbone feature extraction network, a multi-scale feature processing module, an attention enhancement module, and a state detection module. The backbone feature extraction network, as the foundation, includes a first feature extraction and processing module and a second feature extraction and processing module. The first feature extraction and processing module uses a multi-branch convolutional structure to perform preliminary feature extraction on the input disconnector image from multiple angles through different convolutional layers, and then concatenates the results of each branch to obtain a preliminary feature map while maintaining image spatial resolution, providing basic feature information for subsequent processing. The second feature extraction and processing module then uses depthwise separable convolutional layers and dynamic convolutional layers to perform deep feature mining on the preliminary feature map, further refining the key feature maps of the disconnector image, making the features more critical and representative. The multi-scale feature processing module performs multi-scale feature extraction and fusion operations on the key feature maps to obtain multi-scale feature maps, enhancing feature diversity and the ability to capture information at different scales. The attention enhancement module, based on the EA attention mechanism, calculates and adjusts the importance weights of each part in the multi-scale feature map, highlighting key features and suppressing irrelevant information to obtain a feature map that emphasizes key features, allowing the model to focus more on important features. Finally, the state detection module, based on the feature map highlighting key features, analyzes and processes the data to determine the final open / closed state detection result of the disconnector switch, achieving accurate judgment of the disconnector switch's open / closed state by the entire detection network.

[0095] In specific implementation, this invention proposes a multi-scale dynamic multi-branch enhanced network (MDMENet) based on the aforementioned embodiments. Based on this network, a method for detecting the open / closed state of a disconnector switch is designed, including: first, constructing a disconnector switch image dataset; then, preprocessing the image dataset and dividing it into training, validation, and test sets according to a certain ratio; designing an MDMENet network and initializing its training parameters; then, loading the training and validation sets into the MDMENet network for training and updating the network parameters; then, inputting the test set into the trained network to verify its performance; and finally, applying the trained MDMENet network to the disconnector switch open / closed state detection task. The technical route of this method is as follows: Figure 2 As shown, the following describes it in conjunction with some embodiments.

[0096] S1: Construct an image dataset of disconnecting switches.

[0097] High-definition cameras and drones were used in conjunction in the substation to capture static and dynamic images of disconnect switches, with images periodically extracted. To enhance the universality and robustness of the dataset for algorithm training, images of disconnect switches under different weather conditions (sunny, cloudy, rainy, foggy), different light intensities (strong light, soft light, backlight), and different opening and closing states (open, closed) were carefully collected. These images were then filtered and classified to construct a comprehensive and detailed disconnect switch image dataset.

[0098] In one specific implementation, the time interval for capturing images from the video is determined based on the current monitoring task of the disconnector switch status. When the disconnector switch needs to be opened or closed, the interval is usually about 1 second; otherwise, the interval is usually about 30 seconds.

[0099] S2: Perform data preprocessing and preprocess the dataset into partitions.

[0100] The disconnector switch image data underwent preprocessing and dataset partitioning. Gaussian filtering was applied to the acquired disconnector switch images to smooth them, reducing high-frequency noise while preserving edge information, resulting in clearer images of single-column vertical telescopic disconnectors. These processed images constitute a high-quality single-column vertical telescopic disconnector switch image dataset, providing a solid foundation for subsequent image analysis and condition detection. Then, annotation tools were used to label the icing areas. Finally, the overhead power line icing image dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0101] It's important to note that Gaussian filtering is a linear smoothing filtering technique widely used in image processing. It uses the Gaussian function from mathematics as its core, achieving noise reduction and smoothing effects through convolution operations on the image. The core idea of ​​Gaussian filtering is to utilize the properties of the Gaussian function to assign a weight to each pixel in the image. This weight decreases as the pixel's distance from the center point increases, thus removing noise while preserving as much edge and detail information as possible.

[0102] S3: Design the MDMENet network.

[0103] In one embodiment of the present invention, a multi-scale dynamic enhancement fusion network MDMENet is designed for detecting the open / closed state of disconnect switches, and its structure is as follows: Figure 3 As shown in the diagram, in MDMENet, the input image of the disconnector switch is first fed into the Multi-Scale Context Aggregation Convolution (MSCA-Conv) module for feature fusion. Then, the Dynamic Adaptive Spatial Convolution (DASC-Conv) module progressively extracts key features from the disconnector switch image and enhances the model's detection capability through non-linear activation. Simultaneously, the image is fed into the Multi-Scale Feature Extraction & Fusion Convolution (MSFE-Conv) module for multi-scale feature extraction and fusion. An Explanation and Analysis (EA) attention mechanism is then introduced, and finally, the Detect module outputs the detection result of the disconnector switch's open / closed state. The network structure is as follows: Figure 3 As shown.

[0104] Step 3.1: Perform feature fusion on the input disconnect switch image.

[0105] In a further embodiment, the first feature extraction and processing module is designed as an MSCA-Conv module, the structure of which is as follows: Figure 4 As shown, any disconnector image in the disconnector image dataset is denoted as a disconnector feature map. The input is then fed into the MSCA-Conv module for feature fusion.

[0106] In the MSCA-Conv module, the input isolating switch feature map is first processed. Perform a convolution (Conv) operation with a kernel size of 1 and a stride of 1, and then input it into a Batch Normalization (BN) layer for standardization to obtain the output isolating switch feature map. The main purpose of this step is to extract the feature map of the disconnector switch. Local features in the image. This convolution operation enhances the representational power of features while maintaining spatial resolution; at the same time, it also... Perform dilated convolution with a kernel size of 3 and a stride of 1 (dilation rate = 2), and input it into a BN layer for normalization to obtain the isolating switch feature map. The main purpose of this step is to Perform more extensive contextual feature extraction; in addition, A depthwise separable convolution operation with a kernel size of 5 and a stride of 1 is performed, and the result is fed into a BN layer for normalization to obtain the isolating switch feature map. The main purpose of this step is to extract a wider range of local features while reducing the number of parameters.

[0107] After that , , By splicing the channels, the characteristic map of the disconnector switch is obtained. The main purpose of this step is to fuse multi-scale features and enhance feature diversity; then... Perform a convolution operation with a kernel size of 1 and a stride of 1, adjusting the number of channels to match the input. Similarly, the characteristic map of the disconnecting switch is obtained. The main purpose of this step is to integrate multi-scale features while maintaining spatial resolution; then... Perform global average pooling to obtain the feature map. The main purpose of this step is to compress the spatial dimension, retain global information, and reduce computational complexity.

[0108] Then to Attention weights are generated through a fully connected layer and normalized using the Sigmoid function to obtain the isolation switch feature map. The main purpose of this step is to enhance important features and suppress irrelevant features; finally, and Perform a dot product operation to obtain the characteristic map of the disconnector switch. The main purpose of this step is to further enhance the model's sensitivity to key features through the attention mechanism, thereby improving the feature representation capability.

[0109] The MSCA-Conv module enhances the diversity and expressive power of features while maintaining the spatial resolution of the input isolating switch feature map through multi-scale feature extraction and feature fusion, allowing the model to focus more on important features, thereby improving detection accuracy and robustness.

[0110] In a further embodiment, and The transformation formula is as follows:

[0111] (1)

[0112] (2)

[0113] In the formula, It is a convolution with a kernel size of 1 and a stride of 1. It is a convolution with a kernel size of 3 and a stride of 1. It is a convolution with a kernel size of 5 and a stride of 1. It is an average pooling operation. It's an activation function; FC stands for fully connected layer. It is element-wise multiplication.

[0114] It should be noted that the sigmoid function is a commonly used activation function, typically used for non-linear transformations in neural networks. The output value of the sigmoid function is between 0 and 1. When the input x approaches positive infinity, the output is close to 1; when the input approaches negative infinity, the output is close to 0. Its mathematical expression is: .

[0115] Average pooling is a pooling operation commonly used in convolutional neural networks (CNNs). It is mainly used to downsample the input feature map, reduce the size of the feature map, thereby reducing the amount of computation, alleviating overfitting, and extracting more representative features.

[0116] Fully connected layers (FC layers) are a common layer type in deep learning models, widely used in neural networks, especially in convolutional neural networks (CNNs) and fully connected neural networks (MLPs). Their main function is to map input features to the output space, and they are typically used for classification, regression, or other tasks.

[0117] Step 3.2: Extract key features from the feature map of the disconnecting switch output by MSCA-Conv.

[0118] In a further embodiment of the present invention, the second feature extraction and processing module is designed as a DASC-Conv module, the structure of which is as follows: Figure 5 As shown.

[0119] In the DASC-Conv module, firstly... Performing a depthwise separable convolution operation with a kernel size of 1 and a stride of 1 yields the feature map of the isolating switch. The main purpose of this step is to extract local detail features from the feature map of the disconnector switch while reducing the number of parameters; then... Perform a dynamic convolution operation with a kernel size of 3 and a stride of 1 to obtain the feature map of the disconnector switch. The main purpose of this step is to capture the local structural features of the disconnector and enhance the expressive power of the features by dynamically adjusting the weights of the convolution kernel; then... Max pooling is performed using MaxPool to obtain the feature map of the disconnector switch. The main purpose of this step is to compress the spatial dimension while preserving global information. Next, [the following steps are taken]. The feature map of the disconnector switch is obtained by upsampling through bilinear interpolation and activation using the Swish function. The main purpose of this step is to compress the value range of the disconnector switch feature map to between [0, 1], further enhancing the model's sensitivity to the disconnector switch state; finally, and Perform a dot product operation to obtain the characteristic map of the output disconnect switch. The main purpose of this step is to fuse local features with global features through dot product operations, thereby enhancing the model's ability to judge the state of disconnect switches.

[0120] The DASC-Conv module extracts key features from disconnector switch images step by step through operations such as depthwise separable convolution, dynamic convolution, and max pooling, and enhances the model's detection capability through nonlinear activation. This module has a clear structure, is computationally efficient, and can effectively improve the accuracy and robustness of disconnector switch status detection.

[0121] The execution flow of the DASC-Conv module is as follows: assuming The dimension is (Where H is height, W is width, and C is the number of channels), first... Perform a convolution operation with a kernel size of 1, a stride of 1, and a number of convolutions of C, to obtain a dimension of... Disconnecting switch feature diagram Then on Perform a convolution operation with a kernel size of 3, a stride of 1, and a number of convolutions of C, to obtain a dimension of... Disconnecting switch feature diagram ; then Perform max pooling to obtain dimension Disconnecting switch feature diagram In addition, regarding Upsampling is performed using bilinear interpolation to obtain a dimension of Disconnecting switch feature diagram Finally, Activated via the Swish function, and with Perform a dot product operation to obtain the feature map of the disconnector switch. , The dimension is .

[0122] In a further embodiment of the present invention, and The transformation formula is as follows:

[0123] (3)

[0124] (4)

[0125] In the formula, Swish is the activation function, and MaxPool is the max pooling operation. It is a convolution operation with a kernel size of 1 and a stride of 1. It is a convolution operation with a kernel size of 3 and a stride of 1. BI stands for Bilinear Interpolation.

[0126] It's worth noting that max pooling is a common downsampling operation, typically used in convolutional neural networks (CNNs). Its function is to extract the maximum value from each local region (such as a 2x2 or 3x3 window) of the input feature map, thereby reducing the size of the feature map while preserving the most salient features.

[0127] Swish is a smooth and non-monotonic activation function proposed by Google. Its formula is:

[0128]

[0129] Bilinear interpolation is a commonly used image upsampling method to enlarge low-resolution images or feature maps to high resolution. It calculates the target pixel value by weighting the four nearest surrounding pixels, and is characterized by its simplicity and smooth output.

[0130] Step 3.3: Perform multi-scale feature extraction and fusion on the feature map output by DASC-Conv.

[0131] In one embodiment of the present invention, the multi-scale feature processing module is an MSFE-Conv module, the structure of which is as follows: Figure 6 As shown. The dimension to be output from DASC-Conv is Feature diagram of a disconnector switch on a utility pole The input is fed into the MSFE-Conv module for multi-scale feature extraction and fusion.

[0132] In the MSFE-Conv module, firstly... Convolution operations with kernel size 3 and stride 1 and convolution operations with kernel size 1 and stride 1 were performed respectively to obtain the feature maps of the disconnector switch. Disconnecting switch feature diagram The main purpose of this step is to extract local details and global structural features from the feature map of the disconnector switch; then... and Perform a channel concat operation to obtain the feature map of the disconnector switch. The main purpose of this step is to fuse multi-scale features and enhance feature diversity; then... Perform global average pooling to obtain the feature map of the disconnector switch. The main purpose of this step is to compress the spatial dimension and extract global features; then... The GELU activation function is used to perform the activation operation, resulting in the feature map of the disconnector switch. The main purpose of this step is to introduce nonlinearity to enhance the expressive power of features; then... Perform a convolution operation with a kernel size of 1 and a stride of 1, and adjust the number of channels to match the kernel size of 1. Similarly, the characteristic map of the disconnecting switch is obtained. The main purpose of this step is to integrate multi-scale features; finally, and Perform a dot product operation to obtain the characteristic map of the output disconnect switch. The main purpose of this step is to combine global features with the original input features through dot product operations, thereby enhancing the model's sensitivity to key features. The MSFE-Conv module extracts key features from the input isolating switch feature map step by step through multi-scale feature extraction and feature fusion, and enhances the model's expressive power through nonlinear activation and feature fusion.

[0133] In a further embodiment of the present invention, , and The transformation formula is as follows:

[0134] (5)

[0135] (6)

[0136] (7)

[0137] In the formula, It's average pooling, and GELU is the activation function. It is a convolution operation with a kernel size of 1 and a stride of 1. It is a convolution operation with a kernel size of 3 and a stride of 1.

[0138] It's important to note that GELU is an activation function that combines the ideas of ReLU and Dropout, enhancing the expressive power of neural networks by introducing the cumulative distribution function (CDF) of a Gaussian distribution. GELU performs exceptionally well in natural language processing (NLP) and computer vision (CV) tasks, and is particularly widely used in Transformer models such as BERT and GPT.

[0139] Step 3.4: Introduce the EA attention mechanism and perform detection through the Detect module.

[0140] In one embodiment of the present invention, the state detection module is a Detect module, and its structure is as follows: Figure 7 As shown.

[0141] Will The input is fed into the EA module for attention-based feature enhancement to obtain a feature map. The on and off state detection results are output through the disconnector switch state prediction module Detect.

[0142] In the Detect module, the feature map of the disconnect switch is first... The input is fed into the global average pooling module (AvgPooling), which averages the feature values ​​of each channel along the spatial dimension to generate the isolating switch feature map. The feature map of the disconnector switch. A fully connected layer maps the channel dimension to the number of target categories (2 in this case, including the open and closed states of the isolator switch), outputting a feature vector T. A softmax function is then applied to normalize T into a probability distribution. ,in ,in, This represents the probability that the disconnector switch is in the open state. This represents the probability that the disconnector switch is in the closed state. The calculation process is as follows:

[0143] (8)

[0144] (9)

[0145] In the formula, It's an average pooling operation; FC is a fully connected layer. It is a normalization function, and P is the normalized probability distribution of the open and closed states of the disconnector switch.

[0146] It should be noted that the Softmax function is a function that maps any real vector to a probability distribution. Given an input vector... The output of the Softmax function .

[0147] The Enhanced Attention Mechanism (EA) is an attention mechanism used to enhance the feature representation capabilities of deep learning models. It dynamically adjusts the importance of each channel and spatial location in the feature map by combining channel attention and spatial attention, thereby enhancing key features and suppressing irrelevant background information. The EA attention mechanism is widely used in computer vision tasks such as image classification, object detection, and semantic segmentation, and can significantly improve model performance.

[0148] In one specific implementation, the input dimension is a 1024×1024×12 (width×height×channel) isolation switch image. First, in the MSCA-Conv module, Perform a convolution operation with a kernel size of 1 and a stride of 1, and then input it into a BN layer for normalization to obtain a feature map of the isolation switch with an output dimension of 1024×1024×12. At the same time, input Perform dilated convolution with a kernel size of 3 and a stride of 1 (dilation rate = 2), and input it into a BN layer for normalization to obtain a 1024×1024×12 feature map of the isolating switch. In addition, the following will be entered: A depthwise separable convolution operation with a kernel size of 5 and a stride of 1 is performed, and the result is fed into a Batch Normalization (BN) layer for normalization, yielding a 1024×1024×12 dimension isolation switch feature map. .

[0149] After that , , Channel splicing yields a feature map of the disconnector switch with dimensions of 1024×1024×36. Then on Performing a convolution operation with a kernel size of 1 and a stride of 1 yields a 1024×1024×12 dimension isolation switch feature map. Then on Global average pooling is performed to obtain a 1×1×12 feature map of the disconnector switch. .

[0150] Then to Attention weights are generated through a fully connected layer and normalized using the Sigmoid function to obtain a 1×1×12 feature map of the isolation switch. Finally Feature diagram of input isolation switch Perform a dot product and output a feature map of the disconnector switch with dimensions of 1024×1024×12. .

[0151] Then The input is first processed into the DASC-Conv module. Performing a depthwise separable convolution operation with a kernel size of 1 and a stride of 1 yields a 1024×1024×12 feature map of the isolating switch. Then on Perform a dynamic convolution operation with a kernel size of 3 and a stride of 1 to obtain a 1024×1024×12 dimension isolation switch feature map. Then on Max pooling was performed using MaxPool to obtain a feature map of the isolation switch with dimensions of 512×512×12. Then on Upsampling is performed using bilinear interpolation, and activation is applied using the Swish function to obtain a 1024×1024×12 dimension isolation switch feature map. Finally, and Performing a dot product operation yields an output isolating switch feature map with dimensions 1024×1024×12. .

[0152] The feature map of the utility pole disconnect switch with dimensions of 1024×1024×12 will be output from DASC-Conv. The input is fed into the MSFE-Conv module for multi-scale feature extraction and fusion. In the MSFE-Conv module, the first step is to... Perform convolution operations with kernel size 3 and stride 1 and convolution operations with kernel size 1 and stride 1 respectively to obtain a 1024×1024×12 dimension isolation switch feature map. And a feature map of disconnecting switches with dimensions of 1024×1024×12. Then and Perform a channel concat operation to obtain a disconnector feature map with dimensions of 1024×1024×24. Then on Perform global average pooling to obtain a 1×1×24 feature map of the isolation switch. Then on Activation is performed using the GELU activation function to obtain a 1×1×24 feature map of the disconnector switch. Then on Perform a convolution operation with a kernel size of 1 and a stride of 1, and adjust the number of channels to match the kernel size of 1. Similarly, a feature map of the disconnector switch with a dimension of 1×1×12 is obtained. Finally, With disconnect switch Performing a dot product operation yields a feature map of the disconnector switch with an output dimension of 1024×1024×12. .

[0153] Will The input is fed into the ECA module for attention-based feature enhancement, resulting in a 1024×1024×12 dimension disconnector feature map. The on and off state detection results are output through the disconnector switch state prediction module Detect.

[0154] In the disconnector switch state prediction module Detect, firstly... The input is fed into the global average pooling module (AvgPooling), which averages the feature values ​​of each channel along the spatial dimension to generate a 1×1×12 feature map of the isolation switch. The feature map of the disconnector switch. A fully connected layer maps the channel dimension to the number of target categories (2 in this case, including the open and closed states of the disconnector switch), outputting a 1×1×2 feature vector T. A softmax function is then applied to T for activation and normalized to a probability distribution. , The probability that the disconnect switch is in the open state is 0.85. This indicates that the probability of the disconnector being in the closed state is 0.15, because... Therefore, the current state of the disconnect switch is open.

[0155] S4: Load the training and validation sets into the network for training to update the network parameters, and then use the test set as input to verify the network performance.

[0156] During the training of the MDMENet network, parameters are first initialized based on the electrical characteristics of the equipment. A multi-condition alternating optimization algorithm is employed, dynamically adjusting the learning weights for sunny and foggy day samples during training. Model optimization is performed using a composite loss function (including state classification, switch contour, and arc feature losses). A fast alternating optimization strategy is used during training, minimizing the binary cross-entropy loss function to measure the difference between the model's predictions and the true labels. In the training phase, preprocessed training data is input into the network, with a batch size of 2 and 1000 iterations. Backpropagation is performed using stochastic gradient descent, and the Adam optimizer is introduced to dynamically adjust the learning rate to accelerate network convergence and improve training stability.

[0157] After training, the validation set data is input into the trained MDMENet network to calculate the accuracy of detecting the open / closed state of the disconnector switch, thereby evaluating the model's performance on the validation set. Based on the validation results, the network performance is further optimized, including adjusting hyperparameters such as the learning rate and regularization parameters, and fine-tuning the network structure (such as adding attention mechanism branches or introducing residual connections), thereby improving the model's detection accuracy and generalization ability.

[0158] The test set obtained from Step 2 will be input into the MDMENet network to evaluate its performance in the disconnector switch opening / closing state detection task, including metrics such as accuracy and stability. Based on the evaluation results, the model and system will be further improved and optimized.

[0159] It should be noted that the composite loss function is an optimization objective function constructed by weighted combination of multiple basic loss functions. In disconnector switch status detection, it typically includes classification loss (ensuring accurate status judgment), localization loss (precisely locating the disconnector switch position), and edge loss (enhancing contour recognition). The contributions of each loss term are balanced by weight coefficients (e.g., 0.6, 0.3, 0.1), ultimately forming a joint optimization objective that comprehensively optimizes model performance. This design maintains the discriminative ability of the main task while improving fine-grained feature learning through auxiliary loss terms. In practical applications, it can improve detection accuracy by more than 5% while significantly reducing localization errors.

[0160] S5: Apply the obtained network to the disconnector switch opening / closing status detection task.

[0161] The MDMENet network is applied to the task of detecting the open / closed state of disconnectors. Images of the disconnectors are periodically captured from static and dynamic images and then input into the MDMENet network. The MDMENet network processes the images and outputs the probability distribution of the disconnectors. This allows us to determine the open / closed state of the disconnector switch. When If the disconnector is in the open state, it is determined that the disconnector is in the closed state; otherwise, it is in the closed state.

[0162] Based on the same inventive concept, this application also provides a deep learning-based disconnector switch opening / closing state detection device for implementing the aforementioned deep learning-based disconnector switch opening / closing state detection method. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations in the following embodiments of the deep learning-based disconnector switch opening / closing state detection device can be found in the above-described limitations of the deep learning-based disconnector switch opening / closing state detection method, and will not be repeated here.

[0163] Please see Figure 8 This invention also provides a deep learning-based device for detecting the open / closed state of an isolating switch, comprising:

[0164] The image acquisition module is used to acquire images of the disconnecting switch;

[0165] The detection module is used to detect disconnector images using a pre-trained disconnector detection network to obtain the detection results of the disconnector's open and closed state.

[0166] The disconnector detection network utilizes a first feature extraction and processing module and a second feature extraction and processing module for feature extraction. The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnector image using a multi-branch convolutional structure to obtain preliminary features while maintaining the spatial resolution of the image. The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers to obtain the key features of the disconnector image.

[0167] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments 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. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0168] Reference Figure 9 The present invention also provides a computer device, including: a memory and a processor, and a computer program stored in the memory. When the computer program is executed on the processor, it implements the deep learning-based method for detecting the open / closed state of a disconnector switch as described in any of the above methods.

[0169] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 9 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.

[0170] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0171] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0172] This invention also provides a computer-readable storage medium storing a computer program thereon. When the computer program is run by a processor, it implements the deep learning-based method for detecting the open / closed state of a disconnector switch as described in any of the above methods.

[0173] In this embodiment, 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, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0174] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0175] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0176] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or 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 displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0177] 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 the open / closed state of a disconnector switch based on deep learning, characterized in that, Includes the following steps: Acquire images of disconnect switches; The disconnector switch image is detected using a pre-trained disconnector switch detection network to obtain the detection results of the disconnector switch's open / closed state; The disconnector detection network utilizes a first feature extraction and processing module and a second feature extraction and processing module for feature extraction. The first feature extraction and processing module employs a multi-branch convolutional structure to perform preliminary feature extraction on the input disconnector image, thereby obtaining preliminary features while maintaining the image spatial resolution. The second feature extraction and processing module employs depthwise separable convolutional layers and dynamic convolutional layers to perform deep feature mining on the preliminary features, thereby obtaining the key features of the disconnector image. The first feature extraction and processing module is the MSCA-Conv module, and the image processing procedure of the MSCA-Conv module includes: By performing convolution, dilated convolution, and depthwise separable convolution operations on the input isolating switch feature map using a set of Conv modules, the second, third, and fourth isolating switch feature maps are obtained. The feature maps of the second disconnector, the third disconnector, and the fourth disconnector are normalized using a BN layer. pass The layer performs a stitching operation on the standardized disconnector feature map to obtain the fifth disconnector feature map; The fifth disconnector feature map is convolved using a Conv module to obtain the sixth disconnector feature map; The feature map of the sixth disconnector is processed by the global average pooling module to obtain the feature map of the seventh disconnector; Attention weights are generated on the feature map of the seventh disconnector through a fully connected layer, and then normalized using the Sigmoid function to obtain the feature map of the eighth disconnector. The eighth disconnector feature map is multiplied by the input disconnector feature map to obtain the ninth disconnector feature map, and the ninth disconnector feature map is used as the final output of the MSCA-Conv module.

2. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 1, characterized in that, The disconnector switch detection network includes: The system consists of a backbone feature extraction network, a multi-scale feature processing module, an attention enhancement module, and a state detection module. The backbone feature extraction network includes a first feature extraction and processing module and a second feature extraction and processing module; The first feature extraction and processing module is used to perform preliminary feature extraction on the input disconnect switch image. Through different convolutional layers in the multi-branch convolutional structure, features are extracted from different angles, and then the results of each branch are stitched together to obtain a preliminary feature map while maintaining the spatial resolution of the image. The second feature extraction and processing module is used to perform deep feature mining on the preliminary features using depth-separable convolutional layers and dynamic convolutional layers, thereby obtaining the key feature map of the disconnector image; The multi-scale feature processing module is used to extract and fuse the key features at multiple scales to obtain a multi-scale feature map. The attention enhancement module is used to calculate and adjust the importance weights in the multi-scale feature map based on the EA attention mechanism to obtain a feature map that highlights key features; The state detection module is used to determine the final open / closed state detection result of the disconnector switch based on the feature map of the prominent key features.

3. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 1, characterized in that, The transformation formulas for the feature diagrams of the fifth disconnector and the ninth disconnector are as follows: In the formula, This represents a feature diagram of the fifth disconnecting switch. This indicates a splicing operation. Indicates standardized processing; , and These represent convolution operations with a kernel size of 1 and a stride of 1, dilated convolution operations with a kernel size of 3 and a stride of 1, and depthwise separable convolution operations with a kernel size of 5 and a stride of 1, respectively. The feature map representing the input isolating switch; This represents a feature diagram of the ninth isolating switch.

4. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 1 or 2, characterized in that, The second feature extraction and processing module is a DASC-Conv module, and the image processing procedure of the DASC-Conv module includes: The first Conv module performs a depthwise separable convolution operation on the input isolator feature map to obtain the tenth isolator feature map; The eleventh disconnector feature map is obtained by performing a dynamic convolution operation on the tenth disconnector feature map using the second Conv module. The eleventh disconnector feature map is processed by the max pooling module to obtain the twelfth disconnector feature map; The 12th disconnector feature map is obtained by performing bilinear interpolation on the bilinear interpolation module. The Swish function module activates the thirteenth disconnector feature map and multiplies it with the tenth disconnector feature map to obtain the fourteenth disconnector feature map, which is then used as the final output of the DASC-Conv module.

5. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 4, characterized in that, The transformation formulas for the thirteenth disconnect switch feature diagram and the fourteenth disconnect switch feature diagram are as follows: In the formula, This diagram shows the features of the thirteenth disconnector. This represents the bilinear interpolation operation. This indicates a max pooling operation. This represents a convolution operation with a kernel size of 3 and a stride of 1. This represents a convolution operation with a kernel size of 1 and a stride of 1. The feature map representing the input isolating switch; express Activation function This represents the feature diagram of the tenth disconnecting switch. This represents the feature diagram of the fourteenth disconnecting switch.

6. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 2, characterized in that, The multi-scale feature processing module is an MSFE-Conv module, and the image processing procedure of the MSFE-Conv module includes: The input isolating switch feature maps are convolved using two Conv modules of different sizes to obtain the fifteenth isolating switch feature map and the sixteenth isolating switch feature map. The feature map of the fifteenth disconnector and the feature map of the sixteenth disconnector are spliced ​​together by the channel splicing module to obtain the feature map of the seventeenth disconnector. The feature map of the seventeenth disconnector is obtained by performing a global average pooling operation on the feature map of the seventeenth disconnector through the global average pooling module. The 18th disconnector feature map is activated by the GELU activation function to obtain the 19th disconnector feature map; The twentieth disconnector feature map is obtained by performing a convolution operation on the nineteenth disconnector feature map using a Conv module. The 20th disconnector feature map and the input disconnector feature map are multiplied by a dot product to obtain the 21st disconnector feature map, which is then used as the final output of the MSFE-Conv module.

7. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 6, characterized in that, The transformation formulas for the seventeenth disconnector feature diagram, the twentieth disconnector feature diagram, and the twenty-first disconnector feature diagram are as follows: In the formula, This diagram shows the features of the seventeenth disconnecting switch. This indicates a channel splicing operation. This represents a convolution operation with a kernel size of 1 and a stride of 1. This represents a convolution operation with a kernel size of 3 and a stride of 1. The feature map representing the input isolating switch; This represents the feature diagram of the twentieth disconnector. express Activation function This indicates a global average pooling operation. This represents the feature diagram of the 21st disconnecting switch.

8. The method for detecting the open / closed state of a disconnector switch based on deep learning according to claim 2, characterized in that, The state detection module is a Detect module. The feature map highlighting the key features is denoted as the twenty-second disconnector switch feature map. Based on the feature map highlighting the key features, the final disconnector switch opening / closing state detection result is determined, including: The 22nd disconnector feature map is generated by averaging the feature values ​​of each channel along the spatial dimension using the global average pooling module. The feature map of the 23rd disconnector switch is passed through a fully connected layer to map the channel dimension to the number of target categories, and the feature vector is output. The feature vector is normalized into a probability distribution using the Softmax function. ,in ,in, This represents the probability that the disconnector switch is in the open state. This represents the probability that the disconnector switch is in the closed state.

9. A deep learning-based device for detecting the open / closed state of a disconnector switch, characterized in that, include: The image acquisition module is used to acquire images of the disconnecting switch; The detection module is used to detect the disconnector image using a pre-trained disconnector detection network to obtain the detection result of the disconnector's open / closed state; The disconnector detection network utilizes a first feature extraction and processing module and a second feature extraction and processing module for feature extraction. The first feature extraction and processing module employs a multi-branch convolutional structure to perform preliminary feature extraction on the input disconnector image, thereby obtaining preliminary features while maintaining the image spatial resolution. The second feature extraction and processing module employs depthwise separable convolutional layers and dynamic convolutional layers to perform deep feature mining on the preliminary features, thereby obtaining the key features of the disconnector image. The first feature extraction and processing module is the MSCA-Conv module, and the image processing procedure of the MSCA-Conv module includes: By performing convolution, dilated convolution, and depthwise separable convolution operations on the input isolating switch feature map using a set of Conv modules, the second, third, and fourth isolating switch feature maps are obtained. The feature maps of the second disconnector, the third disconnector, and the fourth disconnector are normalized using a BN layer. pass The layer performs a stitching operation on the standardized disconnector feature map to obtain the fifth disconnector feature map; The fifth disconnector feature map is convolved using a Conv module to obtain the sixth disconnector feature map; The feature map of the sixth disconnector is processed by the global average pooling module to obtain the feature map of the seventh disconnector; Attention weights are generated on the feature map of the seventh disconnector through a fully connected layer, and then normalized using the Sigmoid function to obtain the feature map of the eighth disconnector. The eighth disconnector feature map is multiplied by the input disconnector feature map to obtain the ninth disconnector feature map, and the ninth disconnector feature map is used as the final output of the MSCA-Conv module.

Images