A load resource management system based on automatic generation of electrical topology

Abstract

The present invention relates to the technical field of automatic generation of electrical topology, and provides a load resource management system based on automatic generation of electrical topology, including: a convolutional model unit for training a convolutional neural network to generate an image recognition model for identifying a variety of graphic symbols and their terminals based on processed symbols, wherein the processed symbols are a variety of graphic symbols marking terminals; a picture recognition unit for preprocessing the acquired factory electrical wiring diagram, and inputting a variety of image recognition models to identify busbars, feeders, transformers and electrical equipment and their terminals, and generating an electrical topology structure diagram showing their connection relationship according to whether there are connections between the busbars, feeders, transformers and electrical equipment terminals. The present invention realizes the ledger-based and digital management of load resources, improves work quality and efficiency, and also performs target position detection and classification through the LeNet network model, improves model efficiency and simplifies model structure.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic generation of electrical topology, and in particular to a load resource management system based on automatic generation of electrical topology. Background Art

[0002] At present, the special inspection of load resources mainly adopts survey tools such as WeChat applet and load characteristics survey assistant. Since the user's electrical wiring diagram is only archived in the acceptance data, and the survey tool lacks the function of one-click generation of electrical topology structure, and lacks an interface that can systematically display the connection relationship between busbars and feeders and electrical equipment, the load information obtained from the survey, including security load, main production load, auxiliary production load, non-production load, main load-limiting equipment, and control mode (rigid / flexible), cannot be combined with the electrical wiring diagram. Figure 1 Correspondingly, it is not easy to carry out fine management of loads during the peak summer (winter) period. Moreover, due to the lack of visualization tools for electrical topology, the data of previous years' inspections have not been effectively accumulated, counted and analyzed, and thus it is impossible to carry out fine and normalized control of load resources through load resource management.

[0003] The Chinese patent document with application number 201811378384.5 only discloses the completion of inter-station topology identification through intra-station topology identification, plant and station information, and based on the transmission branch-node association matrix. It does not involve the topological structure identification of the electrical wiring diagram, and does not provide a method for load management based on the electrical topology diagram.

[0004] Therefore, how to generate an electrical topology diagram that integrates load information through an electrical wiring diagram to support the refined investigation and management of load resources, and how to improve the LeNet network model so that it can output accurate target location detection results while only outputting target classification results, are technical problems that need to be solved. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the defects of the above-mentioned prior art and provide a load resource management system based on automatic generation of electrical topology. The image recognition model generated by the convolutional neural network automatically converts the electrical wiring diagram into an electrical topology structure diagram that only displays the bus, feeder, transformer and electrical equipment, and annotates the load information to realize the ledger and digital management of load resources, improve work quality and efficiency, and provide strong technical support for the refined investigation and management of load resources. The LeNet network model based on the improved SCYLLA intersection-over-intersection loss function and the fusion attention mechanism is used to output accurate target position detection results based on the improved SCYLLA intersection-over-intersection loss function and the middle feature area of ​​the recognition target, and output accurate target classification results through the activation function, thereby improving the model efficiency of the LeNet network model, enhancing the generalization ability, simplifying the model structure, and realizing the easy deployment and maintenance of the LeNet network model in the load resource management system.

[0006] To achieve the above object, the present invention proposes a load resource management system based on automatic generation of electrical topology, including an electrical topology structure generation module and a management module;

[0007] The electrical topology structure generation module includes a convolution model unit and a picture recognition unit;

[0008] The convolution model unit is used to train a convolutional neural network to generate an image recognition model for identifying a plurality of graphic symbols and their connection terminals based on the processed symbols, wherein the processed symbols are a plurality of graphic symbols marking connection terminals;

[0009] The image recognition unit is used to pre-process the acquired electrical wiring diagram of the factory, and input a plurality of the image recognition models to recognize the busbar, feeder, transformer and electrical equipment and their terminals, and generate an electrical topology diagram showing the connection relationship between the busbar, feeder, transformer and electrical equipment terminals according to whether there are connections between them;

[0010] The management module matches the received load information with the electrical topology diagram it carries, and marks the load information at a corresponding position in the diagram;

[0011] The image recognition model is a LeNet network model based on an improved SCYLLA intersection-over-parameter loss function and a fusion attention mechanism, wherein the first branch of the output fully connected layer of the LeNet network model is set to output the improved SCYLLA intersection-over-parameter loss function for outputting a target position detection result, the second branch of the output fully connected layer is connected to an output activation function, the output activation function is used to output a target classification result, the attention mechanism is set between two convolutional pooling layers of the LeNet network model, the target position detection result is the position data of the terminal in the electrical wiring diagram, and the target classification result is the category of busbars, feeders, transformers and electrical equipment in the electrical wiring diagram;

[0012] The improved SCYLLA intersection-over-union loss function is determined by calculating the square sum of the intersection-over-union ratio and the distance penalty term, the shape penalty term and the angle penalty term;

[0013] The attention mechanism includes a channel attention mechanism and a spatial attention mechanism, wherein the channel attention mechanism is used to perform weighted processing on the feature map output by the convolutional pooling layer, and the spatial attention mechanism is used to make it focus on the terminals and the middle feature area within the graphic range where the busbars, feeders, transformers and electrical equipment in the electrical wiring diagram are located.

[0014] Furthermore, the distance penalty value of the distance penalty item is calculated and determined according to the difference between the coordinates of the center point of the real frame and the coordinates of the center point of the predicted frame;

[0015] The shape penalty term is calculated and determined according to the ratio of the difference between the size parameter of the real frame and the size parameter of the predicted frame and their maximum value;

[0016] The angle penalty term is determined by calculating the trigonometric function value of the angle between the coordinates of the center point of the real frame and the center point of the predicted frame.

[0017] Furthermore, the channel attention mechanism includes a first pooling layer, a first fully connected layer, a channel weight allocation layer and a first feature splicing layer;

[0018] The spatial attention mechanism includes a second pooling layer, a second fully connected layer, a spatial weight allocation layer and a second feature splicing layer.

[0019] Furthermore, the electrical topology structure generation module also includes a character recognition unit;

[0020] The character recognition unit extracts at least the names of the busbar, transformer and electrical equipment through the OCR character recognition model carried by it, and marks the names at corresponding positions in the electrical topology diagram;

[0021] The character recognition unit verifies the connection relationship between the transformer and the electrical equipment and the busbar according to the names thereof, and modifies the electrical topology diagram according to the connection relationship obtained by the character recognition unit.

[0022] Furthermore, the characteristic of the busbar identified by the image recognition unit is that there are multiple straight lines connecting them;

[0023] The character recognition unit extracts a bus name that conforms to the bus name rule from the recognized names, and judges whether the recognition of the bus is accurate according to whether the bus name is close to the recognized bus. If it is inaccurate, the straight line that is closest to the bus name and has multiple connecting lines is corrected as the bus;

[0024] The target classification result of the image recognition unit also includes the circuit breaker and the disconnector, and the target position detection result also includes the connection position data of the circuit breaker and the disconnector in the electrical wiring diagram. In the target classification result and the target position detection result, if both sides of the connection position data are respectively in contact with or overlap with the position data of the two bus sections, the two bus sections are determined to be the same bus;

[0025] The image recognition unit determines whether there is a double bus according to the positional relationship between the multiple busbars that are recognized, and if so, merges the double busbars into one graphic in the electrical topology diagram.

[0026] In the above scheme, the electrical topology diagram is further made concise and accurate, and the quality and efficiency of the troubleshooting and management work are improved.

[0027] Furthermore, the convolution model unit includes a sample expansion subunit, which pre-processes a plurality of graphic symbols in the electrical symbol library through a function to generate a sample data set including all processed symbols, and the sample data set is used to train a convolutional neural network to generate an image recognition model for identifying a plurality of graphic symbols;

[0028] The functions include a noise data function, a size change function, a rotation function, and a reduced integrity function.

[0029] In the above scheme, a sample data set is generated after preprocessing the electrical symbols through a function, which solves the problem that most electrical symbol graphic samples are single and it is difficult to provide a large number of data samples for training the convolutional neural network image recognition model through simple collection.

[0030] Further, the management module includes a load information acquisition unit and a level safety judgment unit;

[0031] The load information acquisition unit acquires the load information obtained through census statistics, and the load information includes the power load level of the electrical equipment; the level safety judgment unit judges whether the number of busbars connected to the electrical equipment and the feeder meets the requirements of the corresponding power load level, and if not, a risk warning is given in the electrical topology diagram.

[0032] Furthermore, the electrical topology structure generation module further includes an expansion unit, which increases the connection of the busbars in the electrical topology structure diagram according to the input feeder number, distribution room number and power supply number;

[0033] The management module also includes a capacity statistics unit, which counts the total maximum production load of all electrical equipment and all feeders connected to the busbar in the electrical topology diagram, and marks the total maximum production load at a corresponding position in the electrical topology diagram.

[0034] Furthermore, the convolution model unit also includes a sample preprocessing subunit, which is used to preprocess the acquired electrical wiring diagram of the factory through a color correction formula, and the color correction formula adjusts the color of the electrical wiring diagram according to the color mean, color variance and color saturation.

[0035] Furthermore, the electrical topology generation module and the management module are mounted on a power grid management cloud platform, and the power grid management cloud platform stores the load information, which also includes the type of the electrical equipment, the main limitable load and the control method of the electrical equipment, and the type is one of the main production load, auxiliary production load, and non-production load.

[0036] In the above scheme, the display content of the electrical topology diagram is further refined, which facilitates the staff to conduct detailed inspection and management of load resources.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The image recognition model generated by the convolutional neural network automatically converts the electrical wiring diagram into an electrical topology diagram that only displays the busbar, feeder, transformer and electrical equipment, and annotates the load information, realizing the ledger-based and digital management of load resources, improving the work quality and efficiency, and providing strong technical support for the refined investigation and management of load resources. In addition, the LeNet network model based on the improved SCYLLA intersection-over-union loss function and the fusion attention mechanism can output accurate target position detection results through the improved SCYLLA intersection-over-union loss function on the basis of focusing on the terminal blocks and the middle feature areas of the identified target, and output accurate target classification results through the activation function, thereby improving the model efficiency of the LeNet network model, enhancing the generalization ability, simplifying the model structure, and realizing the easy deployment and maintenance of the LeNet network model in the load resource management system.

[0039] 2. The electrical topology diagram is further made concise and accurate, improving the quality and efficiency of troubleshooting and management work.

[0040] 3. The sample data set is generated by preprocessing the electrical symbols through functions, which solves the problem that most electrical symbol graphic samples are single and it is difficult to provide a large number of data samples for training the convolutional neural network image recognition model through simple collection.

[0041] 4. The display content of the electrical topology diagram has been further refined, making it easier for staff to conduct detailed inspection and management of load resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of the general structure of a load resource management system automatically generated based on electrical topology according to an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of the general flow of a load resource management system automatically generated based on electrical topology according to an embodiment of the present invention;

[0044] Figure 3 A schematic diagram showing the effect of the electrical topology structure generation module recognizing graphic symbols according to an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of a process for identifying electrical equipment and wiring terminals according to an embodiment of the present invention;

[0046] Figure 5 It is a schematic diagram of a partial electrical topology structure diagram of an embodiment of the present invention;

[0047] Figure 6 Schematic diagram of the structure of the image recognition model of the embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0050] like Figures 1 to 6 As shown, the present invention provides a load resource management system based on automatic generation of electrical topology. Through an image recognition model generated by a convolutional neural network, the electrical wiring diagram is automatically converted into an electrical topology diagram that only displays busbars, feeders, transformers and electrical equipment, and load information is annotated, thereby realizing the ledger-based and digital management of load resources, improving work quality and efficiency, and providing strong technical support for the refined investigation and management of load resources. Moreover, through a LeNet network model based on an improved SCYLLA intersection-over-combination loss function and a fusion attention mechanism, the improved SCYLLA intersection-over-combination loss function is used to output accurate target position detection results on the basis of focusing on the terminal blocks and the middle feature areas of the identified target, and the accurate target classification results are output through the activation function, thereby improving the model efficiency of the LeNet network model, enhancing the generalization ability, simplifying the model structure, and realizing the easy deployment and maintenance of the LeNet network model in the load resource management system.

[0051] like Figures 1 to 6 As shown, this embodiment proposes a load resource management system based on automatic generation of electrical topology, including an electrical topology structure generation module and a management module; the electrical topology structure generation module includes a convolution model unit and an image recognition unit; the convolution model unit is used to train a convolution neural network to generate an image recognition model for identifying a variety of graphic symbols and their terminals based on processed symbols, wherein the processed symbols are a variety of graphic symbols marking terminal blocks; the image recognition unit is used to preprocess the acquired electrical wiring diagram of the factory, and input a variety of the image recognition models to identify the bus, feeder, transformer and electrical equipment and their terminals, and generate an electrical topology structure diagram showing their connection relationship according to whether there are connections between the terminals of the bus, feeder, transformer and electrical equipment; the management module matches the received load information with the electrical topology structure diagram it carries, and marks the load information at the corresponding position in the diagram.

[0052] like Figure 6As shown, the image recognition model is a LeNet network model based on an improved SCYLLA intersection-over-parameter loss function and a fusion attention mechanism, wherein the first branch of the output fully connected layer of the LeNet network model is set to output the improved SCYLLA intersection-over-parameter loss function for outputting a target position detection result, the second branch of the output fully connected layer is connected to an output activation function, the output activation function is used to output a target classification result, the attention mechanism is set between two convolutional pooling layers of the LeNet network model, the target position detection result is the position data of the terminal in the electrical wiring diagram, and the target classification result is the category of busbar, feeder, transformer and electrical equipment in the electrical wiring diagram;

[0053] The improved SCYLLA intersection-over-union loss function is determined by calculating the square sum of the intersection-over-union ratio and the distance penalty term, the shape penalty term and the angle penalty term;

[0054] The attention mechanism includes a channel attention mechanism and a spatial attention mechanism, wherein the channel attention mechanism is used to perform weighted processing on the feature map output by the convolutional pooling layer, and the spatial attention mechanism is used to make it focus on the terminals and the middle feature area within the graphic range where the busbars, feeders, transformers and electrical equipment in the electrical wiring diagram are located.

[0055] It should be noted that the electrical wiring diagram of the factory described in this embodiment is preferably a primary main wiring diagram of factories in typical industries such as textile industry, chemical fiber manufacturing industry, papermaking and products industry, and metal products industry. The primary main wiring diagram must include busbars, transformers and electrical equipment.

[0056] It should be noted that the load information in this embodiment is preferably obtained through the load resource management cloud platform of the application software, and the application software stores the factory's information over the years, including security load, main production load, auxiliary production load, non-production load, main load-limiting equipment, control mode (rigid / flexible) and electrical primary main wiring diagram. Therefore, the load information can be understood as the factory's inspection data directly obtained from i State Grid by setting a data collection tool (such as a website data crawler) or setting a program algorithm. Furthermore, the management module can also directly obtain the load information input by the staff through the application interface.

[0057] It should be noted that the LeNet network model, as a classic and efficient model in convolutional neural networks, has few model parameters and is better for identifying handwritten or unclear electrical wiring diagrams. However, the conventional LeNet network model is only used for image classification tasks and cannot identify a certain feature in the image. This embodiment sets a first branch and a second branch after its output fully connected layer, generates a prediction box in the first branch, and improves the SCYLLA intersection-over-union loss function so that the LeNet network model can perform target position detection, thereby avoiding repeated calculations and resource waste caused by performing target detection and target classification separately. And through the attention mechanism, the LeNet network model can learn richer image features, thereby enhancing the generalization ability of the model.

[0058] It is understandable that there are many types of image recognition algorithms, and the most widely used ones are BP neural network (BackPropagation, BP) and convolutional neural network (Convolutional Neural Networks). The layers in the BP model are all fully connected structures. When the specification of the recognition object is large, a huge amount of data will be generated, which will bring difficulties to computer processing. It takes several days or even months to train a network. For the same recognition task, the convolutional neural network (CNN) model data is greatly reduced compared with the BP model, so it has a wider application in image recognition technology. For basic electrical graphic symbols, image recognition is performed. The shape of the electrical symbol image is relatively simple, and the shape and lines are relatively fixed. The difference between electrical graphic symbols mainly lies in: image pixels, size ratio, background grayscale, number of noise points, image direction, image integrity, straightness and curvature of image lines, etc. Therefore, this embodiment can effectively improve the recognition accuracy of the convolutional neural network model for electrical graphic symbols through the application of convolutional neural network in small sample image recognition of electrical graphic symbols.

[0059] Specifically, the image recognition model is preferably a LeNet-5 network model, see Figure 6 , Figure 6 Conv in the figure is the convolution layer of the LeNet-5 network model, Pooling is the pooling layer, the second and third convolution layers are omitted between the first convolution layer Conv1 and the fourth convolution layer Conv4, Mc and Ms are the channel attention mechanism and spatial attention mechanism respectively, Fullconnection is the output full connection layer, Relu is the Relu activation function, Maxpooling is the maximum pooling layer, and Softmax is the Softmax activation function.

[0060] Furthermore, if Figure 4As shown, the image recognition model is preferably a LeNet-5 network model, which includes a local feature extraction model and a terminal extraction model; the picture recognition unit is used to input the preprocessed electrical wiring diagram into the local feature extraction model, and respectively identify the graphic ranges of the bus, feeder, transformer and electrical equipment; the picture recognition unit is used to input the graphic range into the terminal extraction model corresponding to the bus, feeder, transformer or electrical equipment, and identify the connection terminals within the graphic range.

[0061] Specifically, the working process of the convolutional neural network is divided into two processes, namely the feature extraction process and the classification process. The convolutional neural network self-learns feature extraction parameters by training labeled data. Compared with the manual design method, this feature extraction method is simpler and more popular. The output activation function of the classification process uses the Softmax function to solve the multi-classification problem in visual recognition, and ensures that the classification is specific to the specific probability value of each category, which has a good supporting role in subsequent processing. It should be noted that the local feature extraction model and the terminal extraction model can be a two-level series convolutional neural network.

[0062] More specifically, the image recognition model described in this embodiment preferably adopts a convolutional neural network of the LeNet-5 network model composed of multiple convolutional layers, multiple pooling layers and five fully connected layers. The convolutional neural network of the LeNet-5 network model requires fewer computing resources during training and testing, and at the same time adopts a combination of pooling layers and convolutional layers, which effectively reduces the size of the feature map and reduces the computational complexity of the model.

[0063] Specifically, an image preprocessing module is set before the image recognition model. The image preprocessing module is divided into two parts: picture preprocessing and One-Hot encoding. The picture preprocessing includes picture graying, picture binarization, and image denoising. Graying is to convert the three channels of R, G, and B of the sample image / electrical wiring diagram into a single-channel grayscale image to reduce the image dimension and simplify the subsequent processing steps; binarization is to convert the image into a binary image to improve the clarity of the edges of graphic symbols; image denoising is to remove image noise points that may interfere with recognition through filtering and other methods. One-Hot encoding is used to convert the image after picture preprocessing into a unified 32×32 image.

[0064] Specifically, the convolutional neural network of the LeNet-5 network model is used for training, testing and feature extraction, and its network structure is: input 32×32, convolution filter layer 6×28×28, pooling layer 2×2, fully connected layer 14×14, and output feature 10.

[0065] In the convolution filter layer, the step size is set to 1, and the number of layers is set to 6. In each convolution layer, 32 convolution kernels should be used for calculation to improve the model performance, and the Sigmoid activation method is used. In the convolution filter layer, the 4-channel image is converted to a 3-channel image, and the specific local data of the first channel is shown in the following table:

[0066]

[0067] The pooling layer uses spatial pyramid pooling. For example, after a 28×28 feature map is pooled with a 2×2 window, the size of the feature map is (28 / 2)×(28 / 2), that is, 14×14. In the fully connected layer, global fusion is performed to output the recognition features.

[0068] Some important parameters of the convolutional neural network of the LeNet-5 network model are set as follows: training round epochs100, round patience50 when the accuracy is not significantly improved, training batch size batch16, loss function categorical_crossentropy, optimizer SGD, and initial learning rate 0.01.

[0069] Furthermore, the channel attention mechanism includes a first pooling layer, a first fully connected layer, a channel weight allocation layer and a first feature splicing layer; the spatial attention mechanism includes a second pooling layer, a second fully connected layer, a spatial weight allocation layer and a second feature splicing layer.

[0070] Specifically, the first pooling layer performs global average pooling and global maximum pooling by the following formula:

[0071]

[0072] In the formula, is the average descriptor generated by global average pooling, H and W are the two dimensions of the input feature map, The largest descriptor generated by global maximum pooling;

[0073] Specifically, after the first descriptor and the second descriptor pass through the first fully connected layer, they are input into the channel weight allocation layer represented by the following formula:

[0074]

[0075] In the formula, is the channel attention weight, represents the Sigmoid activation function, ReLU represents the ReLU activation function, and are two learnable weights, is the average descriptor, is the maximum descriptor;

[0076] The channel weight vector is subjected to channel weighting by the first feature concatenation layer represented by the following formula to obtain a first enhanced feature map:

[0077]

[0078] In the formula, is the channel attention weight, F represents the input feature map, represents element-wise multiplication, enhancing a feature map for the channel;

[0079] The channel enhanced feature map is subjected to global average pooling and global maximum pooling through the second pooling layer expressed by the following formula:

[0080]

[0081] In the formula, is the global average descriptor generated by global average pooling, and C is the channel enhancement feature Figure 1 dimensions, The global maximum descriptor generated by global maximum pooling, is the first enhanced feature map;

[0082] After the global average descriptor and the global maximum descriptor pass through the second fully connected layer, the spatial weight allocation layer represented by the following formula performs feature concatenation, convolution layer and activation function operations to generate the spatial attention weight:

[0083]

[0084] In the formula, is the spatial attention weight, is the channel enhanced feature map, represents the Sigmoid activation function, Conv represents the convolutional layer, is the global average descriptor, is the global maximum descriptor, Represents feature concatenation in the channel dimension;

[0085] The spatial attention weight and the channel enhanced feature map are calculated by the second feature concatenation layer represented by the following formula to generate an output feature map:

[0086]

[0087] In the formula, represents the output feature map, is the channel space attention weight, Represents element-wise multiplication.

[0088] Furthermore, the convolution model unit also includes a sample preprocessing subunit, which is used to preprocess the acquired electrical wiring diagram of the factory through a color correction formula, and the color correction formula adjusts the color of the electrical wiring diagram according to the color mean, color variance and color saturation.

[0089] The color correction formula is specifically:

[0090]

[0091] Where c represents one of the three primary colors of RGB, F C represents the output image of the color correction formula, U represents the input electrical wiring diagram, M represents the color mean of the corresponding primary color c in the electrical wiring diagram U, V represents the color variance of the corresponding primary color c in the electrical wiring diagram U, and μ represents the color saturation of the electrical wiring diagram U. Since some electrical wiring diagrams are collected by taking photos with mobile phones and other devices, there are problems of large color deviation and low contrast. Therefore, the color correction formula can restore its normal color and contrast, thereby improving the quality of the input image and thus improving the performance of the entire network model.

[0092] Specifically, a convolution layer and a pooling layer for generating a prediction box are set in the first branch after the output fully connected layer. In the field of deep learning and computer vision, the ground truth box and the prediction box are two crucial concepts, which represent the actual target in the target detection task and the prediction result of the model respectively. The ground truth box is determined by manual annotation or other accurate methods, and contains key information such as the location, size and category of the target. The prediction box is a box of the possible area where the target is located obtained by calculation and reasoning by the output fully connected layer of the LeNet-5 network model, representing the area where the model believes the target is most likely to appear. The accuracy and effect of the prediction box are usually evaluated by comparing it with the actual ground truth box. Commonly used positioning loss functions include intersection over union (IOU) and the like. The intersection over union calculates the ratio of the intersection area to the union area between the prediction box and the ground truth box to evaluate the coverage of the prediction box to the ground truth box. By continuously optimizing the model, the accuracy and intersection over union value of the prediction box can be improved, guiding the model to learn how to accurately predict the location and size of the target, which helps the model better learn how to distinguish between the target and the background.

[0093] The traditional positioning loss function only considers the center point distance between the predicted box and the real box, the minimum bounding box, the aspect ratio and other conditions. The improved SCYLLA intersection-over-union (SIOU) of this embodiment takes the width and height into consideration separately as the loss function. In order to consider the problem of box matching direction, the distance information, shape information and angle information between the boxes are added on the basis of the conventional SCYLLA intersection-over-union loss function to improve the recognition and detection capabilities of partially obscured electrical symbols in the electrical primary main wiring diagram.

[0094] The improved SCYLLA intersection-over-union ratio loss function is calculated based on the intersection-over-union ratio and the sum of squares of the distance penalty term, the shape penalty term and the angle penalty term. The improved SCYLLA intersection-over-union ratio loss function is specifically:

[0095]

[0096] In the formula, is the SCYLLA intersection-over-union loss function, is the intersection-and-union ratio, which is calculated using the conventional intersection-and-union ratio formula, is the distance penalty term, is the shape penalty term, is the angle penalty term.

[0097] It should be noted that the conventional SCYLLA intersection-over-union loss function is:

[0098]

[0099] Therefore, compared with the conventional SCYLLA IoU loss function, the improved SCYLLA IoU loss function can obtain a difference value that more accurately reflects the penalty terms by dividing the sum of the squares of the distance penalty term, the shape penalty term, and the angle penalty term by the quantity 3, thereby enabling the predicted box to converge to the true box faster.

[0100] The distance penalty item calculates the penalty value of each coordinate in the real box and the predicted box by the following formula:

[0101]

[0102] In the formula, is the distance penalty term, e is a natural constant, is the coordinate distance penalty value, including the horizontal coordinate distance penalty value And the vertical axis distance penalty .

[0103] Furthermore, the distance penalty value of the distance penalty item is calculated and determined according to the difference between the coordinates of the center point of the real frame and the coordinates of the center point of the predicted frame, and the distance coordinate penalty value is calculated by the following formula:

[0104]

[0105] In the formula, is the horizontal axis distance penalty value, is the vertical coordinate distance penalty value, x and y are the horizontal and vertical coordinates respectively, and are the horizontal and vertical coordinates of the center point of the real frame, respectively. and are the horizontal and vertical coordinates of the prediction box.

[0106] Furthermore, the shape penalty term is calculated and determined according to the ratio of the difference between the size parameters of the real box and the size parameters of the predicted box to their maximum value. The shape penalty term calculates the penalty values ​​of the length and width of the real box and the predicted box by the following formula:

[0107]

[0108] In the formula, is the shape penalty term, e is a natural constant, is the shape penalty value, including the width shape penalty value and height shape penalty , is the adjustment coefficient.

[0109] The shape penalty item calculates the penalty value of the length and width of the real box and the predicted box by the following formula:

[0110]

[0111] In the formula, is the width shape penalty value, is the height shape penalty value, and are the width and height of the real frame respectively, and are the width and height of the prediction box respectively.

[0112] Furthermore, the angle penalty term is determined by calculating the trigonometric function value of the angle between the coordinates of the center point of the real frame and the center point of the predicted frame. The angle penalty term calculates the gap between the real frame and the predicted frame by the following formula, so that the predicted frame converges quickly to the real frame:

[0113]

[0114] In the formula, is the angle penalty term, and are the horizontal and vertical coordinates of the center point of the real frame, respectively. and is the horizontal and vertical coordinates of the predicted box. Therefore, the predicted box can converge to the x-axis or y-axis of the real box faster.

[0115] First, there are 50 neurons corresponding to 50 categories in the fully connected layer of the local feature extraction model. The fully connected layer compares the features of the extracted graphic symbols with samples or data sets, and then determines which category the graphic symbol belongs to among busbars, transformers, feeder terminals, circuit breakers, disconnectors and electrical equipment, as well as the specific types in the category, such as three-winding transformers and two-winding transformers. After that, the wiring terminals of the graphic symbol are extracted in the terminal extraction model corresponding to the type. Therefore, no matter how many wiring terminals the busbar or transformer has, they can be accurately identified and extracted, and the wiring terminals are extracted through the terminal extraction model corresponding to the type, which further reduces the calculation amount of the model and speeds up the processing speed of using convolutional neural networks to identify electrical wiring diagrams.

[0116] Furthermore, the electrical topology structure generation module also includes a character recognition unit; the character recognition unit extracts at least the names of the bus, transformer and electrical equipment through the OCR (Optical Character Recognition) character recognition model it carries, and marks the names at corresponding positions in the electrical topology structure diagram; the character recognition unit verifies the connection relationship between the transformer and the electrical equipment and the bus according to the names, and corrects the electrical topology structure diagram according to the connection relationship obtained by the character recognition unit.

[0117] Specifically, the OCR character recognition model may adopt a program of a Python resource library.

[0118] As a specific implementation of this embodiment, see Figure 3 The dotted line boxes in the figure are the characters extracted by the OCR character recognition model, and the dotted line boxes in the figure are the extracted busbars, feeders, transformers, circuit breakers, disconnectors and electrical equipment, and the corresponding wiring terminals are identified respectively, for example Figure 3 The display graph can be exported using the program of the Python resource library.

[0119] Furthermore, the image recognition unit recognizes a straight line with multiple connecting lines as a bus; the character recognition unit extracts the bus name that conforms to the bus name rule from the recognized names, and judges whether the recognition of the bus is accurate based on whether the bus name is close to the recognized bus; if it is inaccurate, the straight line with multiple connecting lines that is closest to the bus name is corrected to the bus.

[0120] Furthermore, the characteristic of the busbar identified by the image recognition unit is that there are multiple straight lines connecting them;

[0121] The character recognition unit extracts the bus name that conforms to the bus name rule from the recognized names, and determines whether the recognition of the bus is accurate based on whether the bus name is close to the recognized bus. If it is inaccurate, the straight line with multiple connecting lines that is closest to the bus name is corrected to the bus; the target classification result of the image recognition unit also includes circuit breakers and disconnectors, and the target position detection result also includes the connection position data of the circuit breaker and disconnector in the electrical wiring diagram. In the target classification result and the target position detection result, if the two sides of the connection position data are respectively in contact with or overlap with the position data of two bus segments, the two bus segments are determined to be the same bus.

[0122] It can be understood that the bus naming rules include bus X, bus X, bus No. X, #XM, WX, WLX, etc., where X is an arbitrary number.

[0123] Furthermore, the image recognition unit also identifies circuit breakers and disconnectors through the image recognition model. If it is determined that two adjacent busbars are only connected through circuit breakers and / or disconnectors, the multiple busbars are determined to be the same busbar; the image recognition unit determines whether it is a double busbar based on the positional relationship between the multiple busbars identified. If so, the double busbars are merged into one graphic in the electrical topology diagram.

[0124] It should be noted that since circuit breakers, disconnectors, etc. have no effect on the power load, they are not displayed in the electrical topology diagram. The image recognition model identifies circuit breakers and disconnectors, which are only used to assist in determining the segmented busbars, and to determine the connection relationship between the transformer, electrical equipment, and the busbar through the circuit breaker and disconnector. Specifically, if the image recognition unit determines that the transformer and electrical equipment are connected to the busbar through the circuit breaker and disconnector, the circuit breaker and disconnector are removed, and the wiring terminals at both ends are connected, and the connection relationship after the connection is generated into an electrical topology diagram.

[0125] Specifically, see Figure 3 ,Will Figure 3 W4 and W5 in the diagram are identified as a figure in the electrical topology diagram and named as busbar W4-W5. Figure 3 W1, W2 and W3 are identified as a figure in the electrical topology diagram and named as busbar W1-W2-W3.

[0126] Furthermore, the character recognition unit can also identify the names of circuit breakers and disconnectors, and correct the connection relationship between the circuit breaker and disconnector and the busbar according to the naming rules of their names. Specifically, for example, the disconnector connected to the busbar named 22M is 22M1. For a certain line, the names of all lines are based on the names of the circuit breakers. For example, if a 220kV line switch is named 233, then from the busbar to the line side, the disconnectors are named 2331, 2332, and 2333 in sequence, and the names of the grounding switches are named 2336C, 2336B, and 2336A from the busbar side to the line side. Therefore, by judging whether the names of multiple disconnectors are repeated in the first few digits and whether they are the same as the first few digits of the busbar name, it can be judged whether the connection between the circuit breaker and the disconnector and the busbar is correct.

[0127] In the above scheme, the electrical topology diagram is further made concise and accurate, and the quality and efficiency of the troubleshooting and management work are improved.

[0128] Furthermore, the convolution model unit includes a sample expansion subunit; the sample expansion subunit pre-processes a variety of graphic symbols in the electrical symbol library through a function to generate a sample data set including all processed symbols, and the sample data set is used to train a convolutional neural network to generate an image recognition model for identifying a variety of graphic symbols; the function includes a noise data function, a size change function, a rotation function and a completeness reduction function.

[0129] Specifically, Figure 4 As shown, this embodiment is based on the graphic symbols in the existing electrical symbol library, and uses the functions in the related resource library based on Python to expand the sample size. The functions include adding noise data, changing the original sample pixel size, changing the original sample direction position, changing the original sample integrity, adding Gaussian noise, adding salt and pepper noise, adding Poisson noise, etc.

[0130] In the above scheme, a sample data set is generated after preprocessing the electrical symbols through a function, which solves the problem that most electrical symbol graphic samples are single and it is difficult to provide a large number of data samples for training the convolutional neural network image recognition model through simple collection.

[0131] Furthermore, the management module includes a load information acquisition unit and a level safety judgment unit; the load information acquisition unit acquires the load information obtained through census statistics, and the load information includes the power load level of the electrical equipment; the level safety judgment unit judges whether the number of busbars connected to the electrical equipment and the feeder meets the requirements of the corresponding power load level, and if it does not meet the requirements, a risk warning is given in the electrical topology diagram.

[0132] Furthermore, the management module also includes a level safety recommendation unit; the sample data set includes the correspondence between the graphics of electrical equipment and the electrical types, and the output layer of the image recognition model is used to determine the electrical type corresponding to the graphics; the level safety recommendation unit generates a recommended safety level according to the electrical type, and if the recommended safety level is different from the power load level, the recommended safety level and risk label are marked at the corresponding position in the electrical topology diagram.

[0133] It should be noted that the power load level and its provisions described in this implementation are as follows: When one of the following conditions is met, it should be a first-level power load: 1. When power interruption will cause personal injury or death. 2. When power interruption will cause significant economic losses. For example: major equipment is damaged, major products are scrapped, a large number of products produced using important raw materials are scrapped, the continuous production process of key enterprises in the national economy is disrupted and it takes a long time to recover, etc. 3. Power interruption will affect the normal work of power users with significant economic significance. When one of the following conditions is met, it should be a second-level load: 1. When power interruption will cause significant economic losses. For example: major equipment is damaged, a large number of products are scrapped, the continuous production process is disrupted and it takes a long time to recover, and key enterprises have a large reduction in production, etc. 2. Power interruption will affect the normal work of important power users. Those that do not belong to the first and second level loads should be third-level loads, and short-term power outages are allowed for third-level loads. Based on the power load level, the basic requirements for power supply and distribution are as follows: Level 1 power load: It is required to use two independent power supplies, that is, the electrical equipment in the electrical topology diagram is connected to two different busbars, and the power supply names in the load information of the two busbars are different. Level 2 power load: It is required to use dual circuits or a dedicated line for power supply. Dual circuits are two lines for power supply (including working lines, standby and communication lines), that is, the electrical equipment in the electrical topology diagram is connected to two busbars, or the busbar is only connected to the electrical equipment. Level 3 power load: There are no special requirements for power supply.

[0134] Furthermore, the electrical topology structure generation module also includes an expansion unit; the expansion unit increases the connection of the busbars in the electrical topology structure diagram according to the input feeder quantity, distribution power room quantity and power supply quantity.

[0135] It should be noted that the number of feeders, the number of distribution rooms and the number of power sources can be obtained from the power grid management cloud platform. Since the power grid management cloud platform stores the load information checked in previous years, the expansion unit can generate the corresponding electrical topology diagram based on the different load information in previous years, thereby facilitating the staff's traceability management of load resource usage and providing assistance for the normalized management and control of load resources.

[0136] Furthermore, the management module also includes a capacity statistics unit; the capacity statistics unit counts the total maximum production load of all electrical equipment and all feeders connected to the busbar in the electrical topology diagram, and marks the total maximum production load at a corresponding position in the electrical topology diagram. Figure 5 The current maximum production load in .

[0137] Furthermore, the electrical topology generation module and the management module are mounted on a power grid management cloud platform, and the power grid management cloud platform stores the load information, which also includes the type of the electrical equipment, the main limitable load and the control method of the electrical equipment, and the type is one of the main production load, auxiliary production load, and non-production load.

[0138] Specifically, the power grid management cloud platform may be an application software for internal use, and the electrical topology diagram is planned to be displayed on the application software.

[0139] In the above scheme, the display content of the electrical topology diagram is further refined, which facilitates the staff to conduct detailed inspection and management of load resources.

[0140] In this embodiment, the image recognition model generated by the convolutional neural network automatically converts the electrical wiring diagram into an electrical topology diagram that only displays the busbar, feeder, transformer, and electrical equipment, and annotates the load information, thereby realizing the ledger-based and digital management of load resources, improving the work quality and efficiency, and providing strong technical support for the refined investigation and management of load resources. The conciseness and accuracy of the electrical topology diagram are further realized, and the quality and efficiency of the investigation and management work are improved. The sample data set is generated after the electrical symbols are preprocessed by the function, which solves the problem that most electrical symbol graphic samples are single and it is difficult to provide a large number of data samples for training the convolutional neural network image recognition model through simple collection. The display content of the electrical topology diagram is further refined, which facilitates the staff to conduct refined investigation and management of load resources.

[0141] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A load resource management system based on automatic generation of electrical topology, characterized in that: It includes electrical topology generation module and management module; The electrical topology structure generation module includes a convolution model unit and a picture recognition unit; The convolution model unit is used to train a convolutional neural network to generate an image recognition model for identifying a plurality of graphic symbols and their connection terminals based on the processed symbols, wherein the processed symbols are a plurality of graphic symbols marking connection terminals; The image recognition unit is used to pre-process the acquired electrical wiring diagram of the factory, and input a plurality of the image recognition models to recognize the busbar, feeder, transformer and electrical equipment and their terminals, and generate an electrical topology diagram showing the connection relationship between the busbar, feeder, transformer and electrical equipment terminals according to whether there are connections between them; The management module matches the received load information with the electrical topology diagram it carries, and marks the load information at a corresponding position in the electrical topology diagram; The image recognition model is a LeNet network model based on an improved SCYLLA intersection-over-parameter loss function and a fusion attention mechanism, wherein a fully connected network of an output layer of the LeNet network model connects a first branch and a second branch, the first branch sets a target position detection activation function, the target position detection activation function adopts an improved SCYLLA intersection-over-parameter loss function, the second branch sets a target classification activation function to output a target classification result, the attention mechanism is set between two convolutional pooling layers of the LeNet network model, the target position detection result is the position data of the terminal in the electrical wiring diagram, and the target classification result is the category of busbars, feeders, transformers and electrical equipment in the electrical wiring diagram; The improved SCYLLA intersection-over-union loss function is determined by calculating the square sum of the intersection-over-union ratio and the distance penalty term, the shape penalty term and the angle penalty term; The improved SCYLLA intersection-over-union loss function is specifically: In the formula, is the SCYLLA intersection-over-union loss function, is the intersection-and-union ratio, which is calculated using the conventional intersection-and-union ratio formula, is the distance penalty term, is the shape penalty term, is the angle penalty term; The distance penalty term is calculated by the following formula: In the formula, is the distance penalty term, e is a natural constant, is the coordinate distance penalty value, including the horizontal coordinate distance penalty value And the vertical axis distance penalty ; The distance coordinate penalty value is calculated by the following formula: In the formula, is the horizontal axis distance penalty value, is the vertical coordinate distance penalty value, x and y are the horizontal and vertical coordinates respectively, and are the horizontal and vertical coordinates of the center point of the real frame, respectively. and is the horizontal and vertical coordinates of the prediction box; The shape penalty term is calculated by the following formula: In the formula, is the shape penalty term, e is a natural constant, is the shape penalty value, including the width shape penalty value and height shape penalty , is the adjustment coefficient; The shape penalty value is calculated by the following formula: In the formula, is the width shape penalty value, is the height shape penalty value, and are the width and height of the real frame respectively, and are the width and height of the prediction box respectively; The attention mechanism includes a channel attention mechanism and a spatial attention mechanism, wherein the channel attention mechanism is used to perform weighted processing on the feature map output by the convolutional pooling layer, and the spatial attention mechanism is used to make it focus on the terminals and the middle feature area within the graphic range where the busbars, feeders, transformers and electrical equipment in the electrical wiring diagram are located.

2. The load resource management system based on automatic generation of electrical topology according to claim 1 is characterized in that: The distance penalty value of the distance penalty item is calculated and determined according to the difference between the coordinates of the center point of the real frame and the coordinates of the center point of the predicted frame; The shape penalty term is calculated and determined according to the ratio of the difference between the size parameter of the real frame and the size parameter of the predicted frame and their maximum value; The angle penalty term is determined by calculating the trigonometric function value of the angle between the coordinates of the center point of the real frame and the center point of the predicted frame.

3. The load resource management system based on automatic generation of electrical topology according to claim 1, characterized in that: The channel attention mechanism includes a first pooling layer, a first fully connected layer, a channel weight allocation layer and a first feature splicing layer; The spatial attention mechanism includes a second pooling layer, a second fully connected layer, a spatial weight allocation layer and a second feature splicing layer.

4. The load resource management system based on automatic generation of electrical topology according to claim 1, characterized in that: The electrical topology structure generation module also includes a character recognition unit; The character recognition unit extracts at least the names of the busbar, transformer and electrical equipment through the OCR character recognition model carried by it, and marks the names at corresponding positions in the electrical topology diagram; The character recognition unit verifies the connection relationship between the transformer and the electrical equipment and the busbar according to the names thereof, and modifies the electrical topology diagram according to the connection relationship.

5. The load resource management system based on automatic generation of electrical topology according to claim 4, characterized in that: The characteristic of the busbar identified by the image recognition unit is that there are multiple straight lines connecting them; The character recognition unit extracts a bus name that conforms to the bus name rule from the recognized names, and judges whether the recognition of the bus is accurate according to whether the bus name is close to the recognized bus. If it is inaccurate, the straight line that is closest to the bus name and has multiple connecting lines is corrected as the bus; The target classification result of the image recognition unit also includes the circuit breaker and the disconnector, and the target position detection result also includes the connection position data of the circuit breaker and the disconnector in the electrical wiring diagram. If the two sides of the connection position data in the target classification result and the target position detection result are respectively in contact with or overlap with the position data of the two bus sections, the two bus sections are determined to be the same bus; The image recognition unit determines whether there is a double bus according to the positional relationship between the identified multiple busbars, and if so, merges the double busbars into one graphic in the electrical topology diagram.

6. The load resource management system based on automatic generation of electrical topology according to any one of claims 1 to 5, characterized in that: The convolution model unit includes a sample expansion subunit, which pre-processes multiple graphic symbols in the electrical symbol library through a function to generate a sample data set including all processed symbols, and the sample data set is used to train a convolutional neural network to generate an image recognition model for identifying multiple graphic symbols; The functions include a noise data function, a size change function, a rotation function, and a reduced integrity function.

7. The load resource management system based on automatic generation of electrical topology according to claim 6, characterized in that: The management module includes a load information acquisition unit and a level safety judgment unit; The load information acquisition unit acquires the load information obtained through census statistics, wherein the load information includes the power load level of the electrical equipment; The level safety judgment unit judges whether the number of buses connected to the electrical equipment and the feeder meets the requirements of the corresponding power load level, and if not, a risk warning is given in the electrical topology diagram.

8. The load resource management system based on automatic generation of electrical topology according to claim 7, characterized in that: The electrical topology structure generation module further includes an expansion unit, which increases the connection of the busbars in the electrical topology structure diagram according to the input feeder number, distribution room number and power supply number; The management module also includes a capacity statistics unit, which counts the total maximum production load of all electrical equipment and all feeders connected to the busbar in the electrical topology diagram, and marks the total maximum production load at a corresponding position in the electrical topology diagram.

9. The load resource management system based on automatic generation of electrical topology according to claim 6, characterized in that: The convolution model unit also includes a sample preprocessing subunit, which is used to preprocess the acquired electrical wiring diagram of the factory through a color correction formula, and the color correction formula adjusts the color of the electrical wiring diagram according to the color mean, color variance and color saturation.

10. The load resource management system based on automatic generation of electrical topology according to claim 8, characterized in that: The electrical topology generation module and the management module are mounted on a power grid management cloud platform, which stores the load information. The load information also includes the type of electrical equipment, the main limitable load and control method of the electrical equipment, and the type is one of the main production load, auxiliary production load, and non-production load.

Images