Isolation switch state identification method and device

Abstract

The invention discloses an isolation switch state identification method and device, and the method comprises the steps: S1, constructing an isolation switch shooting unit, and carrying out the shooting of an isolation switch, and obtaining an original isolation switch image data set; s2, constructing a state detection chip, detecting the original isolation switch image, and judging the state of the isolation switch; the state monitoring chip is composed of an image preprocessing unit, an image feature recognition unit and a state judgment unit. S3, constructing a notification and alarm unit, and triggering a local audible and visual alarm in real time according to a state judgment result of the state detection chip on the isolating switch, or sending state information of the isolating switch to a monitoring platform and a mobile operation and maintenance terminal through wireless communication; s4, constructing an energy supply unit, and supplying power to the disconnecting switch shooting unit, the state monitoring chip and the notification and alarm unit; by adopting the technical scheme of the invention, the state of the disconnecting switch can be distinguished.

Description

TECHNICAL FIELD

[0001] The technical scheme belongs to the technical field of smart grid operation and maintenance, and particularly relates to a state discrimination method and device for disconnectors. BACKGROUND

[0002] Traditional disconnector state detection mainly relies on manual inspection or simple sensor detection, which is not only low in efficiency and high in labor intensity, but also difficult to find abnormal states in time, thus existing certain safety hazards. In recent years, image recognition technology based on deep learning is introduced into the field of disconnector state detection, through collecting disconnector images and performing feature analysis, automatic recognition of open and closed states is realized, and the detection efficiency and accuracy are improved.

[0003] The most similar technical scheme to the present application currently available mainly includes: (1) A disconnector state recognition method based on convolutional neural network (CNN). This kind of method usually extracts disconnector features through single-scale convolution, and then performs state discrimination through full connection layer or simple classifier, to realize automatic recognition of open and closed states. However, this kind of scheme has limited ability to extract local details and edge information of the disconnector under complex background, and is difficult to cope with images of different illuminations, different angles and different states, and the application of multi-scale feature fusion and attention mechanism is insufficient, resulting in insufficient precision, robustness and remote monitoring capability of the model; (2) A method based on deep learning. This kind of method is mostly based on single-scale convolution or simple image processing technical idea, and is difficult to fully extract local details and edge information of the disconnector, and is prone to misjudgment under complex background or illumination conditions; and it lacks effective adaptation mechanism for disconnector images of different illuminations, rain, snow, haze or different states, resulting in unstable recognition results. Most methods rely only on single resolution or single channel features, and cannot capture global structure and local details at the same time, and are difficult to be sensitive to slight state changes, in addition, the existing methods lack efficient communication and alarm mechanism, and cannot realize real-time remote monitoring and abnormal prompt of the disconnector state. SUMMARY

[0004] The present application aims to solve the problems of low recognition accuracy, poor adaptability to complex environment, insufficient extraction of local detail information and limited remote monitoring capability of existing disconnector state detection methods. The specific technical purposes include: (1) To improve the accuracy and robustness of disconnector open and closed state recognition, so that the detection can still be stable and reliable under complex background, different illumination and different state conditions; (2) To realize multi-scale feature extraction and adaptive fusion of disconnector state images, to fully retain edge, texture and local detail information, and to improve the sensitivity of the model to slight state changes; (3) Support remote real-time monitoring and alarm, the identification result and the confidence score are reliably transmitted to the mobile terminal through the wireless communication module, and the operation and maintenance personnel can realize the timely grasp and management of the state of the isolating switch.

[0005] To achieve the above object, the technical scheme adopted by the present application is as follows: An isolating switch state recognition method, characterized in that it comprises: S1, an isolating switch shooting unit is constructed to shoot the isolating switch and obtain an original isolating switch image data set; S2, a state detection chip is constructed to detect the original isolating switch image and determine the state of the isolating switch; the state detection chip is composed of an image preprocessing unit, an image feature recognition unit and a state judgment unit; S3, a notification and alarm unit is constructed to trigger a local sound and light alarm in real time according to the state determination result of the isolating switch by the state detection chip, or send the isolating switch state information to a monitoring platform and a mobile operation and maintenance terminal through wireless communication; S4, an energy supply unit is constructed to supply power to the above-mentioned isolating switch shooting unit, state detection chip and notification and alarm unit.

[0006] As a preferred, the execution steps of the isolating switch shooting unit in S1 comprise: The isolating switch shooting unit comprises a high-definition imaging collection device for collecting the original isolating switch image in real time; the high-definition imaging collection device is used to collect images of the isolating switch under multiple angles, multiple illuminations and multiple weather conditions; during the shooting process, the image definition is ensured through automatic exposure, gain adjustment and light supplementing device, and the shooting time, environmental parameters and camera information are recorded; the original isolating switch image is collected to construct an isolating switch image data set.

[0007] As a preferred, the construction process of the state detection chip in S2 comprises: S21, an image preprocessing unit is constructed to process the original isolating switch image; first, an adaptive histogram equalization algorithm is adopted to improve the overall contrast of the image and enhance the detail performance of the isolating switch and its background; then, a bilateral filtering method is applied to remove background noise and texture interference while retaining the edge profile; finally, image size normalization and color space conversion are performed to generate an isolating switch image data of uniform size; S22, an image feature recognition unit is constructed to recognize the state features of the isolating switch from the isolating switch image; the image feature recognition unit is composed of a parallel multi-scale channel attention fusion sub-module, a multi-resolution interactive enhancement convolution sub-module, a frequency domain enhancement channel-space fusion convolution sub-module and a triple attention sub-module; S23. Construct a status judgment unit to determine the status of the disconnect switch based on its status characteristics.

[0008] Preferably, the execution flow of the image feature recognition unit in S22 includes: Step 211: Transfer the image of the disconnect switch. The input is fed into the parallel multi-scale channel attention fusion submodule, where parallel multi-scale channel attention fusion is performed to obtain the isolation switch feature map. ; Step 212, The input is fed into the multi-resolution interactive enhanced convolution submodule to perform multi-resolution interactive enhanced convolution operations, resulting in the isolating switch feature map. ; Step 213, The input is fed into the frequency domain enhancement channel-spatial fusion convolution submodule, where a frequency domain enhancement channel-spatial fusion convolution operation is performed to obtain the isolating switch feature map. Then The input is fed into the triple attention submodule for attention enhancement to obtain the feature map of the disconnector switch. .

[0009] Preferably, the execution steps of the parallel multi-scale channel attention fusion submodule described in S211 include: S2111. Let the input of the parallel multi-scale channel attention fusion submodule be a feature map of the isolation switch with size H×W×C. ,right A convolution operation with a kernel size of 5×5 and a stride of 1 is performed, followed by a normalization layer and a channel scaling submodule, to obtain a feature map of the isolation switch with dimensions H×W×C. At the same time, for Perform grouped convolution operations with a kernel size of 3×3 and a stride of 1, and then pass them through a batch normalization layer to obtain a feature map of the disconnector with dimensions H×W×C. Where H represents the height of the disconnector feature map, W represents the width of the disconnector feature map, and C represents the number of channels in the disconnector feature map; S2112, to and By performing a channel-by-channel splicing operation, a feature map of a disconnecting switch with dimensions H×W×2C is obtained. ;right Perform a convolution operation with a kernel size of 1×1 and a stride of 1 to obtain a feature map of the disconnector switch with dimensions H×W×C. ;right Perform an adaptive average pooling operation to obtain a feature map of the disconnector switch with size H×W×C. ; (1) (2) (3) wherein, represents a convolution operation with a kernel size of 5x5 and a step size of 1; represents a grouped convolution operation with a kernel size of 3x3 and a step size of 1; represents a batch normalization layer; represents a channel scaling module sub-module; represents a channel-wise concatenation operation; represents a convolution operation with a kernel size of 1x1 and a step size of 1; represents an adaptive average pooling operation; S2113, passing through a fully connected layer and a Gelu activation function in sequence, to obtain an isolation switch feature map with a size of HxWx1 ; then passing through a fully connected layer and a Sigmoid activation function in sequence, to obtain an isolation switch feature map with a size of HxWx1 ; S2114, performing a point multiplication operation on and , to obtain an isolation switch feature map with a size of HxWxC : (4) (5) (6) wherein, represents a fully connected layer; represents an activation function; represents an activation function; represents a point multiplication operation.

[0010] As a preferred, the execution steps of the multi-resolution interaction enhancement convolution sub-module of S212 include: S2121, let the input of the parallel multi-scale channel attention fusion sub-module be an isolation switch feature map with a size of HxWxC , perform a convolution operation with a kernel size of 3x3 and a step size of 1 on , to obtain an isolation switch feature map with a size of HxWxC ; at the same time, perform a convolution operation with a kernel size of 3x3 and a step size of 1 on A convolution operation with a kernel size of 5x5 and a step size of 1 is performed to obtain an isolation switch feature map with a size of HxWxC ; S2122, a channel concatenation operation is performed on and to obtain an isolation switch feature map with a size of HxWx2C ; S2123, a global max pooling operation is performed on to obtain an isolation switch feature map with a size of 1x1x2C ; is activated by a Swish activation function to obtain an isolation switch feature map with a size of 1x1x2C ; ; S2124, a convolution operation with a kernel size of 1x1 and a step size of 1 is performed on to obtain an isolation switch feature map with a size of 1x1xC ; S2125, point multiplication is performed on and to obtain an isolation switch feature map with a size of HxWxC : (7) (8) (9) (10) (11) (12) wherein, represents a convolution operation with a kernel size of 3x3 and a step size of 1; represents a convolution operation with a kernel size of 1x1 and a step size of 1; represents a global max pooling operation; represents an activation function.

[0011] As a preferred, the execution steps of the frequency domain enhancement channel-spatial fusion convolution sub-module in S213 include: S2131, the input of the parallel multi-scale channel attention fusion sub-module is denoted as an isolation switch feature map with a size of HxWxC ; a convolution operation with a kernel size of 1x1 and a step size of 1 is performed on to obtain an isolation switch feature map with a size of HxWxC ; S2132, performing two-dimensional Fourier transform, extracting amplitude information of the frequency spectrum, and performing frequency domain feature enhancement through 1x1 convolution+BN+SiLU activation to obtain an isolation switch feature map with a size of HxWxC ; meanwhile, performing depth separable convolution operation with a convolution kernel size of 3x3 and a step size of 1 on to obtain an isolation switch feature map with a size of HxWxC ; S2133, performing channel splicing operation on , to obtain an isolation switch feature map with a size of HxWx2C : (13) (14) ( ) (15) (16) wherein, represents two-dimensional Fourier transform operation; BN represents batch normalization operation; represents activation function; represents depth separable convolution operation with a convolution kernel size of 3x3 and a step size of 1; S2134, performing convolution operation with a convolution kernel size of 1x1 and a step size of 1 on to obtain an isolation switch feature map with a size of HxWxC ; S2135, performing channel attention operation on to obtain an isolation switch feature map with a size of HxWxC ; meanwhile, performing spatial attention operation on to obtain an isolation switch feature map with a size of HxWxC ; S2136, performing point multiplication operation on , , to obtain an isolation switch feature map with a size of HxWxC : (17) (18) (19) (20) wherein, represents a channel attention operation; represents a spatial attention operation.

[0012] Preferably, the execution flow of the state judging unit in S23 comprises: The isolation switch feature map output by the image feature recognition unit is input into the state judging unit to judge the state of the isolation switch. The state judging unit is internally provided with a lightweight discrimination network, which analyzes and reasons through a multi-layer convolution and full connection structure. Two neurons are arranged at the output layer of the network, which respectively output the open state probability and the closed state probability of the isolation switch. When , the state of the isolation switch is determined as the open state, otherwise as the closed state.

[0013] Preferably, the execution steps of the notification and alarm unit in S3 comprise: The notification and alarm unit comprises a wireless communication module and a local alarm module, which are used to receive the state determination result of the isolation switch by the state detection chip and determine whether to send corresponding alarm or notification operation according to the operation and maintenance rules. The wireless communication module is used to upload the state of the isolation switch and related environmental parameters to the monitoring platform or the mobile operation and maintenance terminal in real time.

[0014] The local alarm module comprises an audible and visual alarm device, which can immediately send an alarm prompt when an abnormal state is detected. In the notification and alarm unit, a buffering and retransmission mechanism is internally provided, which is used to enhance the real-time performance and reliability of sending alarm information in the case of network anomaly or communication delay. The application further provides an isolation switch state discrimination device, characterized by comprising: An isolation switch shooting unit shoots the isolation switch to obtain an original isolation switch image data set. A state detection chip detects the original isolation switch image to judge the state of the isolation switch. The state detection chip is composed of an image preprocessing unit, an image feature recognition unit and a state judging unit. A notification and alarm unit outputs the state of the isolation switch according to the state determination result of the isolation switch by the state detection chip.

[0015] An energy supply unit supplies power to the above-mentioned isolation switch shooting unit, state detection chip and notification and alarm unit.The technical purpose of the present application is to construct an isolator state recognition method and device, which significantly improves the accuracy and robustness of isolator state detection. Compared with the traditional method, the present application designs a parallel multi-scale channel attention fusion sub-module (PMCA-Fuse-Conv), which improves the state detection accuracy in complex background through parallel multi-scale feature extraction and channel adaptive weight generation; the present application also designs a multi-resolution interactive enhancement convolution sub-module (MRIE-Conv), which extracts feature information of different scales through parallel multi-resolution convolution paths, fuses local details and global context through a cross-resolution feature interaction mechanism, and then enhances the features through a channel re-calibration mechanism, finally improving the accuracy and robustness of isolator state recognition; the present application also designs a frequency domain enhancement channel-spatial fusion convolution sub-module (GTS-FECA-Conv), which introduces frequency domain information on the basis of traditional convolution feature extraction, obtains the frequency spectrum features of the image through Fourier transform, and selectively enhances the key information through a channel and spatial dual attention mechanism (FCA + SA), thereby effectively improving the sensitivity of the features to texture, edges and local detail changes. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0017] Figure 1 The flow chart of the isolator state recognition technical solution of the present application embodiment is shown in the figure. Figure 2 The schematic diagram of the isolator in the closed state is shown in the figure. Figure 3 The schematic diagram of the isolator in the open state is shown in the figure. Figure 4 The structure diagram of the image feature recognition unit is shown in the figure. Figure 5 The structure diagram of the parallel multi-scale channel attention fusion sub-module is shown in the figure. Figure 6 The structure diagram of the multi-resolution interactive enhancement convolution sub-module is shown in the figure. Figure 7 The structure diagram of the frequency domain enhancement channel-spatial fusion convolution sub-module is shown in the figure. Figure 8 The structure diagram of the isolator open and closed state discrimination device of the present application embodiment is shown in the figure. DETAILED DESCRIPTION

[0018] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0019] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0020] Embodiment 1 As Figure 1 shown, the embodiment of the present application provides a kind of isolator state identification method, comprising: S1, a isolator shooting unit is constructed, and the isolator is shot, and the original isolator image data set is obtained; S2, a state detection chip is constructed, and the original isolator image is detected, and the state of the isolator is judged;The state monitoring chip is composed of image preprocessing unit, image feature recognition unit and state judging unit; S3, a notification and alarm unit is constructed, and according to the state detection chip state determination result of the isolator, local sound and light alarm is triggered in real time, or the isolator state information is sent to monitoring platform and mobile operation terminal by wireless communication; S4, a power supply unit is constructed, and the above isolator shooting unit, state detection chip, notification and alarm unit are powered.

[0021] As an embodiment of the present application, in step S1, the execution steps of the isolator shooting unit include: The isolator shooting unit includes a high-definition imaging acquisition device for real-time acquisition of original isolator images: the high-definition imaging acquisition device is used for image acquisition under multiple angles, multiple illuminations and multiple weather conditions of the isolator, including daytime, nighttime, backlight, rain, snow and haze scenes; during shooting, the image clarity is ensured by automatic exposure, gain adjustment and light supplement device, and the shooting time, environmental parameters (including solar radiation, temperature, humidity, geographic information) and camera information are recorded;Collect original isolator images and construct isolator image data set.

[0022] As an embodiment of the present application, in step S2, the construction process of the state detection chip includes: S21. Construct an image preprocessing unit to process the original disconnector switch image. First, use an adaptive histogram equalization algorithm to improve the overall contrast of the image and enhance the detail of the disconnector switch and its background. Then, apply a bilateral filtering method to remove background noise and texture interference while preserving the edge contour. Finally, perform image size normalization and color space conversion to generate a disconnector switch image of uniform size. The disconnector switch states include: closed and open; the closed state refers to the two conductive arms of the disconnector switch being aligned in a straight line, with the contacts fully in contact and overlapping, such as... Figure 2 As shown; the "open state" refers to the disconnector switch's conductive arms not being in a straight line, and the contacts not being completely overlapped, such as... Figure 3 As shown.

[0023] S22. Construct an image feature recognition unit to recognize the state features of the isolating switch in the isolating switch image; the image feature recognition unit consists of a parallel multi-scale channel attention fusion submodule (PMCA-Fuse-Conv), a multi-resolution interactive enhancement convolution submodule (MRIE-Conv), a frequency domain enhancement channel-spatial fusion convolution submodule (GTS-FECA-Conv), and a triple attention submodule (Triplet Attention); S23. Construct a status judgment unit to determine the status of the disconnect switch based on its status characteristics.

[0024] Note: (1) Adaptive histogram equalization (AHE) is an improved image enhancement method that divides the entire image into multiple small regions (called sub-blocks or context regions) and performs histogram equalization on each region, thereby enhancing local details while maintaining overall contrast. It is especially suitable for images with uneven brightness distribution. (2) Bilateral filtering is an edge-preserving denoising algorithm. While smoothing the image and removing noise, it combines the weights of spatial distance and pixel gray difference to calculate the weighted average only for adjacent pixels with similar brightness, thereby effectively preserving the edge details of the image.

[0025] Further, in step S22, the structural schematic diagram of the image feature recognition unit is as follows: Figure 4 As shown, the execution flow of the image feature recognition unit includes: Step 211: Transfer the image of the disconnect switch. The input is fed into the parallel multi-scale channel attention fusion submodule PMCA-Fuse-Conv to perform parallel multi-scale channel attention fusion operations, resulting in the isolating switch feature map. ; Step 212, The input is input to the multi-resolution interactive enhanced convolution sub-module MRIE-Conv for multi-resolution interactive enhanced convolution operation to obtain the disconnecting switch feature map ; Step 213, the is input to the frequency domain enhanced channel-spatial fusion convolution sub-module GTS-FECA-Conv for frequency domain enhanced channel-spatial fusion convolution operation to obtain the disconnecting switch feature map , and then the is input to the triplet attention sub-module Triplet Attention for attention enhancement operation to obtain the disconnecting switch feature map .

[0026] Further, in step S211, the structure of the parallel multi-scale channel attention fusion sub-module PMCA-Fuse-Conv is as shown in Figure 5 The execution steps of the PMCA-Fuse-Conv sub-module include: S2111, the input of the parallel multi-scale channel attention fusion sub-module is the disconnecting switch feature map with a size of H×W×C , a convolution operation with a convolution kernel size of 5×5 and a step size of 1 is performed on , and then a normalization layer and a channel scaling sub-module are sequentially passed to obtain the disconnecting switch feature map with a size of H×W×C ; at the same time, a grouped convolution operation with a convolution kernel size of 3×3 and a step size of 1 is performed on , and a batch normalization layer is passed to obtain the disconnecting switch feature map with a size of H×W×C ; wherein H represents the height of the disconnecting switch feature map, W represents the width of the disconnecting switch feature map, and C represents the number of channels of the disconnecting switch feature map; S2112, the and are subjected to channel-wise concatenation operation to obtain the disconnecting switch feature map with a size of H×W×2C ; a convolution operation with a convolution kernel size of 1×1 and a step size of 1 is performed on to obtain the disconnecting switch feature map with a size of H×W×C ; a one-time adaptive average pooling operation is performed on to obtain the disconnecting switch feature map with a size of H×W×C ; (1) (2) (3) wherein, This represents a convolution operation with a kernel size of 5×5 and a stride of 1. This represents a grouped convolution operation with a kernel size of 3×3 and a stride of 1; Indicates the batch normalization layer; This indicates a submodule of the channel scaling module; This indicates a channel-based splicing operation; This represents a convolution operation with a kernel size of 1×1 and a stride of 1. This indicates an adaptive average pooling operation; S2113, to By sequentially passing through a fully connected layer and the Gelu activation function, a feature map of the disconnector with size H×W×1 is obtained. Then on By sequentially passing through a fully connected layer and a sigmoid activation function, a feature map of the disconnector switch with size H×W×1 is obtained. ; S2114, will and Performing a dot product operation yields a feature map of a disconnector switch with dimensions H×W×C. : (4) (5) (6) in, Indicates a fully connected layer; express Activation function; express Activation function; This indicates the dot product operation.

[0027] Note: (1) Layer Normalization (LN) is a technique that normalizes all feature dimensions of a single sample. It calculates the mean and variance of the sample's features, adjusts them to a mean of 0 and a variance of 1, and introduces trainable scaling parameter γ and translation parameter β to maintain the model's expressive power. LN does not depend on the batch size, and the calculation method is consistent in the training and inference stages. Therefore, it performs stably and effectively in small-batch or sequence modeling scenarios such as RNN and Transformer. (2) The Channel Shuffle Module (CSM) is a lightweight feature fusion structure. It first unifies the number of channels of multi-scale features and splices them together, and then shuffles the channels to shuffle the order, so as to promote the full interaction of channel information of features of different scales, thereby enhancing the fusion effect. It is often used in tasks that require multi-scale feature integration, such as pose estimation and target detection. (3) Grouped Convolution is a convolution method that divides the input channel into multiple groups and performs convolution operations independently in each group. It can reduce the number of parameters and computation, and support different groups to learn different feature subspaces, thereby improving efficiency and feature diversity. Typical applications include multi-GPU training of AlexNet and multi-branch feature extraction of ResNeXt.

[0028] Further, in step S212, the structure of the multi-resolution interactive enhancement convolutional submodule MRIE-Conv is as follows: Figure 6 As shown, the execution steps of the MRIE-Conv submodule include: S2121. Let the input of the parallel multi-scale channel attention fusion submodule be a feature map of the isolation switch with size H×W×C. ,right Perform a convolution operation with a kernel size of 3×3 and a stride of 1 to obtain a feature map of the disconnector switch with dimensions H×W×C. At the same time, Perform a convolution operation with a kernel size of 5×5 and a stride of 1 to obtain a feature map of the disconnector switch with dimensions H×W×C. ; S2122, to and Perform a channel-by-channel splicing operation to obtain a feature diagram of a disconnecting switch with dimensions H×W×2C. ; S2123, to Perform a global max pooling operation to obtain a feature map of a disconnector switch with a size of 1×1×2C. ;Will Activation is performed using the Swish activation function to obtain a feature map of the disconnecting switch with a size of 1×1×2C. ; S2124, Yes Perform a convolution operation with a kernel size of 1×1 and a stride of 1 to obtain a feature map of the isolating switch with a size of 1×1×C. ; S2125, will and Performing a dot product operation yields a feature map of an isolating switch with dimensions H×W×C. : (7) (8) (9) (10) (11) (12) wherein, represents a convolution operation with a kernel size of 3x3 and a stride of 1; represents a convolution operation with a kernel size of 1x1 and a stride of 1; represents a global max pooling operation; represents an activation function.

[0029] Note: (1) The Swish activation function is a smooth and non-monotonic function that allows small negative values to pass when x<0 and is close to linear when x>0, which can reduce gradient disappearance and improve the convergence and accuracy of deep networks, so it performs better than ReLU in some convolutional networks and Transformers;

[0030] (2) Global Max Pooling (GMP) is a pooling operation that takes the maximum value of each channel in the spatial dimension (usually HxW) of the entire feature map, thereby compressing each channel into a single numerical value. This not only greatly reduces the parameter and computation amount, but also preserves the strongest feature response, and is commonly used in convolutional neural network classification tasks to directly convert two-dimensional feature maps into one-dimensional vectors input to the fully connected layer.

[0031] Further, in step S212, the structure of the frequency domain enhancement channel-spatial fusion convolution sub-module GTS-FECA-Conv is as shown in Figure 7 The execution steps of the GTS-FECA-Conv sub-module include: S2131, record the input of the parallel multi-scale channel attention fusion sub-module as the isolation switch feature map with a size of HxWxC ; perform a convolution operation with a kernel size of 1x1 and a stride of 1 on to obtain an isolation switch feature map with a size of HxWxC ; S2132, perform a convolution operation with a kernel size of 1x1 and a stride of 1 on A two-dimensional Fourier transform is performed to extract the amplitude information of the spectrum, and frequency domain feature enhancement is performed through 1×1 convolution + BN + SiLU activation to obtain a feature map of the disconnector switch with size H×W×C. At the same time, Performing a depthwise separable convolution operation with a kernel size of 3×3 and a stride of 1 yields a feature map of the isolating switch with dimensions H×W×C. ; S2133, to , Perform a channel-by-channel concat operation to obtain a feature map of a disconnecting switch with dimensions H×W×2C. : (13) (14) ( (15) (16) in, BN represents the two-dimensional Fourier transform operation; BN represents the batch normalization operation. express Activation function; This indicates a depthwise separable convolution operation with a kernel size of 3×3 and a stride of 1. S2134, to Perform a convolution operation with a kernel size of 1×1 and a stride of 1 to obtain a feature map of the disconnector switch with dimensions H×W×C. ; S2135, will After channel attention operations, a feature map of the disconnecting switch with dimensions H×W×C is obtained. At the same time After spatial attention operations, a feature map of the disconnecting switch with dimensions H×W×C is obtained. ; S2136, will , , Performing a dot product operation yields a feature map of a disconnector switch with dimensions H×W×C. : (17) (18) (19) (20) in, Indicates channel attention operation; representing spatial attention operations.

[0032] Note: (1) Two-dimensional Fourier transform (2DFFT) is a fast algorithm that converts the spatial domain information of a two-dimensional signal (such as an image) into the frequency domain, which can express the pattern of brightness changes in the image with frequency components: low-frequency parts correspond to smooth areas, and high-frequency parts correspond to edges and details. Through 2DFFT, images can be efficiently analyzed, filtered, or compressed, because it decomposes complex spatial changes into a superposition of a series of sine and cosine waves; (2) Depthwise separable convolution is an operation that splits the standard convolution into depthwise convolution and pointwise convolution (1x1 convolution): depthwise convolution performs spatial convolution on each channel separately to extract local features within the channel; pointwise convolution then performs 1x1 convolution in the channel dimension to achieve information fusion between channels. This significantly reduces the number of parameters and computational complexity compared to ordinary convolution, while maintaining accuracy and speeding up model inference, commonly used in lightweight networks such as Mobile Net; (3) Channel attention (FCA) is a feature weight adjustment mechanism for the channel dimension. It first obtains the global description of each channel through global pooling (such as global average pooling or global maximum pooling), then learns the dependency between channels through a fully connected layer (or MLP), generates a set of weight coefficients between 0 and 1, and finally amplifies or suppresses each channel of the original feature map, so that the network pays more attention to the channel features useful for the task; (4) Spatial attention (SA) is a mechanism that allocates weights in the spatial position dimension. It usually first compresses the input feature map in the channel dimension (such as obtaining two single-channel feature maps through global average pooling and global maximum pooling), then concatenates them and passes them through a convolution layer to generate a spatial weight matrix, and finally enhances or suppresses each position of the original feature map according to the weight, so that the network pays more attention to the important regions in space; (5) Triplet Attention is a lightweight attention mechanism that captures the dependency between features through cross-dimensional triplet interaction. It performs attention calculation on input features in three different perspectives (height x channel, width x channel, height x width) respectively, and fuses the results using residual connection to model the correlation between space and channel. This approach can more effectively improve the model's perception of fine-grained features while keeping the computational overhead low compared to single-channel attention (such as SE module).

[0033] Further, in step S23, the execution flow of the state judgment unit includes: The isolation switch feature map output by the image feature recognition unit is input into the state judgment unit for isolation switch state judgment: the state judgment unit is built-in a lightweight discrimination network, which analyzes and infers the isolation switch feature map through a multi-layer convolution and full connection structure, and sets two neurons at the output layer of the network to output the open state probability and the closed state probability of the isolation switch respectively, and ; when , the state of the isolation switch is determined as the open state, otherwise as the closed state.

[0034] For example, in the experimental scheme of the present application, the published EfficientNet-Lite model is selected as the lightweight discrimination network. EfficientNet-Lite model is a lightweight version of EfficientNet series optimized for mobile and edge devices. It retains the original high-efficiency compound scaling idea while removing operators that rely on hardware accelerators (such as GPU / TPU) and using more general and easier-to-run-on-CPU operations, significantly reducing model parameter and computation, thus achieving faster inference speed and lower power consumption under the premise of ensuring high accuracy, which is very suitable for real-time image classification, detection and other mobile visual tasks.

[0035] As an embodiment of the present application, in step S3, the execution steps of the notification and alarm unit include: ​​The notification and alarm unit includes a wireless communication module and a local alarm module. These modules receive the status determination results of the disconnector from the status detection chip and determine whether to issue corresponding alarms or notifications based on the operation and maintenance rules. The wireless communication module supports multiple networks such as NB-IoT, 4G / 5G, or LoRa, and is used to upload the disconnector status and related environmental parameters to the monitoring platform or mobile operation and maintenance terminal in real time. The local alarm module includes an audible and visual alarm device, which can immediately issue an alarm when an abnormal status is detected. The notification and alarm unit has a built-in buffering and retransmission mechanism to enhance the real-time performance and reliability of alarm information delivery in the event of network anomalies or communication delays. The aforementioned operation and maintenance rules specify whether disconnecting switches should be in the open or closed state at various times during power grid operation procedures. If, at any given moment, the state of the disconnecting switch deviates from the operation and maintenance rules, an alarm or notification should be triggered.

[0036] For example, firstly, the imaging unit is used to construct an original disconnector switch image dataset, resulting in 2500 images; then, the image preprocessing unit of the status monitoring chip is used to preprocess the original disconnector switch image dataset: firstly, an adaptive histogram equalization algorithm is used to improve the overall contrast of the image and enhance the detail of the disconnector switch and its background; then, a bilateral filtering method is applied to remove background noise and texture interference while preserving the edge contours; finally, image size normalization and color space conversion are performed to generate disconnector switch images of uniform size.

[0037] Image of a disconnector switch of any size 1024×1024×3 The image feature recognition unit in the status monitoring chip first processes the image feature recognition data. A convolution operation with a kernel size of 5×5 and a stride of 1 is performed, followed by a normalization layer (LN) and a channel scaling submodule (CSM) to obtain a feature map of the isolating switch with a size of 1024×1024×3. At the same time, A grouped convolution (GConv) operation with a kernel size of 3×3 and a stride of 1 is performed, followed by a batch normalization layer (LN) to obtain a feature map of the isolating switch with a size of 1024×1024×3. ;right and Perform a channel-by-channel concat operation to obtain a feature map of the disconnector with dimensions of 1024×1024×6. ;right Performing a convolution operation with a kernel size of 1×1 and a stride of 1 yields a feature map of the disconnector with dimensions of 1024×1024×3. ;right An adaptive average pooling (Adaptive AVG Pooling) operation is performed to obtain an isolation switch feature map with a size of 1024x1024x3 ; and passing through a full connection layer (FC) and a Gelu activation function in sequence, an isolation switch feature map with a size of 1024x1024x1 is obtained ; and then passing through a full connection layer (FC) and a Sigmoid activation function in sequence, an isolation switch feature map with a size of 1024x1024x1 is obtained ; and and point multiplication operation is performed to obtain an isolation switch feature map with a size of 1024x1024x3 ; input to a multi-resolution interactive enhancement convolution sub-module (MRIE-Conv), first convolution kernel size of 3x3, step size of 1 is performed, to obtain an isolation switch feature map with a size of 1024x1024x3 ; and then convolution kernel size of 5x5, step size of 1 is performed, to obtain an isolation switch feature map with a size of 1024x1024x3 ; and and channel concatenation operation (Concat) is performed, to obtain an isolation switch feature map with a size of 1024x1024x6 ; and global maximum pooling operation (GMP) is performed, to obtain an isolation switch feature map with a size of 1x1x6 ; and is activated through a Swish activation function, to obtain an isolation switch feature map with a size of 1x1x6 ; and convolution kernel size of 1x1, step size of 1 is performed, to obtain an isolation switch feature map with a size of 1x1x3 ; and point multiplication operation is performed with , to obtain an isolation switch feature map with a size of 1024x1024x3 ; ; input to a frequency domain enhancement channel-spatial fusion convolution sub-module (GTS-FECA-Conv), and convolution kernel size of 1x1, step size of 1 is performed, to obtain an isolation switch feature map with a size of 1024x1024x3 ; and ​Perform two-dimensional Fourier transform (FFT), extract the amplitude information of the frequency spectrum, and perform frequency domain feature enhancement through 1*1 convolution + BN + SiLU activation, to obtain an isolation switch feature map with a size of 1024*1024*3 ; meanwhile, the is subjected to a depth separable convolution operation with a convolution kernel size of 3*3 and a step size of 1, to obtain an isolation switch feature map with a size of 1024*1024*3 ; the , is subjected to a channel concatenation operation (Concat), to obtain an isolation switch feature map with a size of 1024*1024*6 ; the is subjected to a convolution operation with a convolution kernel size of 1*1 and a step size of 1, to obtain an isolation switch feature map with a size of 1024*1024*3 ; The is subjected to a channel attention operation FCA, to obtain an isolation switch feature map with a size of 1024*1024*3 ; meanwhile, the is subjected to a spatial attention operation SA, to obtain an isolation switch feature map with a size of 1024*1024*3 ; then, the , , is subjected to a point multiplication operation, to obtain an isolation switch feature map with a size of 1024*1024*3 ; The is input to a triplet attention submodule Triplet Attention for attention enhancement, to obtain a feature map ; the is input to a state judgment output unit, to obtain an isolation switch state judgment result (for example, the judged state is open, and the confidence is 99%), and then the notification and alarm unit is used to transmit the isolation switch state to a mobile communication terminal through a telecom operator network, to realize remote monitoring of the isolation switch state.

[0038] Embodiment 2 The isolation switch state discrimination device provided in the embodiment of the application, as shown in Figure 8 , comprises: an isolation switch shooting unit 1, which is used for shooting an isolation switch to obtain an original isolation switch image data set; a state detection chip 2, which is used for detecting the original isolation switch image to judge the state of the isolation switch; the state detection chip is composed of an image preprocessing unit 21, an image feature recognition unit 22 and a state judgment unit 23; The notification and alarm unit 3 outputs the state of the disconnector according to the state determination result of the disconnector by the state detection chip. The energy supply unit 4 supplies power to the disconnector photographing unit, the state detection chip, and the notification and alarm unit.

Claims

1. A disconnector state recognition method, characterized in that, The application relates to an isolated switch state monitoring method and device. S1, constructing an isolated switch shooting unit to shoot an isolated switch and obtain an original isolated switch image data set; S2, constructing a state detection chip to detect the original isolated switch image and judge the state of the isolated switch; the state detection chip is composed of an image preprocessing unit, an image feature recognition unit and a state judgment unit; S3, constructing a notification and alarm unit, according to the state judgment result of the state detection chip on the isolated switch, triggering a local sound and light alarm in real time, or sending the isolated switch state information to a monitoring platform and a mobile operation terminal through wireless communication; S4, constructing an energy supply unit to supply power to the isolated switch shooting unit, the state detection chip and the notification and alarm unit.

2. The disconnector state recognition method of claim 1, characterized in that, The execution steps of the isolated switch shooting unit in S1 include: The isolated switch shooting unit includes a high-definition imaging collection device for collecting original isolated switch images in real time; the high-definition imaging collection device is used for image collection under multiple angles, multiple illuminations and multiple weather conditions; during shooting, automatic exposure, gain adjustment and light supplementing devices are used to ensure image definition and record shooting time, environmental parameters and camera information; original isolated switch images are collected to construct an isolated switch image data set.

3. The method of isolator state recognition of claim 1, wherein, The construction process of the state detection chip in S2 includes: S21, constructing an image preprocessing unit to process the original isolated switch image; first, an adaptive histogram equalization algorithm is used to improve the overall contrast of the image and enhance the details of the isolated switch and its background; then, a bilateral filtering method is used to remove background noise and texture interference while retaining edge contours; finally, image size normalization and color space conversion are performed to generate an isolated switch image with a uniform size; S22, constructing an image feature recognition unit to recognize the state features of the isolated switch image; the image feature recognition unit is composed of a parallel multi-scale channel attention fusion sub-module, a multi-resolution interactive enhancement convolution sub-module, a frequency domain enhancement channel-space fusion convolution sub-module and a triple attention sub-module; S23, constructing a state judgment unit to judge the state of the isolated switch according to the state features of the isolated switch.

4. The state detection chip of claim 3, wherein, The execution process of the image feature recognition unit in S22 includes: Step 211, input the isolation switch image to the parallel multi-scale channel attention fusion submodule, perform parallel multi-scale channel attention fusion operation, and obtain an isolation switch feature map ; Step 212, The input is fed into the multi-resolution interactive enhanced convolution submodule to perform multi-resolution interactive enhanced convolution operations, resulting in the isolating switch feature map. ; Step 213, The input is fed into the frequency domain enhancement channel-spatial fusion convolution submodule, where a frequency domain enhancement channel-spatial fusion convolution operation is performed to obtain the isolating switch feature map. Then The input is fed into the triple attention submodule for attention enhancement to obtain the feature map of the disconnector switch. .

5. The image feature recognition unit of claim 4, wherein, The execution steps of the parallel multi-scale channel attention fusion sub-module in S211 include: S2111. Let the input of the parallel multi-scale channel attention fusion submodule be a feature map of the isolation switch with size H×W×C. ,right A convolution operation with a kernel size of 5×5 and a stride of 1 is performed, followed by a normalization layer and a channel scaling submodule, to obtain a feature map of the isolation switch with dimensions H×W×C. At the same time, for Perform grouped convolution operations with a kernel size of 3×3 and a stride of 1, and then pass them through a batch normalization layer to obtain a feature map of the disconnector with dimensions H×W×C. Where H represents the height of the disconnector feature map, W represents the width of the disconnector feature map, and C represents the number of channels in the disconnector feature map; S2112, to and performing a channel splicing operation to obtain an isolation switch feature map with a size of HxWx2C ; performing a convolution operation on with a convolution kernel size of 1x1 and a step size of 1 to obtain an isolation switch feature map with a size of HxWxC ; performing a one-time adaptive average pooling operation on to obtain an isolation switch feature map with a size of HxWxC ; (1) (2) (3) wherein, denotes a convolution operation with a kernel size of 5x5 and a stride of 1; denotes a grouped convolution operation with a kernel size of 3x3 and a stride of 1; denotes a batch normalization layer; denotes a channel scaling module submodule; denotes a channel-wise concatenation operation; denotes a convolution operation with a kernel size of 1x1 and a stride of 1; denotes an adaptive average pooling operation; S2113, to passing through a full connection layer and a Gelu activation function in sequence, to obtain an isolation switch feature map with a size of HxWx1 ; and then passing through a full connection layer and a Sigmoid activation function in sequence, to obtain an isolation switch feature map with a size of HxWx1 ; S2114、will with dot product operation is performed to obtain the isolation switch feature map with the size of HxWxC : (4) (5) (6) wherein, denotes a fully connected layer; denotes an activation function; denotes an activation function; denotes a dot product operation.

6. The image feature recognition unit of claim 4, wherein, The execution steps of the multi-resolution interactive enhancement convolution sub-module in S212 include: S2121, the input of the isolation switch feature map with the size of HxWxC is taken as the input of the parallel multi-scale channel attention fusion sub-module , a convolution operation with the kernel size of 3x3 and the step of 1 is performed on the isolation switch feature map to obtain the isolation switch feature map with the size of HxWxC ; meanwhile, a convolution operation with the kernel size of 5x5 and the step of 1 is performed on the isolation switch feature map to obtain the isolation switch feature map with the size of HxWxC ;​​ S2122、to and perform channel concatenation operation to obtain a size of HxWx2C isolated switch feature map ; S2123、to a global max pooling operation is performed to obtain an isolation switch feature map with a size of 1x1x2C ; activated by a Swish activation function to obtain an isolation switch feature map with a size of 1x1x2C ; S2124、to a convolution operation with a convolution kernel size of 1x1 and a step of 1 is performed to obtain a size of 1x1xC isolation switch feature map ; S2125、will be described below. with point multiplication operation is performed to obtain the size of HxWxC isolation switch feature map : (7) (8) (9) (10) (11) (12) wherein, denotes a convolution operation with a kernel size of 3x3 and a stride of 1 ; denotes a convolution operation with a kernel size of 1x1 and a stride of 1 ; denotes a global max pooling operation; denotes an activation function.

7. The image feature recognition unit of claim 4, wherein, The execution steps of the frequency domain enhancement channel-space fusion convolution sub-module in S213 include: S2131. Let the input of the parallel multi-scale channel attention fusion submodule be a feature map of the isolation switch with size H×W×C. ;right Perform a convolution operation with a kernel size of 1×1 and a stride of 1 to obtain a feature map of the disconnector switch with dimensions H×W×C. ; S2132, to performing two-dimensional Fourier transform, extracting amplitude information of the frequency spectrum, and performing frequency domain feature enhancement through 1*1 convolution + BN + SiLU activation to obtain an isolation switch feature map with a size of H*W*C ; at the same time, performing a depth separable convolution operation with a convolution kernel size of 3*3 and a step size of 1 on ; to obtain an isolation switch feature map with a size of H*W*C ; S2133, to , perform channel concatenation operation to obtain the isolation switch feature map with the size of HxWx2C : (13) (14) ( ) (15) (16) wherein, denotes a two-dimensional Fourier transform operation; BN denotes a batch normalization operation; denotes an activation function; denotes a depthwise separable convolution operation with a kernel size of 3x3 and a stride of 1; S2134, to a convolution operation with a kernel size of 1x1 and a step of 1 is performed to obtain an isolation switch feature map with a size of HxWxC ; S2135、will After the channel attention operation, an isolated switch feature map with a size of HxWxC is obtained ; and simultaneously After the spatial attention operation, an isolated switch feature map with a size of HxWxC is obtained ; S2136, the , , point multiplication operation is performed to obtain the isolation switch feature map with the size of HxWxC : (17) (18) (19) (20) wherein, denotes a channel attention operation; denotes a spatial attention operation.

8. The state detection chip of claim 3, wherein, The execution process of the state judgment unit in S23 includes: The isolator feature map output by the image feature recognition unit The isolator state is judged by inputting the isolator feature map to a state judging unit The state judging unit is built-in a lightweight discrimination network, which analyzes and reasons through multi-layer convolution and full connection structure And sets two neurons in the output layer of the network, respectively outputting the open state probability and the closed state probability of the isolator , and When , the state of the isolator is judged as the open state, otherwise, the state of the isolator is judged as the closed state.

9. The method of isolator status recognition of claim 1, wherein, The execution steps of the notification and alarm unit in S3 include: The notification and alarm unit comprises a wireless communication module and a local alarm module, which are used to receive the state determination result of the isolating switch by the state detection chip, and trigger corresponding alarm or notification operation according to the preset threshold or abnormal rule: the wireless communication module is used to upload the isolating switch state and related environmental parameters to the monitoring platform or mobile operation and maintenance terminal in real time; the local alarm module comprises an audible and visual alarm device, which can immediately issue an alarm prompt when an abnormal state is detected; in the notification and alarm unit, a buffering and retransmission mechanism is built-in, which is used to enhance the real-time performance and reliability of the alarm information delivery in the case of network anomaly or communication delay.

10. A disconnector switch status determination device, characterized in that, Comprise: The isolating switch shooting unit shoots the isolating switch to obtain the original isolating switch image data set; The state detection chip detects the original isolating switch image to determine the state of the isolating switch; the state detection chip is composed of an image preprocessing unit, an image feature recognition unit and a state judgment unit; The notification and alarm unit outputs the state of the isolating switch according to the state determination result of the isolating switch by the state detection chip; The energy supply unit supplies power to the isolating switch shooting unit, the state detection chip and the notification and alarm unit.

Images