CN-122024214-A - Transformer substation pointer type instrument reading method based on improved YOLO11-OBB

CN122024214ACN 122024214 ACN122024214 ACN 122024214ACN-122024214-A

Abstract

The invention discloses a substation pointer type instrument reading method based on improved YOLO11-OBB, which comprises the steps of obtaining a substation field pointer type instrument image to be read and preprocessing the substation field pointer type instrument image, inputting the preprocessed instrument image into a trained pointer identification model and performing post-processing to obtain position information of a pointer, wherein the pointer identification model is based on an improved YOLO11-OBB network design, the improved YOLO11-OBB network comprises a main network, a neck network and an OBB preprocessing head, a CBAM attention module is inserted behind a4 th layer C3k2 module in the main network, and a CBAM attention module is respectively inserted behind a 13 th layer C3k2 module and a 16 th layer C3k2 module in the neck network. The method has high precision, high speed and strong robustness, is suitable for an unmanned operation and maintenance system of the transformer substation, and effectively improves the reading efficiency of pointer data of the power equipment.

Inventors

ZOU MINGHE
XIA BINGXIN
GUO YING
LIU YU
XIA CHUNYONG
YANG BIN
YU LIANGLIANG
SUN ZHE
MENG HONGYU
LIU BONAN

Assignees

国网辽宁省电力有限公司沈阳供电公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A substation pointer type meter reading method based on improved YOLO11-OBB, comprising: acquiring a pointer type instrument image of a transformer substation site to be read and preprocessing the pointer type instrument image to obtain a preprocessed instrument image; Inputting the preprocessed instrument image into a trained pointer recognition model and performing post-processing to obtain position information of a pointer, wherein the pointer recognition model is designed based on an improved YOLO11-OBB network, the improved YOLO11-OBB network comprises a main network, a neck network and an OBB pre-measuring head, a CBAM attention module is inserted behind a 4 th layer C3k2 module in the main network and used for enhancing the attention of a P3 feature image output by the C3k2 module, a CBAM attention module is inserted behind a 13 th layer and a 16 th layer C3k2 module in the neck network and used for enhancing the attention of a P4 fusion feature image and a P3 fusion feature image output by the C3k2 module, an anchor frame of a P3 feature layer in the OBB pre-measuring head is [2,12,4,24,6,36], an anchor frame of the P4 feature layer is [3,18,5,30,7,42], and an anchor frame of the P5 feature layer is [4,24,6,36,8,48]; and calculating and correcting the instrument coordinates based on the position information to obtain accurate readings.
2. The method for pointer meter reading in a substation based on modified YOLO11-OBB according to claim 1, wherein said preprocessing comprises adaptive gaussian denoising and CLAHE illumination normalization.
3. The method for reading a pointer instrument of a transformer substation based on improved YOLO11-OBB according to claim 2, wherein said step of adaptive gaussian denoising is as follows: Respectively calculating the horizontal gradient and the vertical gradient of each pixel in the instrument image, and obtaining the comprehensive gradient through a gradient amplitude formula ; According to the integrated gradient Dividing the regions and assigning different Gaussian variances to different regions, wherein the Gaussian variance assignment rule is as follows when When the threshold is preset, the first variance value is obtained When a threshold is preset, a second variance value is taken, wherein the first variance value is smaller than the second variance value; And carrying out adaptive Gaussian filtering on the instrument image based on the Gaussian variance distribution rule.
4. The substation pointer meter reading method based on improved YOLO11-OBB according to claim 2, wherein said step of CLAHE illumination normalization is as follows: Uniformly dividing the denoised image into a plurality of non-overlapping pixel blocks; counting the gray value distribution of each pixel block to obtain a gray histogram of each pixel block; Setting a contrast threshold value T for the gray level histogram of each pixel block, cutting off the part exceeding the contrast threshold value T and evenly distributing the part to other gray levels; and performing equalization operation on the truncated histogram.
5. The method for pointer meter reading of a substation based on improved YOLO11-OBB according to claim 1, wherein training said pointer identification model comprises the steps of: Carrying out multi-scale feature extraction on the received image by utilizing a backbone network to obtain a P3 feature map, a P4 feature map and a P5 feature map; Optimizing the P3 feature map, the P4 feature map and the P5 feature map by using a neck network to obtain three optimized scale feature maps; Obtaining three-scale screened high-quality OBB based on the optimized three-scale feature maps by using an OBB prediction head; And performing cross-scale screening on the high-quality OBB to obtain a pointer detection result.
6. The method for reading pointer type instrument of transformer substation based on improved YOLO11-OBB according to claim 5, wherein the method for multi-scale feature extraction by the backbone network is as follows: Performing first convolution and downsampling on a 640×640×3 standardized image through a Conv (64, 3, 2) convolution layer to output a 320×320×64 feature map; further downsampling by Conv (128, 3, 2) convolutional layers, outputting a 160×160×128 feature map; a 160×160×256 reinforcement feature map is output through a C3k2 (256, false, 0.25) module; Downsampling by a Conv (256, 3, 2) convolution layer to output an 80×80×256 feature map; The channel number is increased to 512 through a C3k2 (512, false, 0.25) module, a P3 layer core feature map of 80 multiplied by 512 is output, and then the P3 feature map after the attention of 80 multiplied by 512 is output through a CBAM (512) attention module; Downsampling the P3 characteristic diagram after the attention is enhanced through Conv (512, 3, 2) convolution layers, outputting a 40 multiplied by 512 characteristic diagram, and outputting a 40 multiplied by 512P 4 layer strengthening characteristic diagram through a C3k2 (512, true) module; The P4 layer strengthening characteristic diagram is downsampled through Conv (1024, 3, 2) convolution layers, the channel number is increased to 1024, a 20 multiplied by 1024 characteristic diagram is output, after being strengthened through a C3k2 (1024, true) module, 5 multiplied by 5 space pyramid pooling is executed through an SPPF (1024, 5) module, global characteristics are optimized through a C2PSA (1024) module, and a 20 multiplied by 1024P 5 layer characteristic diagram is output.
7. The substation pointer meter reading method based on improved YOLO11-OBB according to claim 6, wherein the method for obtaining the optimized three scale feature map using the neck network is as follows: For feature fusion of P5 to P4, nearest neighbor interpolation up-sampling is firstly carried out on the P5 feature map, the nearest neighbor interpolation up-sampling and the P4 layer strengthening feature map are spliced in a channel dimension through Concat operation, the dimension is reduced and the fusion is strengthened through a C3k2 (512, false) module convolution, the P4 fusion feature map is output, and the P4 attention strengthening feature map is output through a CBAM (512) attention module; Aiming at feature fusion of P4 to P3, up-sampling a P4 attention enhancement feature map, splicing the up-sampling feature map with a P3 layer core feature map through Concat operation, convolving to reduce dimensions and strengthen small-scale features through a C3k2 (256, false) module, and outputting a P3 final feature map through a CBAM (256) attention module; Downsampling the P3 final feature map through Conv (256, 3, 2) convolution layers, splicing the downsampled P3 final feature map with the P4 fusion feature map through Concat operation, and outputting the P4 final feature map after strengthening through a C3k2 (512, false) module; The P4 final feature map is downsampled through Conv (512, 3, 2) convolution layers, spliced with the P5 layer final feature map through Concat operation, and the P5 final feature map is output after being strengthened through a C3k2 (1024, true) module.
8. The substation pointer type meter reading method based on improved YOLO11-OBB according to claim 7, wherein the method for obtaining three-scale screened high-quality OBB based on the optimized three-scale feature map by using the OBB prediction head is as follows: integrating the P3, P4 and P5 final feature images to form a multi-scale input tensor, and then loading 3 groups of pointer-adaptive anchor frames and matching with the corresponding scale feature images to generate an initial candidate frame set; Predicting the offset of each candidate frame by a convolution layer , , , , ) , wherein, , The coordinate offsets of the candidate frames are respectively determined, , The offsets of the width and height of the candidate frame respectively, The angular offset of the candidate frame; Calculating final boundary frame parameters based on the original parameters of the candidate frames and the prediction offset, and predicting the confidence that each candidate frame belongs to the pointer class through a convolution layer and a Sigmoid activation function; And executing non-maximum value inhibition operation on the candidate frames of each scale to obtain the high-quality OBB after screening of each scale.
9. The method for reading pointer-type meters of transformer substation based on improved YOLO11-OBB according to claim 1, wherein the loss function used during training is The method comprises the following steps: ; Wherein, the Is a category confidence loss, For the OBB regression loss, Loss for angular regression; category confidence loss is optimized by a cross entropy loss function; The OBB regression loss is optimized by GIoU loss functions; The angular regression loss is optimized by the smoothl 1 loss function.
10. The method for reading pointer type meters of transformer substation based on improved YOLO11-OBB according to claim 1, wherein the method for calculating and correcting the coordinates of the meters based on the position information of the pointer to obtain accurate readings is as follows: calculating an initial included angle between the pointer and the horizontal direction, and adjusting the initial included angle to be 0-360 degrees; According to the mapping relation between the actual angle of the pointer and the measuring range of the instrument, calculating the initial reading of the instrument ; Correcting the initial readings of the meter to eliminate environmental interference to obtain corrected readings ; Take initial readings And correct the reading As the final reading.

Description

Transformer substation pointer type instrument reading method based on improved YOLO11-OBB Technical Field The invention relates to the technical field of intersection of substation equipment state monitoring and computer vision deep learning, and particularly provides an automatic reading method for pointer type meters such as an ammeter, a voltmeter and a power meter in a substation. Background Under the background of rapid development of a smart power grid, the intelligent trend of a transformer substation is that the transformer substation is used as a core hub of a power system, a large number of pointer type meters in the transformer substation are key carriers for reflecting the running state of equipment, the angles of meter pointers directly correspond to core data such as voltage, current and SF6 air pressure, and once pointer readings are abnormal, the fault conditions such as overload and insulation failure of the equipment are possibly predicted, and then operation and maintenance personnel are required to conduct investigation and treatment in time. Therefore, the accurate identification and reading of the pointer type instrument are one of the key links of operation and maintenance work of the transformer substation. However, the current pointer meter reading approach has significant drawbacks: The manual inspection efficiency is low, the error is large, the small-sized and medium-sized transformer stations rely on manual line-by-line check, and the large-sized transformer stations need multi-person cooperation; The suitability of the general deep learning model is insufficient, common models such as YOLOv and YOLOv8 adopt Axis Alignment Bounding Boxes (AABB), a pointer cannot be adapted to an elongated shape, the positioning precision is low, the number of parameters of high-precision models such as YOLOv l is 43M, the frame rate of edge equipment reasoning is low, real-time monitoring cannot be achieved, and the precision of a YOLO11 lightweight model is greatly reduced without being optimized for the pointer. Therefore, the pointer reading method of the transformer substation pointer type instrument, which is accurate in pointer positioning and accurate in angle calculation, is provided, and the problem to be solved is urgent. Disclosure of Invention In view of the above, the invention provides a transformer substation pointer type meter reading method based on improved YOLO11-OBB, which aims to solve the problems of low pointer type meter reading efficiency, poor precision, difficult model deployment and the like in the prior art. The invention provides a substation pointer type instrument reading method for improving YOLO11-OBB, which comprises the following steps: acquiring a pointer type instrument image of a transformer substation site to be read and preprocessing the pointer type instrument image to obtain a preprocessed instrument image; Inputting the preprocessed instrument image into a trained pointer recognition model and performing post-processing to obtain position information of a pointer, wherein the pointer recognition model is designed based on an improved YOLO11-OBB network, the improved YOLO11-OBB network comprises a main network, a neck network and an OBB pre-measuring head, a CBAM attention module is inserted behind a 4 th layer C3k2 module in the main network and used for enhancing the attention of a P3 feature image output by the C3k2 module, a CBAM attention module is inserted behind a 13 th layer and a 16 th layer C3k2 module in the neck network and used for enhancing the attention of a P4 fusion feature image and a P3 fusion feature image output by the C3k2 module, an anchor frame of a P3 feature layer in the OBB pre-measuring head is [2,12,4,24,6,36], an anchor frame of the P4 feature layer is [3,18,5,30,7,42], and an anchor frame of the P5 feature layer is [4,24,6,36,8,48]; and calculating and correcting the instrument coordinates based on the position information to obtain accurate readings. Preferably, the preprocessing includes adaptive gaussian denoising and CLAHE illumination normalization. Further preferably, the step of adaptive gaussian denoising is as follows: Respectively calculating the horizontal gradient and the vertical gradient of each pixel in the instrument image, and obtaining the comprehensive gradient through a gradient amplitude formula ; According to the integrated gradientDividing the regions and assigning different Gaussian variances to different regions, wherein the Gaussian variance assignment rule is as follows whenWhen the threshold is preset, the first variance value is obtainedWhen a threshold is preset, a second variance value is taken, wherein the first variance value is smaller than the second variance value; And carrying out adaptive Gaussian filtering on the instrument image based on the Gaussian variance distribution rule. Further preferably, the step of CLAHE illumination normalization is as follows: Uniformly dividing the den