CN-114463733-B - Detection method, device and processing equipment for electrical protection pressing plate of transformer substation

CN114463733BCN 114463733 BCN114463733 BCN 114463733BCN-114463733-B

Abstract

The application provides a detection method, a detection device and a detection processing device for an electrical protection pressing plate of a transformer substation, which are used for executing individual identification of granularity of the electrical protection pressing plate in a dense electrical protection pressing plate cluster, so that an accurate detection object is provided for intelligent switching state detection of the electrical protection pressing plate, and further intelligent switching state detection of granularity of the electrical protection pressing plate can be completed in the electrical protection pressing plate cluster. The method comprises the steps of acquiring an actual image I acquired from an electrical protection pressure plate cabinet of a transformer substation, configuring different electrical protection pressure plates by the electrical protection pressure plate cabinet in a panel mode, and identifying an electrical protection pressure plate panel prediction mask from the actual image I by a semantic segmentation model M Semantic segmentation Electrical protection clamp plate panel prediction mask The method comprises the steps of dividing an actual image I into areas, indicating the areas where the panel of the electric protection pressing plate is located, and predicting a mask for the panel of the electric protection pressing plate according to dividing distances among different electric protection pressing plates Individual prediction mask divided into different electrical protection pressing plate examples

Inventors

LING FEI
YUAN XIN
CHEN XIAOJIAN
ZHOU MIAOLIN
LIU YANG

Assignees

广东数字生态科技有限责任公司

Dates

Publication Date: 20260505
Application Date: 20220209

Claims (9)

1. A method for detecting an electrical protection pressure plate of a transformer substation, the method comprising: Acquiring actual images acquired from an electrical protection platen cabinet of a substation Wherein, the electric protection pressing plate cabinet is provided with different electric protection pressing plates in a panel mode; Segmentation of models by semantics From the actual image In identifying an electrical protection platen panel predictive mask Wherein the electrical protection platen panel predicts a mask For indicating the actual image The area where the divided electric protection pressing plate panel is located; According to the dividing distance between different electric protection pressing plates, predicting mask by using the electric protection pressing plate panel Individual prediction mask divided into different electrical protection pressing plate examples ; The electric protection pressing plate panel is used for predicting a mask according to the dividing distance between different electric protection pressing plates Individual prediction mask divided into different electrical protection pressing plate examples Thereafter, the method further comprises: through the actual image Individual prediction mask for different electrical protection platen examples Image classification model The corresponding switching state of each electrical protection pressing plate example individual is identified ; Recognition of models by optical characters From the actual image The upper left corner predicted coordinate corresponding to each electric protection pressing plate label is identified Predicted coordinates of lower right corner Tag prediction text; predicting the left upper corner corresponding to each electric protection pressing plate label The lower right corner predicted coordinates And the label prediction text, the switching state corresponding to each electrical protection pressing plate instance unit Correlating to obtain electric protection pressing plate information; For the semantic segmentation model The coding and decoding structure design is used, and the method comprises the following steps: The coding structure consists of a 3-layer convolution module and is used for extracting image features; the decoding structure consists of a 3-layer up-sampling module and is used for converting the features into semantic segmentation masks; In the encoding structure and the decoding structure, the convolution module and the up-sampling module complete cross-module conduction of information through jump connection; The convolution module consists of 2 convolution layers and 1 average pooling layer, and is characterized in that A convolution kernel, 1 convolution step, wherein the tail end of the convolution layer uses a ReLU algorithm to perform nonlinear activation on output characteristics, and the average pooling layer Pooling convolution kernels, and 2 convolution steps; the convolution module is used for extracting high-dimensional characteristics and reducing the characteristic resolution so as to improve the model calculation efficiency; The up-sampling module fuses two input features, the input feature 1 passes through the up-sampling layer and then adjusts the feature resolution to be the same as the input feature 2, then the two features are superimposed pixel by pixel and pass through the two layers After the convolution kernel and the convolution layer with the convolution step of 1, obtaining an output characteristic of an up-sampling module, wherein the up-sampling module is used for representing the content information of the high-dimensional characteristic by using the low-dimensional characteristic; Classifying a model for the image The method comprises the following steps: the main structure is formed by stacking 4 layers of residual modules, and then connecting one layer A convolution kernel, a convolution layer with 1 convolution step and a global average pooling layer, which are used for converting image characteristics into image classification prediction categories, and a residual error module is used for extracting characteristics, and the structure is formed by The convolution kernel is formed by a cascade two-layer convolution layer with 1 convolution step, after feature extraction of the convolution layer, the input features and the extracted features are overlapped in a pixel-by-pixel mode by using jump connection to form output features; For the optical character recognition model The method comprises the following steps: the method comprises the steps of adopting a two-stage design, detecting a text region in an image by using a YOLOv model in the first stage, using a CRNN model in the second stage, using a frame text region in the first stage as input, and converting the image in the text region into a character string by performing sequence processing through a bidirectional LSTM.
2. The method of claim 1, wherein the passing the actual image Individual prediction mask for different electrical protection platen examples Image classification model The corresponding switching state of each electrical protection pressing plate example individual is identified Comprising: individual prediction mask through the different electrical protection platen instances From the actual image Dividing sub-graphs corresponding to the individual examples of the electric protection pressing plates ; Classifying a model by the image Identifying each sub-graph Corresponding to the switching state 。
3. The method of claim 1, wherein said predicting the upper left corner coordinates of each of said electrically protected platen labels The lower right corner predicted coordinates And the label prediction text, the switching state corresponding to each electrical protection pressing plate instance unit Performing association to obtain electrical protection platen information, including: Calculating individual prediction masks for different electrical protection platen instances Is defined by the pixel center point of (2) ; Calculating the left upper corner predicted coordinates corresponding to the electric protection pressure plate labels The lower right corner predicted coordinates To the center point of each pixel Average distance of (2) ; Using a nearest algorithm to predict the label prediction text corresponding to each electric protection pressing plate label and the switching state corresponding to each electric protection pressing plate instance unit And correlating to obtain the electrical protection pressing plate information.
4. The method according to claim 1, wherein the method further comprises: acquiring a sample image acquired from the electrical protection platen cabinet , The sample image is subjected to random shading adjustment, random dynamic blurring, random flipping and random rotation as data enhancement measures Content enhancement is carried out; to the sample image Corresponding sample electrical protection clamp plate panel prediction mask is annotated Individual prediction mask for sample electrical protection pressing plate example State of sample throwing and retreating Predicted coordinates of upper left corner of sample Predicted coordinates of lower right corner of sample And predicting text by the sample label, and then passing through the sample image The sample electrical protection platen panel prediction mask Training the semantic segmentation model Through the sample image Individual prediction mask for example of sample electrical protection pressing plate The sample put-back state Training the image classification model Through the sample image The upper left corner of the sample predicts the coordinates The lower right corner of the sample predicts the coordinates Training the optical character recognition model by the sample tag predictive text 。
5. The method according to any one of claims 1 to 4, wherein after obtaining the electrical protection platen information, the method further comprises: And archiving the electric protection pressing plate information on a system according to a standing book information organization mode.
6. The method of claim 1, wherein the electrical protection platen panel is predicted to mask at the dividing distance between the different electrical protection platens Individual prediction mask divided into different electrical protection pressing plate examples Previously, the method further comprises: the actual image is processed Converting into a gray image; And on the basis of the gray level image, calculating the dividing distance between the different electric protection pressing plates according to the transverse direction and the longitudinal direction by combining an adaptive threshold dividing strategy.
7. An electrical protection clamp plate detection device of a transformer substation, characterized in that the device comprises: An acquisition unit for acquiring an actual image acquired from an electrical protection platen cabinet of a substation Wherein, the electric protection pressing plate cabinet is provided with different electric protection pressing plates in a panel mode; A semantic segmentation unit for segmenting the model by semantic meaning From the actual image In identifying an electrical protection platen panel predictive mask Wherein the electrical protection platen panel predicts a mask For indicating the actual image The area where the divided electric protection pressing plate panel is located; Dividing unit for predicting mask of the panel of the electric protection pressing plate according to dividing distance between different electric protection pressing plates Individual prediction mask divided into different electrical protection pressing plate examples ; The apparatus further comprises: A switching state identification unit for passing the actual image Individual prediction mask for different electrical protection platen examples Image classification model The corresponding switching state of each electrical protection pressing plate example individual is identified ; An optical character recognition unit for recognizing the model by the optical character From the actual image The upper left corner predicted coordinate corresponding to each electric protection pressing plate label is identified Predicted coordinates of lower right corner Tag prediction text; A correlation unit for predicting the left upper corner coordinates corresponding to the electric protection pressure plate labels The lower right corner predicted coordinates And the label prediction text, the switching state corresponding to each electrical protection pressing plate instance unit Correlating to obtain electric protection pressing plate information; For the semantic segmentation model The coding and decoding structure design is used, and the method comprises the following steps: The coding structure consists of a 3-layer convolution module and is used for extracting image features; the decoding structure consists of a 3-layer up-sampling module and is used for converting the features into semantic segmentation masks; In the encoding structure and the decoding structure, the convolution module and the up-sampling module complete cross-module conduction of information through jump connection; The convolution module consists of 2 convolution layers and 1 average pooling layer, and is characterized in that A convolution kernel, 1 convolution step, wherein the tail end of the convolution layer uses a ReLU algorithm to perform nonlinear activation on output characteristics, and the average pooling layer Pooling convolution kernels, and 2 convolution steps; the convolution module is used for extracting high-dimensional characteristics and reducing the characteristic resolution so as to improve the model calculation efficiency; The up-sampling module fuses two input features, the input feature 1 passes through the up-sampling layer and then adjusts the feature resolution to be the same as the input feature 2, then the two features are superimposed pixel by pixel and pass through the two layers After the convolution kernel and the convolution layer with the convolution step of 1, obtaining an output characteristic of an up-sampling module, wherein the up-sampling module is used for representing the content information of the high-dimensional characteristic by using the low-dimensional characteristic; Classifying a model for the image The method comprises the following steps: the main structure is formed by stacking 4 layers of residual modules, and then connecting one layer A convolution kernel, a convolution layer with 1 convolution step and a global average pooling layer, which are used for converting image characteristics into image classification prediction categories, and a residual error module is used for extracting characteristics, and the structure is formed by The convolution kernel is formed by a cascade two-layer convolution layer with 1 convolution step, after feature extraction of the convolution layer, the input features and the extracted features are overlapped in a pixel-by-pixel mode by using jump connection to form output features; For the optical character recognition model The method comprises the following steps: the method comprises the steps of adopting a two-stage design, detecting a text region in an image by using a YOLOv model in the first stage, using a CRNN model in the second stage, using a frame text region in the first stage as input, and converting the image in the text region into a character string by performing sequence processing through a bidirectional LSTM.
8. A processing device comprising a processor and a memory, the memory having stored therein a computer program, the processor executing the method of any of claims 1 to 6 when invoking the computer program in the memory.
9. A computer readable storage medium storing a plurality of instructions adapted to be loaded by a processor to perform the method of any one of claims 1 to 6.

Description

Detection method, device and processing equipment for electrical protection pressing plate of transformer substation Technical Field The application relates to the field of transformer substations, in particular to a method and a device for detecting an electrical protection pressing plate of a transformer substation and processing equipment. Background Along with the continuous expansion of the power grid scale, the number of the electric protection pressing plates of the power transmission line related to the transformer substation is increased sharply, and the corresponding manpower inspection cost is also increased. In order to jump out the limitation of traditional manual inspection, prevent misoperation on the electric protection pressing plate and solve the problem of automatic management of the information of the electric protection pressing plate of the current transformer substation, the intelligent switching state detection method suitable for the electric protection pressing plate of the transformer substation is necessary. Today, intelligent state detection of an electrical protection pressing plate of a transformer substation can be divided into two types, namely a detection technology based on image feature matching and a detection technology based on neural network target detection. The method for detecting the switching state of the electric protection pressing plate based on image feature matching originates from the last century 80 and can be divided into two steps, namely, firstly, the feature point sampling technology is used for extracting the features of the electric protection pressing plate image, and then, the features are matched with the features in the prefabricated data set so as to judge the switching state of the electric protection pressing plate. However, the method is subject to a characteristic point sampling algorithm, and has extremely high requirements on the visual angle of a camera, the brightness and the like when the image of the electric protection pressing plate is shot. The method based on the neural network target detection adopts SSD, mask R-CNN, YOLO and other models as substrates, and uses a large-scale data set to train the models so as to realize the detection of the switching state of the electric protection pressing plate with good robustness. However, the target detection algorithm has good detection effect on the scene of discrete distribution of all the electric protection pressing plates, but the detection effect on the densely distributed electric protection pressing plates is not ideal. In addition, the current two types of electric protection pressing plate switching state detection methods do not deeply study the problem that the electric protection pressing plate corresponds to the standing book, and the automatic management requirement of corresponding information is difficult to meet. It can be found that the intelligent state detection scheme of the existing transformer substation electrical protection pressing plate has the problem of poor detection precision for application scenes of dense electrical protection pressing plates and poor environmental conditions in practical application. Disclosure of Invention The application provides a detection method, a detection device and a detection processing device for an electrical protection pressing plate of a transformer substation, which are used for executing individual identification of granularity of the electrical protection pressing plate in a dense electrical protection pressing plate cluster, so that an accurate detection object is provided for intelligent switching state detection of the electrical protection pressing plate, and further intelligent switching state detection of granularity of the electrical protection pressing plate can be completed in the electrical protection pressing plate cluster. In a first aspect, the application provides a method for detecting an electrical protection pressing plate of a transformer substation, which comprises the following steps: acquiring an actual image I acquired from an electrical protection pressing plate cabinet of a transformer substation, wherein the electrical protection pressing plate cabinet is configured with different electrical protection pressing plates in a panel mode; The prediction mask of the electric protection pressing plate panel is identified from the actual image I through the semantic segmentation model M Semantic segmentation Wherein, the electrical protection clamp plate panel predicts maskThe electric protection pressing plate panel is used for indicating the area where the electric protection pressing plate panel is located and is separated from the actual image I; according to the dividing distance between different electric protection pressing plates, predicting mask for electric protection pressing plate panel Individual prediction mask divided into different electrical protection pressing plate examples With refer