Method for detecting abnormal equipment in transformer substation based on low-illumination image enhancement algorithm

Abstract

The invention discloses a method for detecting abnormal equipment in a transformer substation based on a low-illumination image enhancement algorithm, and the method comprises the steps: 1, collecting transformer substation equipment images under different illuminance, carrying out the preprocessing of image data through a mosaic data enhancement algorithm, and optimizing a data set; 2, building an improved YOLOv8 model by adopting a low illumination enhancement technology; 3, adding a global attention module (GAM) in the YOLOv8 model to optimize the model precision; step 4, acquiring real-time substation equipment image data, performing real-time monitoring by using the improved model, and finally evaluating the model by using indexes such as precision and recall rate; according to the method, the data set is enhanced, the model feature recognition capability is improved, better detection is realized in a low-light environment, the detection speed and accuracy are excellent, the abnormal power equipment in the transformer substation can be effectively recognized and positioned, the real-time detection of the equipment is realized, and the practicability is high. And reliable technical support is provided for maintenance and fault diagnosis of a power system.

Description

Technical Field

[0001] The present invention relates to the field of electric power vision technology, and in particular to a method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm. Background Art

[0002] At present, thermal defects and diagnosis of power equipment have become an important part of ensuring the safe operation of power systems; traditional thermal defect detection methods mainly rely on manual inspections and simple temperature monitoring. Although these methods can detect equipment abnormalities to a certain extent, they have significant limitations; manual inspections are inefficient and labor-intensive, and rely on the experience and skills of operators, which can easily lead to missed detections or misjudgments; although simple temperature monitoring can provide real-time temperature data, it cannot accurately locate the specific location of the defect and is difficult to cope with complex environmental conditions; the particularity of thermal defect detection in power equipment brings new challenges, especially in low-light environments, the decline in image quality will significantly affect the detection effect of the model. Under low-light conditions, images often have problems such as low contrast, high noise, and loss of details, which makes it difficult for traditional deep learning models to extract effective features, thereby affecting the accuracy and reliability of detection.

[0003] In order to solve these problems, researchers have gradually introduced image enhancement technology into deep learning models to improve the detection accuracy of the model in complex environments, especially in low-light environments. How to effectively enhance the quality of images to adapt to the feature extraction requirements of deep learning models has become one of the current research hotspots.

[0004] Chinese patent document number CN115187771A, invention name "A fire identification method based on YOLO-V4 convolutional neural network", discloses a fire identification method based on YOLO, which enhances target detection capability in complex environments and uses loss activation function to improve model generalization ability.

[0005] Chinese patent document number CN118521553A, invention name "A method for detecting abnormality of substation equipment based on lightweight YOLO model", discloses a lightweight YOLO model, which achieves overall efficiency by optimizing the data set and improving the YOLO model to lightweight the model and improve performance.

[0006] In recent years, low-light image enhancement methods based on convolutional neural networks (CNN) have become a research hotspot. The Chinese patent document number CN115937031A, the invention name of which is "A method for enhancing low-light images", discloses the perception mode of brightness and color information, and proposes the Retinex theory for low-light image enhancement. By enhancing the contrast and details of the image, the visual effect of the low-light image is significantly improved.

[0007] Chinese patent document number CN118487711A, invention name "A channel feedback deep learning autoencoder model based on full convolution", discloses a deep learning model based on an autoencoder, which can significantly improve the brightness and contrast of images in low-light environments. It is used to identify signal features from low-light images and adaptively adjust the image brightness to avoid over-enhancement or saturation of bright parts in high dynamic range images.

[0008] Low-light image enhancement is an active research field. The continuous development of related technologies has provided a variety of effective methods for improving image quality, and also provided technical support for application scenarios such as thermal defect detection of power equipment. Although the above-mentioned improved method for the YOLO model has improved certain capabilities of the model, the detection speed and accuracy in low-light environments are still unsatisfactory. Therefore, the present invention proposes a method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm to solve the problems existing in the prior art. Summary of the invention

[0009] In view of the above problems, the purpose of the present invention is to propose a method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm. The method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm uses an improved YOLOv8 algorithm to monitor abnormal conditions of equipment in the substation, thereby realizing real-time, fast and accurate target detection. According to the conditions of different illumination in different time periods, the application of a low-light enhancement algorithm to the model is proposed, which improves the generalization ability of the model and improves the accuracy while ensuring real-time detection.

[0010] To achieve the purpose of the present invention, the present invention is implemented by the following technical solution: a method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm, comprising the following steps:

[0011] Step 1: First collect substation equipment images under different illumination intensities, and use the mosaic data enhancement algorithm to preprocess the image data and optimize the data set;

[0012] Step 2: Use low-light enhancement technology to build an improved YOLOv8 model;

[0013] Step 3: Add the global attention module GAM to the YOLOv8 model to optimize the model accuracy;

[0014] Step 4: Obtain real-time substation equipment image data and use the improved model for real-time monitoring. Finally, evaluate the model using indicators such as precision and recall.

[0015] A further improvement is that in step one, mosaic data enhancement is used to randomly select four training images, then scale the four images and splice them into a new result image, and then the four images are combined and placed in four quadrants of the result image in sequence to form a result image set, that is, an optimized data set.

[0016] A further improvement is that the 3930 power equipment image data with a size of 640×640 collected in step 1 are divided into seven categories, namely, a training set, a test set and a validation set, wherein the training set contains 3435 image data, the test set contains 165 image data, and the validation set contains 330 image data.

[0017] Further improvements are: in step 2, low-light enhancement technology is used to focus on both the spatial features and channel features of the image to enhance contrast and detail performance; the low-light observation image is equal to the dot product of the clear image and the light, which is expressed by the following formula:

[0018]

[0019] where u t With x t Represent the residual and illumination of the tth stage, H m Represents the lighting estimation network.

[0020] Further improvement is that: the low light enhancement technology in step 2 introduces a self-correction module to reduce the amount of model calculation, and the self-correction module introduces the Retinex theory, which is expressed by the following formula:

[0021]

[0022] where v t is the calibrated input for the next stage, whereby the basic unit of the lighting optimization process is reformulated as

[0023] F(x t )→F(G(x t )).

[0024] Further improvements are as follows: in step three, a GAM global attention module is added to the neck network layer of the YOLOv8 model; the attention score of each input position is calculated through a feedforward neural network or a convolution operation, and the data features are processed by combining the CA module with the SA module.

[0025] Further improvements are: in step 4, the image data of the equipment in the substation is captured in real time by an infrared camera device; the evaluation indicators include precision, recall, F1 index, Score ) and mAP indicators, the formula is as follows:

[0026]

[0027] Among them, TP, FP, TN and FN represent the true positive samples predicted as positive by the model, the actual negative samples predicted as positive by the model, the actual negative samples predicted as negative by the model and the true positive samples predicted as negative by the model respectively; P represents the proportion of correctly detected positive samples to all predicted positive samples, R represents the proportion of correctly detected positive samples to all positive samples, and N is the number of objects; the mAP indicator value calculation formula first interpolates the precision-recall curve to calculate the difference precision, then calculates the area, and finally averages the areas under the subsequent precisions to obtain the mAP value.

[0028] The beneficial effects of the present invention are as follows: the present invention uses the improved YOLOv8 algorithm to monitor abnormal conditions of equipment in the substation, realizes real-time, fast and accurate target detection, and proposes the application of a low-light enhancement algorithm to the model for conditions with different illumination in different time periods, thereby improving the generalization ability of the model and improving the accuracy while ensuring real-time detection;

[0029] The method of the present invention enhances the data set, improves the ability of model feature recognition, and performs better in low-light environment detection. It excels in detection speed and accuracy, can effectively identify and locate abnormal power equipment in the substation, realize real-time detection of equipment, and provide reliable technical support for maintenance and fault diagnosis of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the present invention.

[0031] Figure 2 This is a network structure diagram of YOLOv8 of the present invention.

[0032] Figure 3 This is the improved YOLOv8 network structure diagram of the present invention.

[0033] Figure 4 This is a flow chart of the low-light image enhancement algorithm of the present invention.

[0034] Figure 5 Schematic diagram of the global attention mechanism GAM structure of the present invention.

[0035] Figure 6 This is a diagram of the detection results of the improved model of the present invention. DETAILED DESCRIPTION

[0036] In order to deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with examples. The examples are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0037] according to Figure 1-Figure 6 As shown, this embodiment provides a method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm, comprising the following steps:

[0038] Step 1: Collect substation equipment images under different illumination levels, and use the mosaic data enhancement algorithm to preprocess the image data and optimize the data set;

[0039] First, we collect substation equipment image data and use mosaic data augmentation technology to optimize the dataset, including selecting images: randomly selecting four images as source images for scaling and cropping. Each source image will be scaled to a different size and may be randomly cropped to fit the size of the final composite image; then performing image synthesis to merge the adjusted four images into a new image canvas. Usually, the four images are placed in the four quadrants of the new canvas; adjusting the object annotations in each source image to ensure that their position and size in the new image are correct, and finally obtaining an optimized dataset.

[0040] Step 2: Use low-light enhancement technology (SCINet) to build an improved YOLOv8 model;

[0041] A low-light image enhancement algorithm is used. The algorithm consists of a network composed of multiple convolutional layers and deconvolutional layers, including two sub-networks: one for estimating background brightness and the other for estimating the transmittance and reflectivity of objects. SCINet mainly optimizes these two sub-networks to find the optimal parameters suitable for the current scene and used for image illumination. In addition, it is also a network model designed specifically for low-light image enhancement. It processes image data by utilizing cascaded illumination learning steps and weight sharing mechanisms to achieve the purpose of optimizing image quality.

[0042] According to Retinex theory, the low-light observation image is equal to the dot product of the clear image and the illumination, as shown in the following formula:

[0043]

[0044] where u t With x t Represent the residual and illumination of the tth stage, H m represents the illumination estimation network;

[0045] The self-calibration module introduces physical laws (Retinex theory) to gradually correct the input of each stage to indirectly affect the output of each stage, thereby achieving convergence between each stage; the SCINet algorithm flow is as shown in the attached manual. Figure 1 shown.

[0046] Step 3: Add the global attention mechanism GAM to the model to improve model accuracy;

[0047] Using the global attention mechanism, GAM is an attention mechanism used in deep learning models, which aims to improve the performance of the model by capturing the global information of the input data; the core idea of ​​GAM is to give the model the ability to allocate attention based on the overall content of the input, rather than just relying on local features;

[0048] Calculate attention scores: Calculate the attention scores for each input position through a feedforward neural network or convolution operation; these scores reflect the importance of the position in the entire input, as shown in the appendix of the manual. Figure 2 As shown in the figure, GAM processes data features by combining the CA module with the SA module. This module adopts a new method to process channel and spatial information to improve the performance of deep neural networks.

[0049] In general, the global attention mechanism combines 3D permutation with multi-layer perceptron for channel attention, and also includes a convolutional spatial attention submodule. These components work together to reduce information loss and amplify global dimensional interaction features. The global attention mechanism processes the input feature data to obtain intermediate states and output states, where the channel attention map and spatial attention map are used to capture important features of the channel and spatial dimensions respectively. Through element-level multiplication operations, these attention maps interact with the input feature map to achieve global cross-dimensional information interaction.

[0050] From the perspective of network structure, YOLOv8 adopts the network structure idea of ​​YOLOv5, including the design of the backbone network and part of the neck network based on CSP, as well as the design for models of different sizes; in the backbone network layer, the C2f module is used instead of the C3 module to make the model further lightweight, while retaining the CSP structure and adopting the SPPF module; in the neck layer, the PAN idea is retained, the convolution structure in the upsampling is removed, and the C3 module is replaced by C2f. The head structure includes a detection head and a classification head, including convolution and deconvolution layers, and a global average pooling layer is used for classification detection; the detection head uses a Decoupled-Head structure to enhance the network reasoning speed and effect; VFL is used as the classification loss function, and DFL and CIOU are used as regression loss functions; the above improvements have enabled YOLOv8 to achieve its peak at debut, pursuing lightweight while achieving high precision.

[0051] Step 4: Obtain real-time substation equipment image data and use the improved model for real-time monitoring. Finally, evaluate the model using indicators such as precision and recall rate.

[0052] The image data of equipment in the substation is detected in real time. The experimental data of the present invention includes 3930 pictures in total. The picture size is 640x640, the initial learning rate is set to 0.01, the final learning rate is set to 0.01, the IOU is set to 0.75, the number of iterations is 300 rounds, the default optimizer is used, and the evaluation indicators include precision index, recall rate index, mAP index, and F1 score index.

[0053] This method for detecting abnormal equipment in substations based on low-light image enhancement algorithm integrates the low-light image enhancement algorithm SCINet into the YOLOv8s model structure, adopts the global attention mechanism GAM in the neck network, and finally uses low-light power equipment image data, and uses relevant evaluation indicators and experimental results to verify the advancement and superiority of the invention through visualization.

[0054] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm, characterized in that: The following steps are involved: Step 1: First collect substation equipment images under different illumination intensities, and use the mosaic data enhancement algorithm to preprocess the image data and optimize the data set; Step 2: Use low-light enhancement technology to build an improved YOLOv8 model; Step 3: Add the global attention module GAM to the YOLOv8 model to optimize the model accuracy; Step 4: Obtain real-time substation equipment image data and use the improved model for real-time monitoring. Finally, evaluate the model using indicators such as precision and recall.

2. According to claim 1, a method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm is characterized in that: In the step 1, mosaic data enhancement is used to randomly select four training images, then scale the four images and splice them into a new result image, and then the four images are sequentially combined and placed in four quadrants of the result image to form a result image set, that is, an optimized data set.

3. The method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm according to claim 1, characterized in that: The step 1 collects a total of 3930 power equipment image data of size 640×640 and is divided into seven categories, namely, a training set, a test set and a validation set, wherein the training set contains 3435 image data, the test set contains 165 image data, and the validation set contains 330 image data.

4. The method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm according to claim 1, characterized in that: In step 2, low-light enhancement technology is used to focus on both the spatial features and channel features of the image to enhance contrast and detail performance; the low-light observation image is equal to the dot product of the clear image and the light, which is expressed by the following formula where u t With x t Represent the residual and illumination of the tth stage, H m Represents the lighting estimation network.

5. The method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm according to claim 1, characterized in that: The low light enhancement technology in step 2 introduces a self-correction module to reduce the amount of model calculation. The self-correction module introduces the Retinex theory, which is expressed by the following formula: where v t is the calibrated input for the next stage, whereby the basic unit of the lighting optimization process is reformulated as F(x t )→F(G(x t ))。 6. The method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm according to claim 1, characterized in that: In the step 3, a GAM global attention module is added to the neck network layer of the YOLOv8 model; The attention score of each input position is calculated through a feedforward neural network or convolution operation, and the data features are processed by combining the CA module with the SA module.

7. The method for detecting abnormal equipment in a substation based on a low-light image enhancement algorithm according to claim 1, characterized in that: In the step 4, the image data of the equipment in the substation is captured in real time by an infrared camera device; the evaluation indicators include precision, recall rate, F1 index and mAP index, and the formula is as follows: Among them, TP, FP, TN and FN represent the true positive samples predicted as positive by the model, the actual negative samples predicted as positive by the model, the actual negative samples predicted as negative by the model and the true positive samples predicted as negative by the model respectively; P represents the proportion of correctly detected positive samples to all predicted positive samples, R represents the proportion of correctly detected positive samples to all positive samples, and N is the number of objects; the mAP indicator value calculation formula first interpolates the precision-recall curve to calculate the difference precision, then calculates the area, and finally averages the areas under the subsequent precisions to obtain the mAP value.

Images