A method and device for calculating the shift angle of an off-excitation tap changer

Abstract

The present invention proposes a method and device for calculating the shift angle of an off-circuit tap changer. The method comprises: obtaining vertical view image samples of the off-circuit tap changer, using the vertical view image samples to train a shift angle detection model; the shift angle detection model extracts features from the vertical view image, then uses a circular edge detection branch to detect the geometric parameters of each circular cylinder, and a slot center detection branch to detect the geometric parameters of each rectangular slot; finally, the angle between adjacent rectangular slots is calculated based on the geometric parameters of each circular cylinder and the geometric parameters of adjacent rectangular slots; the vertical view image of the off-circuit tap changer is obtained and input into the trained shift angle detection model to obtain a predicted angle between adjacent rectangular slots, which is used to calculate the shift angle of the off-circuit tap changer's switch paddle from an initial gear position to a target gear position. The present invention utilizes a dual-branch task structure to optimize the center and gear slots of the rotary switch to achieve high-precision detection.

Description

Technical Field

[0001] The present invention relates to image recognition technology, and in particular to a method and device for calculating a shift angle of an off-excitation tap changer. Background Art

[0002] With the accelerated construction of new power systems, medium and low voltage distribution networks are being fully transformed into active distribution networks. Power supply reliability and voltage quality have become core concerns in the power distribution field. To this end, power grid companies must monitor and adjust the transformer outlet voltage in real time. They need to make 2-6 adjustments each year based on seasonal, load and distributed power source changes, and implement dynamic management.

[0003] The current gear shifting operation process is cumbersome, involving power outages and grounding on the high-voltage and low-voltage sides, manual climbing to the distribution transformer installation platform, and manual gear shifting of the de-energized tap changer. During the gear shifting process, if the gear shifting angle deviation is too large, it will lead to poor contact of the joints inside the distribution transformer, which can easily cause the distribution transformer to burn out.

[0004] To address the shortcomings of manual shifting, such as poor precision and low efficiency, some current technical solutions propose robotic CNC shifting and apply artificial intelligence technologies such as YOLO to identify and calculate shifting angles. However, in the specific scenario of rotary switch detection, traditional methods have the following shortcomings:

[0005] Existing deep learning methods for predicting the angle of rotary switches typically employ traditional object detection strategies without specific optimization for angle prediction. These methods primarily focus on object location and classification, lacking specialized design for accurate prediction of rotation angles. While conventional object detection methods can detect the switch center and slot position when processing rotary switches, their angle prediction accuracy is limited. These methods typically first detect the position of a target box and then calculate the relevant angle information based on the target box's position. However, due to a lack of targeted optimization, their prediction errors can be large, making them unable to meet the requirements of high-precision industrial control.

[0006] like Figure 1 As shown, the off-circuit tap changer comprises a first circular cylinder 1, a second circular cylinder 2, a third circular cylinder 3, a fourth circular cylinder 4, a switch paddle 5, a center column clamp 6, a first paddle slot 7, a second paddle slot 8, a first paddle position indicator 9, and a second paddle position indicator 10. It includes a central component with distinct circular features and a rectangular position slot. Due to the differences in shape and size between these two components, a single detection strategy often fails to simultaneously address both characteristics, resulting in unstable detection results or significant errors. Conventional rectangular detection methods struggle to capture the edge features of circular features due to their symmetry, resulting in reduced positioning accuracy. Rectangular slots have straight edges and clear angles, so detection methods designed specifically for circular features may result in missed or false detections. Summary of the Invention

[0007] The technical problem to be solved by the present invention is: In response to the technical problems existing in the prior art, the present invention provides a method and device for calculating the gear shift angle of a non-excited tap changer, which adopts a dual-branch task structure and specifically optimizes the center and gear slot of the rotary switch to achieve high-precision detection.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A method for calculating a shift angle of an off-excitation tap changer comprises the following steps:

[0010] Obtain vertical view image samples of the de-energized tap changer and use the vertical view image samples to train a gear angle detection model. The gear angle detection model extracts features from the vertical view image to obtain a high-dimensional feature map. Then, a circular edge detection branch is used to detect the geometric parameters of each circular cylinder in the high-dimensional feature map. Simultaneously, a slot center detection branch is used to detect the geometric parameters of each rectangular slot in the high-dimensional feature map. Finally, the angle between adjacent rectangular slots is calculated based on the geometric parameters of each circular cylinder and the geometric parameters of adjacent rectangular slots.

[0011] A vertical perspective image of the off-circuit tap changer is obtained and input into a trained gear angle detection model to obtain the angle between adjacent rectangular slots. Based on the number of adjacent rectangular slots and the angle between adjacent rectangular slots between the initial and target gear positions of the off-circuit tap changer, the gear shift angle of the off-circuit tap changer from the initial gear position to the target gear position is calculated.

[0012] Furthermore, the geometric parameters of the circular cylinder include the center coordinates and radius of each circular cylinder of the off-excitation tap changer, and the geometric parameters of the rectangular slot include the center coordinates, width, height and direction angle of each rectangular slot of the off-excitation tap changer. The detection head of the circular edge detection branch and the detection head of the slot center detection branch are both regression networks composed of multiple fully connected layers.

[0013] Furthermore, when using vertical view image samples to train the angle detection model, the following steps are included: calculating the circle center loss, and adjusting the model parameters of the detection head of the circular edge detection branch according to the circle center loss. The circle center loss expression is as follows:

[0014]

[0015] in, and is the weight parameter, is the coordinate loss of the center of the cylinder, is the radius loss of the circular cylinder, is the geometric constraint loss of the circular cylinder, and its expression is as follows:

[0016]

[0017] in, is the angle between the center of the i-th circular cylinder and the line connecting the centers of the two rectangular slots, calculated based on the predicted coordinate values of the center of the i-th circular cylinder and the predicted coordinate values of the centers of the two adjacent rectangular slots. is the true value of the angle.

[0018] Furthermore, the center coordinate loss expression is as follows:

[0019]

[0020] in, are the predicted values of the horizontal and vertical coordinates of the center of the i-th circular cylinder, are the true values of the abscissa and ordinate of the center coordinates of the i-th circular cylinder, respectively, and N is the number of circular cylinders of the off-circuit tap changer;

[0021] The radius loss expression is as follows:

[0022]

[0023] in, is the predicted value of the radius of the i-th circular cylinder, is the true value of the radius of the i-th circular cylinder.

[0024] Furthermore, when using vertical view image samples to train the angle detection model, the following steps are included: calculating the slot loss, and adjusting the model parameters of the detection head of the slot center detection branch according to the slot loss. The slot loss expression is as follows:

[0025]

[0026] in, and is the weight parameter, is the center coordinate loss of the slot, is the geometric parameter loss of the slot, is the angular loss of the slot.

[0027] Furthermore, the center coordinate loss expression is as follows:

[0028]

[0029] in, are the predicted values of the horizontal and vertical coordinates of the center coordinates of the i-th slot, are the true values of the abscissa and ordinate of the center coordinate of the i-th slot, respectively, and N' is the number of slots of the off-circuit tap changer;

[0030] The geometric parameter loss expression is as follows:

[0031]

[0032] in, are the predicted values of the width and height of the i-th slot, and are the true values of the width and height of the i-th slot respectively;

[0033] The direction angle loss expression is as follows:

[0034]

[0035] in, is the predicted value of the direction angle of the i-th slot, is the true value of the direction angle of the i-th slot.

[0036] Furthermore, after detecting the geometric parameters of each circular cylinder in the high-dimensional feature map by the circular edge detection branch, the method further includes: calculating an average of the predicted values of the center coordinates of all the circular cylinders to obtain a predicted value of the center coordinates of the off-excitation tap changer; and calculating the angle between adjacent rectangular slots based on the geometric parameters of each circular cylinder and the geometric parameters of the adjacent rectangular slots, including:

[0037] Calculate the vectors of the predicted center coordinate value of the off-circuit tap changer and the predicted center coordinate value of each rectangular slot in the adjacent rectangular slots respectively;

[0038] The angle between the vectors corresponding to each rectangular slot in the adjacent rectangular slots is calculated to obtain the predicted value of the angle between the adjacent rectangular slots.

[0039] Furthermore, when the profile angle detection model extracts features from the vertical viewing angle image, specifically, the vertical viewing angle image is input into the convolutional neural network to extract common features in the image. When the profile angle detection model is trained using the vertical viewing angle image samples, the steps include: calculating the loss function of the profile angle detection model, and adjusting the model parameters of the convolutional neural network according to the loss function of the profile angle detection model. The loss function expression of the profile angle detection model is as follows:

[0040]

[0041] in, , and They are weight hyperparameters, is the slot loss, is the center loss, is the angle loss, and its expression is as follows:

[0042]

[0043] in, is the predicted value of the angle between adjacent rectangular slots, is the actual value of the angle between adjacent rectangular slots.

[0044] The present invention also provides a device for calculating a shift angle of an off-circuit tap changer, comprising a microprocessor and a computer-readable storage medium connected to each other, wherein the microprocessor is programmed or configured to execute any one of the methods for calculating a shift angle of an off-circuit tap changer.

[0045] The present invention further provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium. The computer program is programmed or configured to execute any one of the above-mentioned methods for calculating the shift angle of a de-energized tap changer through a microprocessor.

[0046] Compared with the prior art, the advantages of the present invention are:

[0047] The present invention uses two parallel branches to predict the central component and gear slot of a rotary switch respectively, and then fuses and processes the prediction results of the two branches to achieve high-precision prediction of the rotation angle of the rotary switch. The two branches are specifically optimized for the center and gear slot of the rotary switch, and detection heads are designed for circular and rectangular features respectively to ensure that the central component and gear slot of the rotary switch can be effectively identified simultaneously. In addition, an angle loss function is specially designed to solve the problem of insufficient angle prediction accuracy, ultimately achieving high-precision prediction of the rotation angle. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a structural diagram of the off-circuit tap changer.

[0049] Figure 2 The following is a brief flow chart of a method according to an embodiment of the present invention.

[0050] Figure 3 Schematic diagram of the gear angle detection model architecture according to an embodiment of the present invention.

[0051] Figure 4 This is a training flowchart of the circular edge detection branch according to an embodiment of the present invention.

[0052] Figure 5 This is a training flowchart of the slot center detection branch according to an embodiment of the present invention.

[0053] Figure 6 This is a training flowchart of the gear angle detection model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the scope of protection of the present invention is not limited thereby.

[0055] Example 1

[0056] This embodiment proposes a method for calculating the shift angle of a non-excited tap changer, specifically a detection method based on a dual-branch task structure, which is specifically optimized for the center and the shift slot of the rotary switch to achieve high-precision detection. Figure 2 As shown, the method of this embodiment includes the following steps:

[0057] Model training phase: vertical view image samples of the de-energized tap changer are obtained and used to train a gear angle detection model. The gear angle detection model extracts features from the vertical view image to obtain a high-dimensional feature map. The circular edge detection branch then detects the geometric parameters of each circular cylinder in the high-dimensional feature map. The slot center detection branch also detects the geometric parameters of each rectangular slot in the high-dimensional feature map. Finally, the angle between adjacent rectangular slots is calculated based on the geometric parameters of each circular cylinder and the geometric parameters of adjacent rectangular slots.

[0058] Model application stage: Obtain a vertical perspective image of the off-circuit tap changer and input it into the trained gear angle detection model to obtain the angle between adjacent rectangular slots. Based on the number of adjacent rectangular slots between the initial gear position and the target gear position of the off-circuit tap changer and the angle between adjacent rectangular slots, calculate the gear adjustment angle of the off-circuit tap changer from the initial gear position to the target gear position.

[0059] The gear angle detection model of this embodiment is as follows Figure 3 As shown in the figure, it consists of two parallel branches, which respectively predict the central component and gear slot of the off-excitation tap changer. The feature extraction modules of the two branches use a weight-sharing network. Finally, the prediction results of the two branches are fused and processed to achieve high-precision prediction of the rotation angle of the rotary switch. Figure 1 The working process of the off-circuit tap changer shown is as follows:

[0060] First, feature extraction is performed. The image acquisition unit obtains the image of the vertical viewing angle of the switch unit at the current moment. After normalization, resizing, and denoising, the vertical viewing angle image is pre-processed and then input into the convolutional neural network to extract common features in the image. In this embodiment, a convolutional neural network suitable for extracting common features is used as the backbone network of the shared feature extraction module. The backbone network can be any suitable convolutional neural network structure, such as ResNet, VGG, MobileNet, etc., or other network architectures with feature extraction capabilities. The backbone network extracts common features from the image through a series of operations such as convolutional layers, pooling layers, and activation functions. These features will be used for the tasks of the subsequent two branches.

[0061] Then, the simultaneous prediction of the two branches is performed, including the following steps:

[0062] S1) Using the circular edge detection branch, simultaneously detect the first circular cylinder 1, the second circular cylinder 2, the third circular cylinder 3, and the fourth circular cylinder 4, obtain the center coordinates and radius information of each circular cylinder, and calculate the average value based on the center coordinates of the four detected circular cylinders to determine the center coordinates of the off-circuit tap changer;

[0063] S2) Using the slot center detection branch, simultaneously identify the center coordinates of the first paddle slot 7 and the second paddle slot 8, as well as the width, height, and orientation angle information of the slots;

[0064] Finally, the prediction results of the two branches are combined to predict the angle between adjacent paddle slots, which includes the following steps:

[0065] S3) Combining the center coordinates of the off-circuit tap changer obtained in S1 and the geometric information of the first paddle slot 7 and the second paddle slot 8 obtained in S2, calculate the angle between the center points of the two slots and the line connecting the unit center of the off-circuit tap changer.

[0066] Based on the prediction results obtained by the gear angle detection model through the above process, during the model training process, the angle predicted by the model is compared with the manually labeled angle data, and the loss function is calculated to optimize the gear angle detection model's prediction ability for the angle, and finally the prediction of the rotation angle required for the switch paddle 5 to switch from the current position to the first paddle slot 7 and the second paddle slot 8 is completed, realizing the entire end-to-end angle prediction process.

[0067] The relevant content is further explained below.

[0068] In this embodiment, the geometric parameters of the circular cylinder include the center coordinates and radius of each circular cylinder of the off-circuit tap changer. When predicting the circular geometric parameters in the circular edge detection branch, a detection head consisting of three sequentially connected fully connected layers (referred to as the first detection head in this embodiment for distinction) is used to output information such as the center coordinates and radius of the circular cylinder of the central component, as well as the category information of the central component. The first detection head receives as input the high-dimensional feature map extracted by the convolutional neural network. The calculation process of each fully connected layer in the first detection head is as follows:

[0069]

[0070] in, is the layer index, For the The weight matrix of the layer, is the bias vector, is the activation function, That is, convolutional neural networks extract high-dimensional feature maps of features.

[0071] The last layer of the first detection head is a 12-dimensional fully connected layer that does not use an activation function and is used to regress and output the parameter information of each detected circular cylinder, including the center coordinates. and radius , so in step S1, the final output of the circular edge detection branch is:

[0072]

[0073] in, are the predicted values of the horizontal and vertical coordinates of the center of the first circular cylinder 1, is the predicted value of the radius of the first circular cylinder 1, are the predicted values of the horizontal and vertical coordinates of the center of the second cylindrical surface 2, is the predicted value of the radius of the second circular cylinder 2, are the predicted values of the abscissa and ordinate of the center coordinates of the third cylindrical surface 3, is the predicted value of the radius of the third circular cylinder 3, are the predicted values of the horizontal and vertical coordinates of the center of the fourth cylindrical surface 4, is the predicted value of the radius of the fourth circular cylinder 4.

[0074] After the circular cylinder detection is completed, the center coordinates of each circular cylinder detected are combined , calculate the coordinates of the center of the switch unit as a whole. Based on the center coordinates of the four detected circular cylinders, the coordinates of the center of the off-circuit tap changer can be calculated using the average value formula:

[0075]

[0076] in, and They represent the predicted values of the abscissa and ordinate of the center of the de-energized tap changer circle, respectively.

[0077] In this embodiment, the slot center detection branch is responsible for detecting the geometric parameters of each rectangular slot of the de-energized tap changer in the multi-dimensional feature image, based on the straight edge and angle features of the rectangular slot. In step S2, the slot center detection branch outputs the following results:

[0078] Center point coordinates of the first paddle slot 7 and the second paddle slot 8 and , used to describe the specific position of the slot center in the image;

[0079] The width of the first paddle slot 7 and the second paddle slot 8 and height , respectively represent the size of the slot along the horizontal and vertical directions;

[0080] Direction angles of the first paddle slot 7 and the second paddle slot 8 , which is used to represent the rotation angle of the slot in the image (that is, the degree of inclination of the long axis of the slot relative to the horizontal direction).

[0081] Therefore, the detection head of the slot center detection branch (referred to as the second detection head in this embodiment for distinction) is similar to the first detection head, and is also a regression network composed of multiple fully connected layers connected in sequence. The last layer of the second detection head is a 10-dimensional fully connected layer to predict the above-mentioned geometric parameters of the first paddle slot 7 and the second paddle slot 8.

[0082] After obtaining the off-circuit tap changer's center parameters through the circular edge detection branch and the slot geometry parameters through the slot center detection branch, the rotation angle between the slot center and the off-circuit tap changer's center can be calculated. This is accomplished in step S3 by defining a vector from the circle center to the slot center and analyzing its characteristics.

[0083] Specifically, when calculating the angle between adjacent rectangular slots based on the geometric parameters of each circular cylinder and the geometric parameters of adjacent rectangular slots, Figure 1 The center coordinates of the adjacent first paddle slot 7 and the second paddle slot 8 are and , the coordinates of the center of the circle of the off-circuit tap changer are Based on these coordinates, the vectors of the predicted center coordinate value of the off-circuit tap changer and the predicted center coordinate value of each rectangular slot in the adjacent rectangular slots can be calculated respectively. In this embodiment, the following two vectors are defined:

[0084] Center of circle Vector to first paddle slot 7 , the two components of this vector represent the distances from the center of the circle to the first paddle slot 7 in the horizontal and vertical directions respectively.

[0085] Center of circle Vector to second paddle slot 8 Similarly, this vector describes the direction and distance from the center of the circle to the second paddle slot 8. Vector and These are the core data that describe the spatial relationship between the slot center and the circle center. They not only provide the positional relationship of the slot relative to the circle center, but also lay the foundation for subsequent angle calculations.

[0086] The angle between adjacent rectangular slots can be calculated based on the angle between the vectors defined for each rectangular slot. The angle between the first paddle slot 7 and the second paddle slot 8 is calculated by the vector and The angle between vectors can be calculated using the dot product formula:

[0087]

[0088] in: is the dot product of the vectors. The modulus of the vector , are the lengths of the two vectors. Substituting the dot product and the modulus into the formula, the final angle can be calculated using the inverse cosine function:

[0089]

[0090] Calculated angle It is used to quantify the relative spatial position relationship between the first paddle slot 7 and the second paddle slot 8 , and provide a mathematical basis for the subsequent control action of the switch paddle 5 .

[0091] Based on the workflow of the profile angle detection model, this embodiment performs supervised learning training on the two branches of the profile angle detection model during the model training phase. Simultaneously, the convolutional neural network used for feature extraction in the profile angle detection model is trained. Specifically, when training the profile angle detection model using vertical viewing angle image samples, the following steps are included:

[0092] (1) Circular edge detection branch training

[0093] like Figure 4As shown, during the training process, after obtaining the geometric parameters of the circular cylinder, such as the center coordinates and radius, the center loss is calculated according to the geometric parameters of the circular cylinder, and the model parameters of the first detection head of the circular edge detection branch are adjusted according to the center loss.

[0094] In the circular edge detection branch, in order to improve the detection accuracy of the model in practical applications, this embodiment performs supervised learning training on the circular edge detection branch of the gear angle detection model. , the target value is the real cylinder center coordinate . The detection error is calculated using the L2 loss function:

[0095]

[0096] in, are the predicted values of the horizontal and vertical coordinates of the center of the i-th circular cylinder, are the true values of the horizontal and vertical coordinates of the center of the i-th circular cylinder, respectively. N is the number of circular cylinders of the non-excited tap changer to be detected. In this embodiment, = 4. By minimizing the coordinate loss of the center of the cylinder , the model can predict the center position of the cylinder more accurately.

[0097] For each cylinder radius , the target value is the true radius The radius error is also calculated using the L2 loss function:

[0098]

[0099] in, is the radius loss of the circular cylinder, is the predicted value of the radius of the i-th circular cylinder, is the true value of the radius of the i-th circular cylinder. The radius loss calculation helps the model accurately identify the geometric size of the circular cylinder.

[0100] In addition to the center coordinate loss and radius loss of the circular cylinder, in this embodiment, the circular edge detection branch also introduces a geometric constraint term in the loss function This constraint is mainly used to enhance the model's ability to learn directional information and ensure that the predicted angle is consistent with the directionality of the circular geometric characteristics. The geometric constraint measures the deviation between the predicted angle and the true angle and uses the cosine function as an error measure to ensure the accuracy of the direction prediction. Compared to the Euclidean distance, the cosine function has periodic characteristics, making this method more robust in directional learning. The geometric constraint loss function is in the form of:

[0101]

[0102] in, is the angle between the center of the i-th circular cylinder and the line connecting the centers of the two rectangular slots, calculated based on the predicted coordinate values of the center of the i-th circular cylinder and the predicted coordinate values of the centers of the two adjacent rectangular slots. is the true value of the angle manually marked. , constraining the model’s output to have consistent angular directionality, ensuring it accurately reflects the circle’s true geometric properties. Geometric constraints directly incorporate the circle’s symmetry and directional characteristics into the optimization objective, effectively guiding model learning.

[0103] In this embodiment, during the circular edge detection branch training phase, the geometric constraint item As part of the overall loss, it is jointly optimized with the traditional center point and radius errors:

[0104]

[0105] in, and is a weight parameter used to balance the importance of the two losses. During training, by adjusting the weight value, the model's performance can be dynamically optimized according to the task requirements. The geometric constraint term effectively reduces the error in predicting circle parameters and improves detection performance in complex scenes.

[0106] (2) Slot center detection branch training

[0107] like Figure 5 As shown, during the training process, after obtaining the geometric parameters of the paddle slot, such as the center coordinates, width, height, and direction angle, the slot loss is calculated based on the geometric parameters of the paddle slot, and the model parameters of the second detection head of the slot center detection branch are adjusted based on the slot loss.

[0108] The training goal of the slot center detection branch is to allow the model to continuously improve the prediction accuracy of the slot geometric parameters by optimizing the difference between the predicted value and the true labeled value. The center coordinate loss measures the error between the slot center point predicted by the model and the true labeled center point. By calculating the Euclidean squared error between the two points, the accuracy of the slot center coordinate prediction can be effectively quantified. The slot center coordinate loss function The expression is:

[0109]

[0110] in, are the predicted values of the horizontal and vertical coordinates of the center coordinates of the i-th slot, are the true values of the horizontal and vertical coordinates of the center coordinates of the i-th slot, is the number of slots in the de-energized tap changer. This part of the loss focuses on the positioning accuracy of the slots, ensuring that the model can accurately predict the location of the slots in the image.

[0111] Slot width and height are the key parameters that determine the slot shape, and the prediction errors of these two parameters are calculated through the geometric parameter loss optimization:

[0112]

[0113] in, are the predicted values of the width and height of the i-th slot, and are the true values of the width and height of the i-th slot, respectively. The error of geometric parameters directly affects the shape description accuracy of the slot, which further affects the accuracy of the angle calculation.

[0114] Direction It is an important parameter to describe the slot orientation, which can reflect the rotation angle of the slot in space. The error of the orientation angle is calculated by the following loss function calculate:

[0115]

[0116] in, is the predicted value of the direction angle of the i-th slot, is the true value of the orientation angle of the i-th slot. The optimization of the orientation angle error ensures the prediction accuracy of the slot orientation and provides support for further analysis of the spatial relationship between the slot center points.

[0117] Final loss function of the slot center detection branch It is defined as the weighted sum of each sub-loss function:

[0118]

[0119] in, and is a weight coefficient used to adjust the contribution of each subtask to the total loss. These weights are usually adjusted based on the importance of the task and its impact on the overall performance.

[0120] (3) Feature extraction module training

[0121] like Figure 6As shown in the figure, during the training process, after obtaining the angle between adjacent paddle slots, the angle loss function is calculated based on the angle prediction value and the actual value manually labeled. Then, the model parameters of the convolutional neural network of the feature extraction module are adjusted according to the center loss, slot loss and angle loss functions.

[0122] In this embodiment, the formula The angle between the first paddle slot 7 and the second paddle slot 8 is calculated, and the angle loss is designed. To further improve the model's sensitivity to the relative position of the detection frame, the calculation formula is as follows:

[0123]

[0124] in, is the predicted value of the angle between adjacent rectangular slots, is the actual value of the angle between adjacent rectangular slots, which is the manually labeled angle. This loss term focuses on the prediction accuracy of the angle, ensuring that the model can accurately capture the spatial relationship between the two slots.

[0125] The feature extraction module of the gear angle detection model adopts slot loss , center loss and angle loss Perform hybrid training to ensure that the features perform well in both slot and circle center detection tasks. It is mainly used to optimize the detection accuracy of the center of the switch unit, while the slot loss Focus on the prediction of slot-related parameters. This joint training strategy enables the model to achieve a good balance between global and local features, thereby improving the overall accuracy of angle calculation. Taking into account the detection tasks of slots and circle centers, the total loss function is defined in this embodiment Slot loss , center loss and angle loss The weighted sum of:

[0126]

[0127] in, , and Slot loss , center loss and angle loss The weight hyperparameter is used to control the impact of each component loss on the overall training. Through joint training, the model can simultaneously optimize the detection performance of slots and circle centers, achieving an effective combination of global and local features.

[0128] By optimizing the total loss function The gear angle detection model can accurately predict the geometric parameters of the slot and its spatial relationship with the center of the circle, thereby accurately calculating Figure 1 The included angle between the adjacent first paddle slot 7 and the second paddle slot 8.

[0129] During the model application phase, given the prior condition that the angle between each pair of slots is equal, the angle between the rectangular slots corresponding to a pair of adjacent gear positions (e.g., gears 2 and 3) can be calculated to deduce the angles of other adjacent gear positions (e.g., gears 1 to 2). To calculate the shift angle of the off-circuit tapchanger's switch paddle from the initial gear position to the target gear position, the number of adjacent rectangular slots and the angle between adjacent rectangular slots are used to identify the initial gear position of the switch paddle. Methods such as YOLO can then be used to determine the number of adjacent rectangular slots between the initial and target gear positions based on the difference between the initial and target gear positions. Finally, the shift angle is calculated by multiplying the angle between the detected adjacent rectangular slots and the number of adjacent rectangular slots.

[0130] For example, it is recognized that the initial gear position of the switch paddle is first gear and the target gear position is third gear. Since the gear positions correspond one-to-one to the rectangular slots, there are two pairs of adjacent rectangular slots between the initial gear position (first gear) and the target gear position (third gear) (first gear to second gear, second gear to third gear). At the same time, the gear angle detection model recognizes that the angle between a pair of adjacent rectangular slots is 15°, so the gear shift angle is 15°×2=30°.

[0131] It can be seen that the angle data identified by the gear angle detection model can be directly used to guide the paddle control action and achieve precise gear adjustment of the rotary switch.

[0132] Example 2

[0133] This embodiment provides a device for calculating a shift angle of a de-energized tap changer, comprising a microprocessor and a computer-readable storage medium connected to each other. The microprocessor is programmed or configured to execute the method for calculating a shift angle of a de-energized tap changer described in the first embodiment.

[0134] This embodiment further provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium. The computer program is programmed or configured to execute the method for calculating the shift angle of the de-energized tap changer according to the first embodiment through a microprocessor.

[0135] In summary, the present invention proposes a method and device for calculating the shift angle of a non-excited tap changer, which meets the requirements of robot numerical control shifting and applies artificial intelligence technology to identify and calculate the shift angle. Based on a dual-branch task structure, the present invention specifically optimizes the center and gear slot of the rotary switch to achieve high-precision detection. To address the problem of insufficient angle prediction accuracy, the present invention specifically designs an angle loss function to enable the model to more accurately predict the rotation angle. To address the shape matching problem, by designing detection heads for circular and rectangular features respectively, it is ensured that the model can effectively identify the center component and gear slot of the rotary switch at the same time.

[0136] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed above with reference to the preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent variations, and modifications to the above embodiment that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for calculating the shift angle of an off-excitation tap changer, characterized in that: The following steps are involved: Obtain vertical view image samples of the de-energized tap changer and use the vertical view image samples to train a gear angle detection model. The gear angle detection model extracts features from the vertical view image to obtain a high-dimensional feature map. Then, a circular edge detection branch is used to detect the geometric parameters of each circular cylinder in the high-dimensional feature map. Simultaneously, a slot center detection branch is used to detect the geometric parameters of each rectangular slot in the high-dimensional feature map. Finally, the angle between adjacent rectangular slots is calculated based on the geometric parameters of each circular cylinder and the geometric parameters of adjacent rectangular slots. A vertical perspective image of the off-circuit tap changer is obtained and input into a trained gear angle detection model to obtain the angle between adjacent rectangular slots. Based on the number of adjacent rectangular slots and the angle between adjacent rectangular slots between the initial and target gear positions of the off-circuit tap changer, the gear shift angle of the off-circuit tap changer from the initial gear position to the target gear position is calculated.

2. The method for calculating the shift angle of a non-excited tap changer according to claim 1, characterized in that: The geometric parameters of the circular cylinder include the center coordinates and radius of each circular cylinder of the off-circuit tap changer, and the geometric parameters of the rectangular slot include the center coordinates, width, height and orientation angle of each rectangular slot of the off-circuit tap changer. The detection head of the circular edge detection branch and the detection head of the slot center detection branch are both regression networks composed of multiple fully connected layers.

3. The method for calculating the shift angle of a non-excited tap changer according to claim 2, characterized in that: When using vertical view image samples to train the angle detection model, it includes: calculating the circle center loss, and adjusting the model parameters of the detection head of the circular edge detection branch based on the circle center loss. The circle center loss expression is as follows: in, and is the weight parameter, is the coordinate loss of the center of the cylinder, is the radius loss of the circular cylinder, is the geometric constraint loss of the circular cylinder, and its expression is as follows: in, is the angle between the center of the i-th circular cylinder and the line connecting the centers of the two rectangular slots, calculated based on the predicted coordinate values of the center of the i-th circular cylinder and the predicted coordinate values of the centers of the two adjacent rectangular slots. is the true value of the angle.

4. The method for calculating the shift angle of a non-excited tap changer according to claim 3, characterized in that: The center coordinate loss expression is as follows: in, are the predicted values of the horizontal and vertical coordinates of the center of the i-th circular cylinder, are the true values of the abscissa and ordinate of the center coordinates of the i-th circular cylinder, respectively, and N is the number of circular cylinders of the off-circuit tap changer; The radius loss expression is as follows: in, is the predicted value of the radius of the i-th circular cylinder, is the true value of the radius of the i-th circular cylinder.

5. The method for calculating the shift angle of a non-excited tap changer according to claim 2, characterized in that: When using vertical view image samples to train the angle detection model, the following steps are involved: calculating the slot loss and adjusting the model parameters of the detection head of the slot center detection branch based on the slot loss. The slot loss expression is as follows: in, and is the weight parameter, is the center coordinate loss of the slot, is the geometric parameter loss of the slot, is the angular loss of the slot.

6. The method for calculating the shift angle of a non-excited tap changer according to claim 5, characterized in that: The center coordinate loss expression is as follows: in, are the predicted values of the horizontal and vertical coordinates of the center coordinates of the i-th slot, are the true values of the abscissa and ordinate of the center coordinate of the i-th slot, respectively, and N' is the number of slots of the off-circuit tap changer; The geometric parameter loss expression is as follows: in, are the predicted values of the width and height of the i-th slot, and are the true values of the width and height of the i-th slot respectively; The direction angle loss expression is as follows: in, is the predicted value of the direction angle of the i-th slot, is the true value of the direction angle of the i-th slot.

7. The method for calculating the shift angle of a non-excited tap changer according to claim 2, characterized in that: After detecting the geometric parameters of each circular cylinder in the high-dimensional feature map by the circular edge detection branch, the method further includes: calculating an average of the predicted values of the center coordinates of all the circular cylinders to obtain a predicted value of the center coordinates of the off-excitation tap changer; and calculating the angle between adjacent rectangular slots based on the geometric parameters of each circular cylinder and the geometric parameters of the adjacent rectangular slots, including: Calculate the vectors of the predicted center coordinate value of the off-circuit tap changer and the predicted center coordinate value of each rectangular slot in the adjacent rectangular slots respectively; The angle between the vectors corresponding to each rectangular slot in the adjacent rectangular slots is calculated to obtain the predicted value of the angle between the adjacent rectangular slots.

8. The method for calculating the shift angle of a non-excited tap changer according to claim 7, characterized in that: When the profile angle detection model extracts features from the vertical viewing angle image, specifically, the vertical viewing angle image is input into the convolutional neural network to extract common features in the image. When the profile angle detection model is trained using the vertical viewing angle image samples, the steps include: calculating the loss function of the profile angle detection model, and adjusting the model parameters of the convolutional neural network according to the loss function of the profile angle detection model. The loss function expression of the profile angle detection model is as follows: in, , and They are weight hyperparameters, is the slot loss, is the center loss, is the angle loss, and its expression is as follows: in, is the predicted value of the angle between adjacent rectangular slots, is the actual value of the angle between adjacent rectangular slots.

9. A device for calculating a shift angle of an off-circuit tap changer, comprising a microprocessor and a computer-readable storage medium connected to each other, characterized in that: The microprocessor is programmed or configured to execute the method for calculating the shift angle of the off-circuit tap changer according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is programmed or configured to execute the method for calculating the shift angle of a de-energized tap changer according to any one of claims 1 to 8 through a microprocessor.

Images