A Deep Learning-Based Method for Component Detection in Transmission Areas
By using a deep learning-based method for detecting transformer substation components, combined with binocular ranging technology, the problem of time-consuming and labor-intensive transformer substation inspections has been solved. This method enables rapid and accurate detection and identification of components in the substation area, thereby improving inspection efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-14
- Publication Date
- 2026-03-10
AI Technical Summary
The current inspection of transformer substations mainly relies on manual measurement, which is time-consuming and labor-intensive, and it is difficult to efficiently determine whether the components are installed in accordance with regulations.
A deep learning-based method for detecting transformer components, combined with binocular ranging technology, is adopted to achieve intelligent identification and quantitative measurement of transformer components through sample collection, annotation, and model training.
It enables rapid and accurate detection of components in the transformer substation, reduces the burden on inspection personnel, improves detection quality, shortens inspection time, and achieves high levels of recall rate and average overlap.
Smart Images
Figure CN113963172B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of deep learning, in particular to a transformer part detection method based on deep learning. BACKGROUND
[0002] According to the technical novelty report, the existing standard transformer inspection method is still mainly manual inspection, mainly through visual observation of transformer parts and manual measurement to judge the quality of the transformer, and according to the result to judge whether the transformer installation is standard. Manual inspection judges the quality of the transformer by visual observation of transformer parts and manual measurement, which requires time and effort. SUMMARY
[0003] The present application is to overcome the problem of high cost and long time of manual measurement in the prior art, and provides a transformer part detection method based on deep learning, which realizes the saving of inspection cost, the shortening of inspection time and the reduction of the burden of inspection personnel. The intelligent detection and identification of transformer parts can be realized by using deep learning technology, and the measurement of part of the quantitative index of the transformer can be realized by combining the binocular distance measurement technology. Another object of the present application is to collect samples by binocular camera, which can standardize the measurement of the distance of the installation position of the parts, and facilitate the next step of processing the picture data.
[0004] In order to achieve the above purpose, the present application adopts the following technical scheme:
[0005] A transformer part detection method based on deep learning, comprising the following steps:
[0006] S1, sample collection, collecting picture data of multiple distances, angles and directions of multiple standard transformers in multiple areas;
[0007] S2, sample annotation of the collected samples;
[0008] S3, after the sample is adjusted, input the transformer part detection model for model training and pre-processing of the sample picture to be tested;
[0009] S4, model inference.
[0010] The present application can complete the quality judgment of the transformer part by processing and calculating the picture, and judge whether the transformer installation is standard according to the result, which can shorten the inspection time and realize the measurement of the quantitative index of the transformer part. The sample collection is carried out by binocular camera, and the distance measurement of the installation position of the parts can be standardized, which facilitates the next step of processing the picture data.
[0011] The GPU performs sample labeling on the collected pictures, first identifies the target type of the components to be detected in the pictures, and calculates and processes the distance coordinate data. The pictures are input into the trained model after being resized, and the sample pictures with target detection boxes are output. According to the types and positions of the target detection boxes on the pictures, it is judged whether the installation of the transformer area is reasonable, and the quantitative inspection work is realized to achieve fast inspection and improve the detection quality.
[0012] As preferred, the S2 includes the following content:
[0013] Define the target categories that need to be labeled in the sample pictures, including JP cabinets, signboards, transformers, warning signs, pole number plates, phase plates, and the areas of each cross arm between the poles.
[0014] Design sample labeling rules, and label the target area as the minimum circumscribed rectangle of the actual target area. Complete the labeling work of the collected samples.
[0015] As preferred, the target types in the JP cabinet, signboard, transformer, warning sign, pole number plate, phase plate, and the area of each cross arm between the poles are defined as the target set W={w1, w2, …, w n}。
[0016] As preferred, the S2 further includes the following content: each target area is a polygon, the polygon includes i vertices Si, and the diagonal lines formed by connecting the i vertices with other vertices have 1 / 2*i(i-3) diagonal lines, the intersection points between the diagonal lines are used as a plurality of feature points, the distance ratio of each feature point to each vertex is identified, and the GPU of the transformer component detection model compares the distance ratio of the feature points of the target area with the existing target types to obtain the target type in the sample picture and construct the target set.
[0017] As preferred, the GPU model is GTX2080Ti.
[0018] As preferred, the S3 includes the following content: after resizing the sample, the corresponding target set in the sample is input into the transformer component detection model to predict and regress each type of target area in the sample, the deviation value of the position of each type of target area predicted by the sample target set from the actual labeled position is evaluated according to the yolov3 loss function, and the deviation value is used to optimize the network parameters through back propagation;
[0019] Repeat the above operation to iteratively optimize until the deviation value converges, and complete the training of the model.
[0020] As preferred, the network structure of the transformer component detection model includes Data, CBM, a plurality of DEG and DTL, UTL, and OP.
[0021] Data represents data input;
[0022] CBM represents a conforming operation, CBM is composed of three parts in series: convolution, batch normalization and activation function; DeG represents dense organization, each DeG is composed of a plurality of dense units in series, and the output of each dense unit and the input of other dense units are spliced by channel;
[0023] DTL represents a down-sampling transition layer;
[0024] UTL represents an up-sampling transition layer;
[0025] OP represents an output layer.
[0026] As a preferred, the S4 comprises the following contents: inputting the picture pre-processed to-be-tested sample into the transformer detection model, the transformer detection model is internally provided with a component segmentation model, and the component segmentation model outputs a predicted display of each transformer component target detection frame on the sample picture.
[0027] As a preferred, each transformer detection component target detection frame corresponds to different display frame colors according to different target categories.
[0028] Therefore, the present application has the following beneficial effects:
[0029] The present application proposes a transformer component detection method based on deep learning technology aiming at the problems caused by the artificial inspection mode in the standard transformer inspection process. The method designs a transformer component detection network based on deep learning, and the evaluation result of the model trained based on the collected samples for the collected transformer test samples is a recall rate of 99.99%, an average overlap degree of 92.5%, and a detection speed of 45ms per picture in the GPU GTX2080Ti environment, which realizes accurate and rapid detection and identification of transformer components. BRIEF DESCRIPTION OF DRAWINGS
[0030] Figure 1 is a flowchart of the present embodiment.
[0031] Figure 2 is a transformer component detection model training flowchart of the present embodiment.
[0032] Figure 3 is a transformer component detection model network structure diagram of the present embodiment.
[0033] Figure 4 is a structure diagram of CBM, DU, DeG, DTL and UTL of the present embodiment.
[0034] Figure 5 is a detection result and part of the detail diagram of the present embodiment. DETAILED DESCRIPTION
[0035] The application will be further described in conjunction with the accompanying drawings and specific embodiments.
[0036] Embodiments:
[0037] The embodiment provides a transformer area component detection method based on deep learning, as shown in Figure 1 、 2 , comprising the following steps:
[0038] S1, using a binocular camera to collect samples, collecting picture data of multiple distances, angles and directions of multiple regional standard transformer areas, and the maximum range of the picture of each transformer area including all components thereof;
[0039] S2, sample labeling is performed on the collected samples;
[0040] The S2 comprises the following contents:
[0041] Defining the target categories to be labeled in the sample picture, including a JP cabinet, a signboard, a transformer, a warning board, a pole number plate, a phase plate, an abnormal line merger clamp and a pole area between each cross arm;
[0042] The S2 comprises the following contents:
[0043] Defining the target categories to be labeled in the sample picture, including a JP cabinet, a signboard, a transformer, a warning board, a pole number plate, a phase plate, an abnormal line merger clamp and a pole area between each cross arm;
[0044] Each target area is a polygon, the polygon comprises i vertices Si, the diagonal lines formed by the i vertices and other vertices have 1 / 2*i(i-3) in total, the intersection points between the diagonal lines are used as a plurality of feature points, the distance ratios of each feature point to the vertices are identified, the GPU of the transformer area component detection model compares the distance ratios of the feature points of the target area with existing target types to obtain the target types in the sample picture and construct a target set;
[0045] The target types in the JP cabinet, the signboard, the transformer, the warning board, the pole number plate, the phase plate, the abnormal line merger clamp and the pole area between each cross arm are defined as the target set W={w1, w2,..., w n}.
[0046] Designing a sample labeling rule, labeling the target area as the minimum circumscribed rectangle of the actual target area; and completing the labeling work of the collected samples.
[0047] S3, the sample is input into the substation component detection model after adjustment for model training and pre-processing of the sample picture to be tested; S3 includes the following: after adjusting the size of the sample, input the corresponding target set in the sample into the substation component detection model to predict and regress each target area in the sample, evaluate the deviation value of the position of each target area in the sample target set currently predicted from the yolov3 loss function, and optimize the network parameters by using the deviation value. Repeat the above operation, and iterate and optimize until the deviation value converges, and complete the training of the model.
[0048] As shown in Figure 3 , the network structure of the substation component detection model includes Data, CBM, a plurality of DEG and DTL, UTL and OP.
[0049] Define num to represent the number of convolution kernels, k to represent the size of the convolution kernel, s to represent the step, and p to represent the number of zero-padding.
[0050] In the above structure diagram, DeG represents a dense group (DeG), which is composed of a plurality of dense units (DeU) in series, and the output of each dense unit and its input will perform a channel splicing operation; DTL represents a downsample transition layer (DTL), and UTL represents an upsample transition layer (UTL); CBM is a composite operation composed of three parts in series: convolution (Conv), batch normalization (BN) and activation function Mish; OP represents an output layer.
[0051] In the dense unit, 1x1 CBM represents that the convolution kernel size of the convolution layer is 1, which is configured as [num=64, k=1, s=1, p=0], and 1x1 CBM represents that the convolution kernel size of the convolution layer is 3, which is configured as In the above diagram (3), c represents channel splicing. In the DTL, the maximum pooling layer is configured as [k=2, s=2], and in the UTL, the upsample multiple is 2 times.
[0052] The structure of CBM, DU, DeG, DTL and UTL is shown in Figure 4 .
[0053] The configuration and output of the network structure of the substation component detection model are shown in Table 1:
[0054] Table 1
[0055]
[0056]
[0057] If the detection component type is N, then M=(5+N)*3 in the above table.
[0058] S4, model inference is performed.
[0059] The image preprocessed sample to be tested is input into the transformer detection model, and the transformer detection model is internally provided with a component segmentation model. Figure 5 As shown in the figure, from left to right are a standard transformer picture, a detection result, and a result part detail. The component segmentation model outputs a prediction of each transformer component target detection frame on the sample picture. Each transformer detection component target detection frame corresponds to a different display frame color according to the target category. The installation of the transformer is judged according to the type and position of the target detection frame on the picture. The quantitative inspection work realizes fast inspection and improves the detection quality.
[0060] The above embodiments are only used for further description of the present application and cannot be understood as a limitation on the protection scope of the present application. The skilled in the art can make some non-essential improvements and adjustments to the present application according to the above content of the present application, which falls within the protection scope of the present application.
Claims
1. A deep learning-based transformer component detection method, characterized in that, Comprise the following steps: S1, sample collection, collect a variety of distance, angle, azimuth of a plurality of regional standard area picture data; S2, sample labeling of the collected sample: Define the target categories that need to be labeled in the sample picture, including JP cabinet, signboard, transformer, warning sign, pole sign, phase sign, and the area of each cross arm between the poles; Each target area is a polygon, the polygon includes i vertices Si, the intersection of the 1 / 2*i(i-3) diagonal lines connected by the i vertices is a plurality of feature points, the distance ratio of each feature point to each vertex is identified, and the GPU of the substation component detection model compares the distance ratio of the feature points of the target area with the existing target type to obtain the target type in the sample picture and construct a target set; Design sample labeling rules, and label the target area as the minimum circumscribed rectangle of the actual target area; S3, the adjusted sample is input into the substation component detection model for model training and pre-processing of the sample picture to be tested; S4, model inference, realize the measurement of substation quantitative indicators.
2. The method of claim 1, wherein the method comprises: The target types including JP cabinet, signboard, transformer, warning sign, pole sign, phase sign, and the area of each cross arm between the poles are defined as a target set W={w1, w2,..., wn}.
3. The deep learning-based method for detecting transformer components according to claim 1, characterized in that, The GPU model is GTX2080Ti.
4. The method of claim 1, wherein the method comprises: The S3 Comprise the following contents: S3.1: After adjusting the size of the sample, input the corresponding target set in the sample into the substation component detection model to predict and regress each target area in the sample, evaluate the deviation value of the position of each target area predicted by the sample target set from the actual labeled position according to the yolov3 loss function, and optimize the network parameters by using the deviation value back propagation; S3.2: Repeat S3.1, iterate and optimize until the deviation value converges, and complete the training of the model.
5. The method of claim 4, wherein the method further comprises: The network structure of the substation component detection model comprises Data, CBM, a plurality of DEG and DTL, UTL and OP; Data represents data input; CBM represents a composite operation, which is composed of convolution, batch normalization and activation function in series; DEG represents dense organization, each DEG is composed of a plurality of dense units in series, and the output of each dense unit and the input of other dense units are spliced by channel; DTL represents down-sampling transition layer; UTL represents up-sampling transition layer; OP represents output layer.
6. The method of claim 1, wherein the method comprises: The S4 Comprise the following contents: input the sample to be tested after picture preprocessing into the substation detection model, the substation detection model is provided with a component segmentation model, and the component segmentation model outputs the display of each substation component target detection frame in the sample picture.
7. The method of claim 3, wherein the method is characterized by, Each substation detection component target detection frame has different display frame colors according to different target categories.
Citation Information
Patent Citations
Distribution transformer district acceptance method and device based on image recognition and computer storage medium
CN110490105A
Kiln surface defect detection method and system and medium
CN112288043A
Character detection method and device, electronic equipment and storage medium
CN112348025A