CN-121982601-A - One-key sequential control knife switch state identification method and system based on ViT and GRU fusion

CN121982601ACN 121982601 ACN121982601 ACN 121982601ACN-121982601-A

Abstract

A one-key sequential control knife switch state identification method and system based on ViT and GRU fusion comprises the following steps of 1, obtaining a knife switch monitoring video, extracting continuous frame image sequences from the knife switch monitoring video, 2, extracting features of the multi-frame image sequences through a ViT model to obtain frame-level feature vectors, performing action clustering and phase division on the frame-level feature vectors to obtain an enhanced frame feature sequence, 3, inputting the enhanced feature sequence into a gating circulation unit network to obtain final time sequence features, 4, inputting the final time sequence features into a fully-connected classification layer, and processing and outputting state types of the knife switch through a normalization index function. According to the invention, the attention weighting mechanism is used for carrying out self-adaptive weighting fusion on the time characteristics, the rule constraint item is introduced, and abnormal weight suppression is carried out, so that the time sequence perceptibility and robustness of the model can be obviously improved while the detection precision is ensured.

Inventors

HAN MINGLEI
ZHANG LIQIANG
WEI JIAOLONG
ZHENG DANIAN
LI NAN
YANG ZITAO
YU JIMIN
LIU GANG
XU YANMING

Assignees

北京四方继保自动化股份有限公司
北京四方继保工程技术有限公司

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (10)

1. A one-key sequential control knife switch state identification method based on ViT and GRU fusion is characterized by comprising the following steps: step 1, acquiring a disconnecting link monitoring video, and extracting a continuous frame image sequence from the disconnecting link monitoring video; Step 2, extracting features of the frame image sequence through ViT models to obtain frame-level feature vectors, and performing action clustering and phase division on the frame-level feature vectors to obtain an enhanced frame feature sequence; Step 3, inputting the enhanced feature sequence into a gating circulation unit network to obtain a final time sequence feature; And step 4, inputting the final time sequence characteristics into a full-connection classification layer, and processing and outputting the state types of the disconnecting link through a normalized exponential function.
2. The method for identifying a one-button sequential control knife switch state based on ViT and GRU fusion according to claim 1, wherein, The step 1 specifically includes: Extracting continuous multi-frame images from a knife switch monitoring video, and carrying out size normalization processing on each frame of images to obtain processed frame images X t = [ B, T, C, H and W ], wherein B is the batch size of the images, C is the number of channels of each frame, and H and W are the height and width of the frame images respectively; And obtaining a continuous frame image sequence X, X= [ X 1 ,X 2 ,...,X T ] according to each processed frame image, wherein T is the total frame number, and T is [1, T ].
3. The method for identifying a one-button sequential control knife switch state based on ViT and GRU fusion according to claim 1, wherein, The step 2 specifically includes: feature extraction and image block embedding processing are carried out on each frame image in a frame image sequence by utilizing ViT model, and a t frame image is obtained Corresponding frame-level feature vectors : Wherein, the (. Cndot.) represents ViT model treatments.
4. The method for identifying a one-button sequential control knife switch state based on ViT and GRU fusion according to claim 1, wherein, The step of performing action clustering and phase division on the frame-level feature vector specifically comprises the following steps: acquiring the representation priori of the initial state and the terminal state according to the frame-level feature vector; And carrying out cluster analysis on the feature vector set of the whole action sequence based on the representation priori, and dividing the feature vector set into three types of state spaces in a self-adaptive manner, wherein the three types of state spaces correspond to an initial stage, an intermediate transition stage and a termination stage respectively.
5. The method for identifying a one-touch sequential control knife switch state based on ViT and GRU fusion according to claim 4, The step of obtaining the characterization priori of the initial state and the end state according to the frame-level feature vector specifically comprises the following steps: Calculating an average feature vector of the first n frames of the frame-level feature vectors And the average eigenvector of the next n frames : Wherein, T total frame number, Representing the frame-level feature vector corresponding to the i-th frame.
6. The method for identifying a one-touch sequential control knife switch state based on ViT and GRU fusion of claim 5, wherein, The feature vector set of the whole action sequence is subjected to cluster analysis based on the characterization priori, and the method specifically comprises the following steps: from the average eigenvector of the previous n frames And the average eigenvector of the next n frames Clustering action phases including an initial phase, an intermediate phase and a final phase to obtain a clustering result : Wherein, the 、、 Respectively representing an initial stage class, an intermediate stage class and a termination stage class; wherein, according to the relation between the cluster center and the anchor point distance, the semantics are automatically mapped to the average feature vector of the previous n frames The frame image with the smallest distance is classified as the initial stage and is combined with the average eigenvector of the following n frames The frame images with the smallest distance are classified as termination phases, and the rest frame images are classified as intermediate phases; calculating a difference vector for n frames And splice with the current frame vector to obtain spliced vector : Wherein, the Representing frame-level feature vectors corresponding to the t frame and the t+n frame; Assigning an action phase embedding vector to each type of action phase And fused with the frame features to obtain enhanced frame features : Wherein, the The expression vector is spliced to obtain the enhanced frame characteristics fused with the space information, the motion change and the motion phase semantic information As input for subsequent GRU timing modeling; wherein the motion phase embeds the vector The construction of (2) comprises: for the t frame image, for the label of the action phase Extract the first from the phase parameter matrix Lines, get embedded vectors : Wherein the matrix K is the number of categories of the action phase, Dimensions are embedded for the phases.
7. The method for identifying a one-touch sequential control knife switch state based on ViT and GRU fusion of claim 6, wherein, Further comprises: calculating the attention weight of the t-th frame The following are provided: wherein T is the total frame number, For the attention weighting parameter exp (·) is an exponential function, R (t) is a rule constraint function, Is an intermediate parameter; Combining the attention weight and the enhanced frame characteristics to obtain optimized enhanced frame characteristics : And taking the optimized enhanced characteristic as the input of the gating circulating unit network.
8. A one-key sequential control knife switch state identification system based on ViT and GRU fusion, for implementing the one-key sequential control knife switch state identification method based on ViT and GRU fusion according to any one of claims 1-7, comprising: the video acquisition module is used for acquiring a disconnecting link monitoring video; the frame extraction and preprocessing module is used for extracting a continuous frame image sequence from the disconnecting link monitoring video; The time sequence identification module is used for carrying out feature extraction on the frame image sequence through the ViT model to obtain a frame-level feature vector, and carrying out action clustering and phase division on the frame-level feature vector to obtain an enhanced frame feature sequence; the state classification and output module is used for inputting the final time sequence characteristics into the full-connection classification layer and processing the state classification of the output disconnecting link through the normalized exponential function.
9. A terminal comprises a processor and a storage medium, and is characterized in that: The storage medium is used for storing instructions; the processor is operative to perform the steps of the one-touch sequential knife switch status identification method of any one of claims 1-7 based on ViT and GRU fusion according to the instructions.
10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor performs the steps of the one-touch sequential control knife switch state identification method based on ViT and GRU fusion of any one of claims 1 to 7.

Description

One-key sequential control knife switch state identification method and system based on ViT and GRU fusion Technical Field The invention relates to the technical field of knife switch state identification, in particular to a one-key sequential control knife switch state identification method and system based on ViT and GRU fusion. Background Along with the continuous promotion of the intelligent level of the power grid, the automation and unattended operation of the transformer substation gradually become trend. In the operation of a transformer substation, the on-off state of a disconnecting switch (also called a disconnecting link) is one of the important factors affecting the safe operation of equipment. In order to improve the operation efficiency and the safety, a power grid dispatching system generally adopts a one-key sequential control (namely one-key sequential control) technology to realize automatic sequential control and state monitoring of equipment such as a breaker, a disconnecting switch, a grounding switch and the like. In the operation process of 'one-key sequential control', the system can automatically execute a series of opening and closing actions according to a preset operation ticket, and in order to ensure the action safety, the actual state of the isolating switch (disconnecting link) needs to be doubly confirmed: 1. the control system judges the on-off state according to the equipment signal (such as the contact quantity of the switch position) by one-time confirmation; 2. the secondary confirmation usually carries out recognition and verification on the physical position state of the field disconnecting link through a video image recognition or visual analysis means so as to prevent misjudgment caused by signal mistransmission or equipment jamming. The existing one-key sequential control knife switch identification method generally adopts the following procedures: And the video host or the algorithm server processes the acquired knife switch picture, identifies the actual on-off state of the knife switch, and returns the identification result to the dispatching system. The dispatching platform compares the returned result with the primary signal, if the returned result is consistent with the primary signal, the state of the disconnecting link is considered to be confirmed successfully, otherwise, the dispatching platform alarms to prompt or stops the sequential control operation. For identifying the opening and closing states of the disconnecting link, a plurality of disconnecting link identification methods are applied in the industry at present, and comprise the following steps: State recognition based on conventional image processing (e.g., edge detection, template matching); detecting visual identification of the network (e.g., YOLO, centerNet) based on deep learning; Based on the manner in which the data is fused. The prior art has at least the following drawbacks: Traditional single frame image recognition models (such as convolutional neural network CNN, vision Transformer ViT, etc.) can extract spatial features of the device in the still image. In practical application, the disconnecting link opening and closing processes have obvious time sequence continuity, and single-frame images are difficult to accurately reflect the movement trend of the disconnecting link. When the angle of the camera, the illumination condition or the shielding factor are changed, the situation that misjudgment or unstable identification is easy to occur on the basis of a single-frame identification model is faced to the situation that the knife switch is switched from 'on' to 'off' or from 'on' to 'off', and effective modeling of the time sequence relationship between frames is lacking, so that the identification accuracy of the 'false on' state and the 'false off' state is low. For example, the state recognition is performed through a single frame, and the accuracy is lower for the situation that the separation and the combination are not in place. Although partial researches try to improve the judging accuracy of the state of the knife switch by constructing a sequence model and predicting the action trend, the methods still have certain limitations in terms of feature expression precision, computational complexity or migration, and are not suitable for scenes with low cost, such as edge equipment. For example, in the prior art, a method for predicting a knife switch trend through modeling through information such as angle, speed and time exists, but the modeling of the method is complex, and because of different ageing degrees of different devices, the modeling of the devices needs to be corrected, so that the accuracy of the method in use in the different devices is greatly different. Disclosure of Invention In order to solve the defects in the prior art, the invention provides a one-key sequential control knife switch state identification method based on integration of a visual transducer a