Method and device for identifying on-off state of hard pressing plate of transformer substation
By using the dual CNN model and weighted average fusion algorithm in the state recognition of the substation hard plate, the problem of low identification accuracy in the prior art is solved, and high accuracy recognition in complex environments is achieved, and operation and maintenance costs are reduced.
Patent Information
- Application Number
- CN202411628300.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-14
- Publication Date
- 2025-03-18
AI Technical Summary
In the prior art, the accuracy of the status recognition of the hard plate of the substation is low, especially under complex lighting conditions, which is prone to false alarms and missed alarms.
The dual CNN model is adopted, the lightweight CNN model is used for preliminary classification and feature extraction, and the complex CNN model is used for verification and correction. The output results of the two models are integrated through the weighted average fusion algorithm to obtain the recognition results of the hard pressure plate drop state.
It significantly improves the accuracy and robustness of hard press plate status recognition, adapts to complex environments, reduces the dependence of manual inspection, reduces operation and maintenance costs, and improves overall work efficiency.
Smart Images
Figure CN119649091A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of status monitoring of hard pressure plates in substations, and in particular to a method and device for identifying the status of hard pressure plates in substations. Background Art
[0002] As an important part of modern power grid, substation plays a key role in ensuring the stable operation of power system. As a common control element, the hard pressure plate is used to connect or disconnect the circuit, and its switching status is directly related to the safety and reliability of the substation. At present, the status monitoring of the hard pressure plate mainly relies on manual inspection, which is inefficient and prone to errors.
[0003] With the development of computer vision technology and deep learning algorithms, it has become possible to automatically identify the insertion and retraction status of the hard pressure plate. However, the automatic recognition solutions in the existing technology are not accurate enough and have limited application scope, especially under complex lighting conditions, the recognition accuracy still needs to be improved.
[0004] Therefore, there is an urgent need for a solution that can accurately identify the insertion and withdrawal status of the hard pressure plate in various environments. Summary of the invention
[0005] In view of this, it is necessary to provide a method and device for identifying the status of the hard pressure plate of the substation, so as to solve the problems of low accuracy in the identification of the status of the hard pressure plate of the substation in the prior art, especially the low identification accuracy under complex conditions, which is prone to false alarms and missed alarms.
[0006] A method for identifying the activation and deactivation status of a hard pressure plate in a substation, comprising: Get real-time images of the hard platen; Processing the real-time image through a lightweight CNN model to obtain a preliminary recognition result; Processing the preliminary recognition result through a complex CNN model to obtain a refined recognition result; A weighted average fusion algorithm is used to fuse the preliminary recognition result and the refined recognition result, and a recognition result of the hard pressure plate insertion and retraction state is obtained.
[0007] Preferably, a weighted average fusion algorithm is used to integrate the output results of the two models, satisfying: Pfinal=m×PMobileNet+n×PResNet in, Pfinal It represents the final result of integrating the outputs of two models using the weighted average fusion algorithm; PMobileNet Indicates the preliminary recognition result. PResNet Indicates the refined recognition result; m,n are the weights of the preliminary recognition result and the refined recognition result respectively.
[0008] Preferably, obtaining the hard pressing plate insertion and retraction state recognition result includes: comparing the final result with a preset hard platen state threshold; If the final result is less than the preset hard pressure plate state threshold, determining that the hard pressure plate is in the exit state; If the final result is not less than the preset hard pressure plate state threshold, it is determined that the hard pressure plate is in the engaged state.
[0009] Preferably, obtaining a real-time image of the hard pressing plate further comprises: Performing an image enhancement operation on the real-time image; wherein the image enhancement operation includes: Adjust contrast, adjust brightness, and sharpen.
[0010] Preferably, the real-time image is processed by a lightweight CNN model to obtain a preliminary recognition result, including: A lightweight CNN model is used to perform preliminary classification processing to identify the main features of the hard pressing plate in the real-time image; and a feature extraction process is performed to extract key features in the image.
[0011] Preferably, the preliminary recognition result is processed by a complex CNN model to obtain a refined recognition result, including: The complex CNN model re-features the image of the preliminary recognition result, and verifies and corrects the result of the preliminary classification process.
[0012] Preferably, the method further includes training the lightweight CNN model and the complex CNN model: Acquire a training set, wherein the training set includes an image dataset of a hard press plate under different lighting conditions; Annotating the training set, wherein the annotation content includes the position, size, and insertion and retraction status of the hard pressing plate; The lightweight CNN model and the complex CNN model are trained using the labeled training set.
[0013] Preferably, labeling the training set includes: Manually annotate a preset number of manually annotated data; Training a labeling model based on the manually labeled data; The training set is labeled using the labeling model.
[0014] Compared with the prior art, the beneficial effects of this application are: using dual CNN models, the lightweight CNN model is used for preliminary classification and feature extraction, and the complex CNN model is used for verification and correction. The output results of the two models are integrated through a weighted average fusion algorithm, which significantly improves the accuracy of recognition. In addition, this application can realize automatic recognition, which can effectively reduce labor costs compared to traditional manual inspection. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 It is a flow chart of a method for identifying the activation and deactivation status of a hard pressure plate in a substation provided in an embodiment of the present application.
[0016] Figure 2 It is a flow chart of another method for identifying the activation and deactivation status of a hard pressure plate in a substation provided in an embodiment of the present application.
[0017] Figure 3 It is a structural schematic diagram of a device 300 for identifying the activation and deactivation status of a hard pressure plate in a substation provided in an embodiment of the present application. DETAILED DESCRIPTION
[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0019] See also Figure 1 , Figure 1 1 is a flow chart of a method for identifying the status of a substation hard pressure plate provided by an embodiment of the present application. The method for identifying the status of a substation hard pressure plate includes the following steps: A method for identifying the activation and deactivation status of a hard pressure plate in a substation, comprising: S101: Acquire a real-time image of a hard pressing plate.
[0020] The real-time image of the hard pressing plate can be obtained by an industrial camera, a camera of a mobile terminal device (such as a smart phone or a tablet computer), etc.
[0021] Preferably, the real-time image of the hard pressing plate is acquired, and then the method further comprises: performing an image enhancement operation on the real-time image; wherein the image enhancement operation comprises: adjusting contrast, adjusting brightness and sharpening.
[0022] Adjust the contrast to increase the contrast of the image and make the details of the hard plate more obvious. You can use the functions in the OpenCV library to adjust the contrast of the image. For example, you can use the cv2.convertScaleAbs function to increase the contrast by multiplying an enhancement factor. The contrast enhancement factor can be set to 1.5.
[0023] The purpose of adjusting brightness is to increase or decrease the brightness of the image so that it can maintain good visibility under different lighting conditions. You can use functions in the OpenCV library to adjust the brightness of an image. For example, you can use the cv2.convertScaleAbs function to adjust the brightness by adding a brightness increase value. The brightness increase value can be set to 50.
[0024] Sharpening aims to enhance the edges and details of the image and make the outline of the hard plate clearer. You can use filters in the OpenCV library to sharpen the image. For example, you can use the cv2.filter2D function to enhance the details of the image by applying a sharpening filter. The sharpening filter size can be set to 3x3.
[0025] In this embodiment, image enhancement is performed to ensure that clear image input can be obtained even in complex situations such as insufficient light or excessive light.
[0026] S102: Processing the real-time image through a lightweight CNN model to obtain a preliminary recognition result.
[0027] A lightweight CNN model is used to perform preliminary image classification and feature extraction on real-time images. The purpose of preliminary image classification is to quickly identify the main features of the hard pressing plate and determine the approximate state of the hard pressing plate. Among them, the lightweight CNN model can use MobileNet.
[0028] S103: Processing the preliminary recognition result through a complex CNN model to obtain a refined recognition result.
[0029] The complex CNN model verifies and corrects the preliminary recognition results of the lightweight CNN model, especially in complex backgrounds, to further refine the recognition results. The complex CNN model (ResNet model) re-extracts the features of the image, especially the more detailed local features. The preliminary classification results are verified and corrected, especially in complex backgrounds, to further refine the recognition results. Through a more complex model structure, more details and features are captured, false positives and false negatives are reduced, and accuracy is improved.
[0030] Among them, the complex CNN model can be a pre-trained ResNet model with powerful feature extraction capabilities.
[0031] S104: using a weighted average fusion algorithm to fuse the preliminary recognition result and the refined recognition result, and obtaining a recognition result of the hard pressure plate insertion and retraction state.
[0032] The output results of the two models are integrated using a weighted average fusion algorithm to obtain the final recognition result of the hard pressure plate insertion and retraction status.
[0033] Furthermore, the output results of the two models are integrated using a weighted average fusion algorithm to meet the following requirements: Pfinal=m×PMobileNet+n×PResNet in, Pfinal It represents the final result of integrating the outputs of two models using the weighted average fusion algorithm; PMobileNet Indicates the preliminary recognition result. PResNet Indicates the refined recognition result; m,n are the weights of the preliminary recognition result and the refined recognition result respectively.
[0034] The above m,n That is, the weights of the MobileNet model and the ResNet model can be 0.4 and 0.6 respectively. m,n The specific value can be obtained through experiments and experience, etc. This application is not m,n The specific value is limited.
[0035] Furthermore, the recognition result of the hard pressure plate insertion and retraction status is obtained, including: comparing the final result with a preset hard platen state threshold; If the final result is less than the preset hard pressure plate state threshold, determining that the hard pressure plate is in the exit state; If the final result is not less than the preset hard pressure plate state threshold, it is determined that the hard pressure plate is in the engaged state.
[0036] For example, suppose the threshold T=0.5 is set. MobileNet and ResNet The model's recognition results for the "investment" state are PMobileNet=0.7 and PResNet=0.6 respectively, so the final weighted average probability Pfinal It can be expressed as: Pfinal=0.4×PMobileNet+0.6×PResNet=0.4×0.7+0.6×0.6=0.28+0.36=0.64 Judgment: Pfinal ≥ T, the hard pressure plate image is considered to be in the "investment" state.
[0037] This embodiment realizes the automatic identification of the status of the hard pressure plate of the substation, significantly improves the accuracy and robustness of the identification, reduces the reliance on manual inspection, reduces the operation and maintenance costs, and improves the overall work efficiency.
[0038] Furthermore, the method further includes training the lightweight CNN model and the complex CNN model, including: Acquire a training set, wherein the training set includes an image dataset of a hard press plate under different lighting conditions; Annotating the training set, wherein the annotation content includes the position, size, and insertion and retraction status of the hard pressing plate; The lightweight CNN model and the complex CNN model are trained using the labeled training set.
[0039] By using the TensorFlow deep learning framework, a lightweight CNN model based on MobileNet and a complex CNN model based on ResNet can be built and trained. The MobileNet model is used to quickly identify the main features of the hard plate, and the ResNet model is used for precise verification, which improves the recognition speed and accuracy.
[0040] Furthermore, the training set is labeled, including: Manually annotate a preset number of manually annotated data; Training a labeling model based on the manually labeled data; The training set is labeled using the labeling model.
[0041] The labeling can be done manually, using the LabelImg labeling tool, or using a pre-trained (neural network) labeling model. The labeling content includes the position, size, and retraction status of the hard pressure plate. Labeling the training set ensures that these features can be accurately learned during model training, thereby improving recognition accuracy.
[0042] The following is a further introduction to the method for identifying the status of a substation hard pressure plate provided by the present application in conjunction with a specific embodiment. Figure 2 As shown, Figure 2 1 is a flow chart of another method for identifying the status of a substation hard pressure plate provided by an embodiment of the present application, comprising the following steps: S201: Image dataset collection.
[0043] Collect image datasets of the hard platen under different lighting conditions to ensure that the dataset covers a variety of working environments. For example, collect images of the hard platen under different lighting conditions such as daytime, nighttime, indoors, and outdoors. Collect at least 1,000 images under each condition, and a total of about 5,000 images. Ensure that the status of the hard platen (in or out) is clearly visible in each image.
[0044] S202: Annotation of image dataset.
[0045] The LabelImg annotation tool is used to manually annotate the image dataset. The annotation content includes the position, size, and in / out status of the hard platen. The position of each hard platen is represented by a bounding box, and its status (in or out) is recorded. When annotating, ensure that the bounding box fits the edge of the hard platen accurately to avoid including too much background interference.
[0046] S203: Train the CNN model.
[0047] The model was developed using the TensorFlow deep learning framework. First, the image dataset was preprocessed, including scaling and normalization. Then, a lightweight CNN model based on MobileNet and a complex CNN model based on ResNet were built. During the model training process, the following parameters were set: MobileNet model: Use ImageNet pre-trained weights to initialize the model, freeze the first 13 convolutional layers, and only train the 14th and later layers and the fully connected layers. Set the learning rate to 0.001, the batch size to 32, and the number of training rounds to 30.
[0048] ResNet model: Also use the ImageNet pre-trained weights to initialize the model, freeze the first 6 layers, and unfreeze the last few layers, usually including the global average pooling layer and the fully connected layer, which are responsible for the classification task. Set the learning rate to 0.0005, the batch size to 16, and the number of training rounds to 50.
[0049] S204: Setting the hard pressure plate state recognition method.
[0050] Use the two trained CNN models to identify the insertion and retraction status of the hard pressure plate. Use the weighted average fusion algorithm to integrate the output results of the two models to obtain the final recognition result of the insertion and retraction status of the hard pressure plate. Set the weight of the MobileNet model to 0.4; the weight of the ResNet model to 0.6.
[0051] When the probability that the two models predict that a certain hard pressure plate is in the "engaged" state is higher than the threshold, the hard pressure plate is considered to be in the engaged state. When the probability is lower than the threshold, the hard pressure plate is considered to be in the exited state.
[0052] S205: Real-time image acquisition.
[0053] Use a camera, such as a mobile device camera, to obtain a real-time image of the hard platen. Make sure the camera resolution is no less than 12 megapixels and enable the autofocus function to ensure image clarity.
[0054] S206: Perform image enhancement processing on the real-time image.
[0055] Perform image enhancement, including adjusting contrast, brightness, and sharpening. You can use the OpenCV library and set the parameters as follows: ▪Contrast enhancement factor: 1.5 ▪Brightness increase value: 50 ▪Sharpening filter size: 3x3 By adjusting these parameters, you can ensure clear image input even in low or high light conditions.
[0056] S207: Lightweight CNN model for preliminary classification and feature extraction.
[0057] Use the trained MobileNet model to perform preliminary image classification and feature extraction on real-time images. Set the input image size to 224x224 pixels during inference and ensure that the input image is normalized.
[0058] S208: Verify and calibrate complex CNN models.
[0059] Use the trained ResNet model to verify and correct the preliminary recognition results. Especially in complex backgrounds or low-light environments, the ResNet model will further confirm the status of the hard plate through more detailed feature analysis. Set the input image size for inference to 299x299 pixels and ensure that the input image is normalized.
[0060] S209: Using a weighted average fusion algorithm to integrate the output results of the two models.
[0061] The output results of the two models are integrated using a weighted average fusion algorithm to obtain the final recognition result of the hard platen "in" state. The weights are set as follows: The weight of the MobileNet model is 0.4; the weight of the ResNet model is 0.6; the threshold T is set to 0.5; The recognition results of the "investment" state by the MobileNet and ResNet models are: PMobileNet=0.7; PResNet=0.6; Pfinal==0.4×0.7+0.6×0.6=0.28+0.36=0.64 According to Pfinal≥T, since 0.64≥0.5, the hard pressure plate image is considered to be in the "investment" state.
[0062] Through the above weight setting, it is ensured that the final recognition result is both fast and accurate.
[0063] In summary, this application significantly improves the accuracy and robustness of substation hard plate state recognition by introducing deep learning technology, especially the dual CNN model and image enhancement function, and can adapt to complex environments such as weak light and strong light. At the same time, it reduces the reliance on manual inspection, reduces operation and maintenance costs, and improves overall work efficiency.
[0064] Based on the same inventive concept, the present application also provides a device 300 for identifying the status of a hard pressure plate in a substation. Figure 3 The structural diagram of the device 300 for identifying the insertion and withdrawal status of the hard pressure plate of a substation shown in FIG. 1 includes: An image acquisition module 301 is used to obtain a real-time image of the hard pressing plate; A first recognition module 302 is used to process the real-time image through a lightweight CNN model to obtain a preliminary recognition result; The second recognition module 303 processes the preliminary recognition result through a complex CNN model to obtain a refined recognition result; The result recognition module 304 is used to fuse the preliminary recognition result and the refined recognition result by using a weighted average fusion algorithm, and obtain a recognition result of the hard pressing plate insertion and retraction state.
[0065] Furthermore, it also includes an image enhancement module 305, which is used to perform image enhancement operations on the real-time image.
[0066] In the above embodiment, the specific implementation method of each functional module of the device 300 for identifying the status of a hard pressure plate in a substation can be referred to the introduction in the above embodiment of the method for identifying the status of a hard pressure plate in a substation, which will not be described in detail in this embodiment.
Claims
1. A method for identifying the status of a substation hard pressure plate, characterized in that: include: Get real-time images of the hard platen; Processing the real-time image through a lightweight CNN model to obtain a preliminary recognition result; Processing the preliminary recognition result through a complex CNN model to obtain a refined recognition result; A weighted average fusion algorithm is used to fuse the preliminary recognition result and the refined recognition result, and a recognition result of the hard pressure plate insertion and retraction state is obtained.
2. The method according to claim 1, characterized in that The output results of the two models are integrated using a weighted average fusion algorithm to meet the following requirements: Pfinal=m×PMobileNet+n×PResNet in, Pfinal It represents the final result of integrating the outputs of two models using the weighted average fusion algorithm; PMobileNet Indicates the preliminary recognition result. PResNet Indicates the refined recognition result; m,n are the weights of the preliminary recognition result and the refined recognition result respectively.
3. The method according to claim 2, characterized in that Obtain the hard pressure plate throw-in and throw-out status recognition results, including: comparing the final result with a preset hard platen state threshold; If the final result is less than the preset hard pressure plate state threshold, determining that the hard pressure plate is in the exit state; If the final result is not less than the preset hard pressure plate state threshold, it is determined that the hard pressure plate is in the engaged state.
4. The method according to claim 1, characterized in that: Acquire a real-time image of the hard pressing plate, and then further include: Performing an image enhancement operation on the real-time image; wherein the image enhancement operation includes: Adjust contrast, adjust brightness, and sharpen.
5. The method according to claim 1, characterized in that: The real-time image is processed by a lightweight CNN model to obtain preliminary recognition results, including: A lightweight CNN model is used to perform preliminary classification processing to identify the main features of the hard pressing plate in the real-time image; and a feature extraction process is performed to extract key features in the image.
6. The method according to claim 1, characterized in that The preliminary recognition results are processed by a complex CNN model to obtain refined recognition results, including: The complex CNN model re-features the image of the preliminary recognition result, and verifies and corrects the result of the preliminary classification process.
7. The method according to claim 1, characterized in that: It also includes training the lightweight CNN model and the complex CNN model: Acquire a training set, wherein the training set includes an image dataset of a hard press plate under different lighting conditions; Annotating the training set, wherein the annotation content includes the position, size, and insertion and retraction status of the hard pressing plate; The lightweight CNN model and the complex CNN model are trained using the labeled training set.
8. The method according to claim 7, characterized in that The training set is labeled, including: Manually annotate a preset number of manually annotated data; Training a labeling model based on the manually labeled data; The training set is labeled using the labeling model.
9. A device for identifying the status of a substation hard pressure plate, characterized in that: include: An image acquisition module, used to obtain a real-time image of the hard pressing plate; A first recognition module, used to process the real-time image through a lightweight CNN model to obtain a preliminary recognition result; The second recognition module processes the preliminary recognition result through a complex CNN model to obtain a refined recognition result; The result recognition module is used to fuse the preliminary recognition result and the refined recognition result by using a weighted average fusion algorithm, and obtain a recognition result of the hard pressing plate's insertion and retraction state.
10. The device according to claim 9, characterized in that It also includes an image enhancement module, which is used to perform image enhancement operations on the real-time image.