A power grid defect device identification algorithm based on redundancy information decoupling

Abstract

The application discloses a kind of power grid defect equipment identification algorithm based on redundancy information decoupling, comprising the following steps: obtaining aerial transmission line image, making dataset;Depth learning model is constructed, data set is input into depth learning model, and power grid defect equipment identification is realized;Depth learning model is composed of feature extractor, redundancy information decoupling module, feature projector and classifier, feature extractor is used to the feature extraction of input data;Redundancy information decoupling module is used to the redundancy information decoupling of extracted feature, distinguish out the feature with discriminative and the feature without discriminative, and only the feature with discriminative is used for subsequent identification task;Feature projector and classifier are used to the classification identification of obtained feature.The application separates out redundancy information in data set using decoupling learning, only retains the feature with discriminative, speeds up the speed of model, and improves the identification accuracy of power equipment defect target.

Description

Technical Field

[0001] The present invention belongs to the fields of computer vision and information retrieval, and in particular relates to a power grid defective equipment identification algorithm based on redundant information decoupling. Background Art

[0002] The power system is an integrated system for producing and consuming electricity, comprising power plants, transmission and transformation lines, power distribution networks, and power user networks. Its function is to convert natural primary energy into electrical energy through power generation devices, and then supply this energy to users through transmission, transformation, and distribution, thus completing the entire process of electricity production, transmission, and consumption. Ensuring the reliability and stability of transmission is crucial for both national security and daily life. Therefore, reliability testing of power grid equipment is essential to ensure the stability of the entire power network.

[0003] my country's power grid is vast, and much of its equipment is located in remote, remote areas with complex terrain. Perpetually exposed to the elements, grid equipment is subject to natural hazards such as intense sunlight, strong winds, heavy rain, and snow. It also endures high voltages from electric shocks and lightning strikes, as well as internal and external stress fatigue damage and limited lifespans. Insulator breakage and disconnection, conductor breakage, and the presence of loose foreign objects can all cause power failures, leading to widespread power outages.

[0004] Therefore, during routine power inspections, rapid location of faulty equipment in the power grid helps quickly locate the fault point and analyze the cause, providing an important reference for subsequent power inspection and maintenance. However, this also places high demands on the accuracy and timeliness of locating faulty components in inspection images.

[0005] The traditional inspection model relies on manual inspections. However, since most power transmission networks are located in complex mountainous and forested terrain, this inspection model is labor-intensive and costly. Furthermore, inspection results are significantly affected by personnel skills and natural conditions such as weather, terrain, and sunlight. For transmission lines that are difficult to reach with manual inspections, drone inspections are a relatively effective method, optimizing the inspection process and addressing gaps in the existing system. Drone inspections of transmission lines are primarily designed to capture fault-prone components, promptly identifying defects and faults, and eliminating potential safety hazards. This effectively addresses the difficulties of manual inspections and improves their efficiency, quality, and effectiveness. Summary of the Invention

[0006] In response to the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a grid defective equipment identification algorithm based on redundant information decoupling. By effectively identifying aerial images of transmission lines, the subsequent maintenance efficiency is improved and maintenance costs are saved. Compared with manual visual inspection, it has the advantages of reliability, efficiency, and accuracy, and greatly reduces labor costs.

[0007] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0008] A grid defective device identification algorithm based on redundant information decoupling includes the following steps:

[0009] S1. Obtain aerial transmission line images and create a dataset.

[0010] S2. Build a deep learning model and input the dataset into the built deep learning model for training;

[0011] The constructed deep learning model consists of a feature extractor, a redundant information decoupling module, a feature projector and a classifier, wherein the feature extractor is used to extract features from the input data; the redundant information decoupling module is used to decouple the extracted features from redundant information, distinguish between discriminative features and non-discriminative features, and only use the discriminative features for subsequent recognition tasks; the feature projector and classifier are used to classify and identify the obtained features, thereby realizing the identification of defective power grid equipment.

[0012] Preferably, before making a data set, the pixel size of the aerial transmission line image needs to be scaled to limit the long side of the image to no more than 1000 and the short side to no more than 600, and the number of channels of the image is converted to a 3-channel color image to make a data set.

[0013] Preferably, during the feature extraction process, ResNet50, pre-trained on ImageNet, is used as a feature extractor to extract features from the input data. By inputting the input data X into ResNet50, the network will perform multi-level convolution and pooling operations on the image, gradually extracting abstract features of the image. These features not only include low-level edge and texture information, but also higher-level semantic information, such as the shape, structure, and spatial relationship of the object.

[0014] Preferably, in the process of decoupling the extracted features, features directly related to the equipment defect category are extracted and strengthened, and decoupled from the background information, so that the model can more accurately judge and locate defects; the image processed by the decoupling method only retains information related to the equipment defect, while eliminating complex background interference, so that the model can more focusedly and accurately identify defects, and the two parts obtained are respectively features with human discriminability and features without human discriminability.

[0015] Preferably, in step S2, the feature f extracted by the feature extractor E is decoupled into two parts: the feature f with character discrimination u and features f that are not person-discriminative n , specifically expressed as follows:

[0016] f=f u +f n

[0017] In order to ensure that f u With f n Be separated as much as possible, the present invention adopts loss per This makes the two tend to be orthogonal in the representation space, loss per The formula is as follows:

[0018] loss per =|cos(f u -f n )|

[0019] Among them, cos represents cosine distance, |·| represents absolute value, and loss per Continuous constraints to achieve f u With f n The purpose of the cosine distance between them gradually tends to zero, and the feature distributions of the two parts tend to be orthogonal, which means that f u With f n decoupled;

[0020] Adopt a new function loss inv To constrain, function loss inv is defined as follows:

[0021]

[0022] Among them, n is the number of categories in an ibatch, m is the number of samples in the same category, and Indicates the discriminative features among all sample features in the same category. represents the two-norm.

[0023] Preferably, in step S2, the feature projector and classifier are used to classify and identify the obtained features, and a feature projector composed of a double-layer linear is used to classify the discriminative feature f u Mapped to the projected feature υ:

[0024] υ=P(f u )

[0025] Then a classifier is introduced to classify the projected feature υ and obtain the predicted label pred:

[0026] pred=C(υ)

[0027] In order to evaluate the accuracy of the model, the present invention compares the predicted label pred with the true label L and calculates the difference between them using the cross entropy loss function, as follows:

[0028] loss ce =-∑L×log(pred)

[0029] Thus, the loss function of the entire model is written as:

[0030] loss=loss d +loss ce Through multiple training iterations, the model is able to continuously update itself under the optimization of the loss function. In each iteration, the model adjusts its parameters based on the gradient information of the loss function through the backpropagation algorithm to minimize the difference between the predicted label and the true label.

[0031] The present invention has the following beneficial effects:

[0032] (1) Considering that even if the collected defect data express defective devices of the same category, the locations and backgrounds of the defective devices are different, which greatly increases the difficulty of identification, the present invention provides a grid defective device identification algorithm based on redundant information decoupling. By separating the above-mentioned redundant information that is not conducive to the identification task, only the features with the discriminativeness of the defective devices are retained, thereby achieving more effective identification. By separating the discriminative part of the data features, the model strengthens the distinction between different grid defective device categories, thereby improving the accuracy of identification.

[0033] (2) The present invention integrates the advantages of decoupling learning and deep learning algorithms. By utilizing the advantages of decoupling learning, it separates redundant information from the data set and retains only discriminative features, greatly accelerating the overall speed of the model and improving the accuracy of identifying power equipment defect targets. It solves the problem of low speed and limited accuracy of manual detection of grid components and defects. In addition, the present invention has a good recognition effect on a variety of different power equipment defects and anomalies. It can effectively replace the process of manual identification of abnormal defects in practical applications, greatly reducing the consumption of manpower and material resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 A schematic diagram of the process framework of a power grid defective device identification algorithm based on redundant information decoupling provided by the present invention;

[0036] Figure 2 Schematic diagram of the feature decoupling module;

[0037] Figure 3 It is the loss-epoch curve;

[0038] Figure 4 This is a curve of the change of accuracy ACC(%) with epoch. DETAILED DESCRIPTION

[0039] In the following description, specific details such as specific system structures and technologies are provided for the purpose of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may also be implemented in other embodiments without these specific details.

[0040] Example 1

[0041] For example, a dataset of drone-generated images of power transmission equipment in .jpg format includes three categories: conductors, insulators, and floating foreign objects. The experimental hardware configuration includes an Ubuntu 20.04 operating system, 7.7GB of RAM, and an NVIDIA 3080Ti graphics card. The software environment uses the PyTorch deep learning framework, running on a CUDA 10.3 architecture platform for parallel computing.

[0042] An embodiment of the present invention provides a grid defective device identification algorithm based on redundant information decoupling, comprising the following steps:

[0043] S1. Obtain aerial transmission line images and create a dataset.

[0044] Specifically, aerial transmission line images are obtained. In order to ensure the universality of the model, before model training, the present invention scales the pixel sizes of all aerial transmission line image data, limiting the long side of the image to no more than 1000 and the short side to no more than 600, and converts the number of channels of the image into a 3-channel color image to produce a data set.

[0045] S2. Build a deep learning model and input the data set into the built deep learning model for training to identify defective devices in the power grid;

[0046] The deep learning model constructed in the present invention is composed of a feature extractor, a redundant information decoupling module, a feature projector and a classifier. Figure 1 As shown in the figure, the present invention first uses a feature extractor to extract features from the input data. Then, it decouples redundant information from the extracted features to distinguish between discriminative features and non-discriminative features, and only uses the discriminative features for subsequent recognition tasks. Finally, it uses a feature projector and a classifier to classify and identify the obtained features, thereby realizing the identification of defective devices in the power grid.

[0047] (1) In the above-mentioned feature extraction process, the present invention uses ResNet50 pre-trained on ImageNet as a feature extractor for effective feature extraction of the input data. ResNet50 is a deep residual network that has been trained on a large number of image recognition tasks and has become sensitive to various visual features and can effectively capture important information in the image. By inputting the input data X into ResNet50, the network will perform multi-level convolution and pooling operations on the image, gradually extracting the abstract features of the image. These features not only contain the underlying edge and texture information, but also contain higher-level semantic information, such as the shape, structure and spatial relationship of the object. By using ResNet50, the present invention can make full use of its deep convolutional neural network structure and training experience on large-scale image databases, thereby obtaining features f with high representation capabilities:

[0048] f=E(X)

[0049] Because ResNet50 is pre-trained on the large-scale ImageNet dataset, it has already learned visual feature representations for many common objects and scenes. This approach allows a wide range of tasks and domains to benefit from ResNet50's powerful feature extraction capabilities. By using a pre-trained ResNet50 as the feature extractor E, the present invention ensures that the features f extracted from the input data X have high discriminability and generalization capabilities, thereby providing strong support for subsequent tasks (such as classification, detection, or recognition).

[0050] (2) Decouple the redundant information of the extracted features, distinguish between discriminative features and non-discriminative features, and only use the discriminative features for subsequent recognition tasks.

[0051] Because drones face different weather conditions and environments when collecting images of defective power grid equipment, the images obtained vary greatly. These differences may be due to factors such as lighting conditions, weather conditions, and shooting angles. In the case of the same equipment defect category, the information contained in the image not only represents the defect category, but also covers other information related to the scene. This additional information may include the surrounding environment, surrounding structures of the equipment, and other features. Since the equipment in the image is not pre-positioned and segmented, the model will be interfered with by complex background information during the processing process. This complex background information may include other power grid components, trees, buildings, etc., which are mixed with the characteristics of the equipment defects, making it difficult for the model to accurately distinguish and identify the defects.

[0052] In order to solve the above problems, the present invention adopts a decoupling method to remove redundant information in image features, such as Figure 2 As shown, only discriminative key information is retained, thereby enhancing the model's ability to distinguish different grid defective devices. Specifically, this method extracts and strengthens features directly related to the device defect category, decoupling them from background information, enabling the model to more accurately judge and locate defects.

[0053] The decoupling method processes images, retaining only information related to equipment defects while eliminating complex background interference, enabling the model to more accurately and focusedly identify defects. This approach can improve the effectiveness of drones in power grid defect detection, providing the power industry with a reliable and efficient equipment monitoring and maintenance solution.

[0054] Specifically, the present invention decouples the feature f extracted by the feature extractor E into two parts: the feature f with character discrimination u and features f that are not person-discriminative n , specifically expressed as follows:

[0055] f=f u +f n

[0056] The present invention only uses f u They can perform recognition tasks because they have the ability to distinguish different defective devices and can contribute to the current recognition task, while f n The features do not have any beneficial effect on the current recognition task.

[0057] 1. Feature Orthogonality Loss

[0058] In order to ensure that f u With f n Be separated as much as possible, the present invention adopts loss per This makes the two tend to be orthogonal in the representation space, loss per The formula is as follows:

[0059] loss per =|cos(f u -f n )|

[0060] Among them, cos represents cosine distance, |·| represents absolute value, and loss per Continuous constraints to achieve f u With f n The purpose of the cosine distance between them gradually tends to zero, and the feature distributions of the two parts tend to be orthogonal, which means that f u With f n Decoupled.

[0061] 2. Inter-class similarity loss

[0062] The above formula does not have a standard f u With f n What are the specific shapes respectively. u With f n When it is not defined, the model cannot distinguish which part of the information is f u , which part is f n . In this way, it is impossible to decouple the redundant information. Therefore, the present invention adopts a new function loss inv To constrain, function loss inv is defined as follows:

[0063]

[0064] Among them, n is the number of categories in an ibatch, m is the number of samples in the same category, and Indicates the discriminative features among all sample features in the same category. represents the two-norm.

[0065] Function loss inv Based on the f of the same category u The calculation is based on the features, because f u There is a high degree of similarity between the same categories, indicating the same defective device. Ideally, in the case of the same defective device type, f u Remain consistent and not affected by changes in redundant information such as background, while f n This function can be regarded as a regularization term, which reduces the average f under the same label. n The distance between features is used to strengthen the similarity between them.

[0066] The overall loss function loss of the redundant information decoupling part d It can be written as:

[0067] loss d =λ d (loss per +loss inv )

[0068] In loss d With the optimization, redundant information can be decoupled and will not be used in subsequent recognition tasks. The model can effectively capture the discriminative information of different defective devices.

[0069] (3) The obtained features are classified and identified using feature projectors and classifiers, thereby realizing the identification of defective equipment in the power grid.

[0070] In order to ensure that the model has good recognition ability, the present invention adopts a feature projector P composed of a double-layer linear to project the discriminative feature f u Mapped to the projected feature υ:

[0071] υ=P(f u )

[0072] Then the present invention introduces a classifier to classify the projected feature υ and obtain the predicted label pred:

[0073] pred=C(υ)

[0074] In order to evaluate the accuracy of the model, the present invention compares the predicted label pred with the true label L and calculates the difference between them using the cross entropy loss function, as follows:

[0075] loss ce=-∑L×log(pred)

[0076] The cross entropy loss function is a commonly used loss function that measures the difference between the predicted results and the actual results and provides a direction for model training. The overall loss function of the model can be written as:

[0077] loss=loss d +loss ce

[0078] Through multiple training iterations, the model continuously updates itself while optimizing the loss function. In each iteration, the model uses the backpropagation algorithm to adjust parameters based on the gradient of the loss function to minimize the difference between the predicted label and the true label. This iterative training process allows the model to gradually learn more accurate feature representations and classification decision rules, thereby improving its recognition capabilities.

[0079] Experimental results

[0080] As the number of training iterations increases, the grid equipment defect recognition model in the present invention is gradually optimized, and the loss function gradually converges to a stable level, such as Figure 3 By continuously adjusting the model parameters, the present invention can make the performance of the model reach the optimal state.

[0081] The present invention selects the accuracy rate ACC as an indicator to evaluate the performance of the model. Accuracy rate is an important indicator to measure the accuracy of model classification. It indicates the proportion of samples correctly classified by the model on the test data set. A high accuracy rate means that the model can accurately distinguish different defect samples, thereby providing reliable defect detection results. After multiple experiments and training, the model of the present invention can achieve an accuracy rate of 93.37%. Figure 4 As shown, Figure 4 The results show that the model achieves good results in defect identification.

[0082] In addition, we also compared the algorithm of the present invention with other recognition algorithms, as shown in Table 1. From the results in Table 1 below, it can be seen that the algorithm of the present invention achieves the best recognition effect.

[0083] Table 1

[0084]

[0085] Through the training and evaluation process of this invention, the model has demonstrated good effectiveness and superiority in identifying defects in power grid equipment. This provides important technical support for automated defect identification in related fields and is expected to play a significant role in practical applications.

[0086] The present invention is not limited to the above-mentioned specific implementation methods. Various changes made by ordinary technicians in this field based on the above-mentioned concept without creative work are all within the scope of protection of the present invention.

Claims

1. A grid defective equipment identification algorithm based on redundant information decoupling, characterized by: The following steps are involved: S1. Obtain aerial transmission line images and create a dataset. S2. Build a deep learning model and input the data set into the built deep learning model for training to identify defective devices in the power grid; The constructed deep learning model consists of a feature extractor, a redundant information decoupling module, a feature projector and a classifier, wherein the feature extractor is used to extract features from the input data; the redundant information decoupling module is used to decouple the redundant information of the extracted features, distinguish between discriminative features and non-discriminative features, and only use the discriminative features for subsequent recognition tasks; the feature projector and classifier are used to classify and identify the obtained features; The features extracted by feature extraction Decoupled into two parts: discriminative features and non-discriminative features , specifically expressed as follows: To ensure and are separated as much as possible, the present invention adopts This makes the two tend to be orthogonal in the representation space. The formula is as follows: Among them, cos represents the cosine distance, Represents the absolute value, through The constant constraints to achieve and The purpose of the cosine distance between them gradually tends to zero, and the feature distributions of the two parts tend to be orthogonal, which means that and decoupled; Adopt a new function To constrain, function is defined as follows: Among them, n is the number of categories in an ibatch, m is the number of samples in the same category, and Indicates the discriminative features among all sample features in the same category. represents the two-norm; In step S2, the feature projector and classifier are used to classify and identify the obtained features. A feature projector composed of a double-layer linear is used to classify the discriminative feature f u Mapping to projection features : Then introduce a classifier to project the features Perform classification and get the predicted label : In order to evaluate the accuracy of the model, the present invention predicts the label Compare it with the true label L and calculate the difference between them using the cross entropy loss function, as follows: Thus we get the loss function of the entire model Written as: The overall loss function of the redundant information decoupling part Written as: Through multiple training iterations, the model can continuously update itself under the optimization of the loss function. In each iteration, the model adjusts the parameters according to the gradient information of the loss function through the backpropagation algorithm to minimize the difference between the predicted label and the true label.

2. The grid defective device identification algorithm based on redundant information decoupling according to claim 1 is characterized in that: Before inputting into the model, the pixel size of the aerial transmission line image data was scaled to limit the image length to no more than 1000 and the short side to no more than 600. The image channels were converted into a 3-channel color image to create a data set.

3. The grid defective device identification algorithm based on redundant information decoupling according to claim 1 is characterized in that: In the feature extraction process, ResNet50 pre-trained on ImageNet is used as a feature extractor to extract features of the input data. By inputting the input data into ResNet50, the network will perform multi-level convolution and pooling operations on the image, and gradually extract the abstract features of the image.

4. The grid defective device identification algorithm based on redundant information decoupling according to claim 1 is characterized in that: The abstract features extracted from the image include bottom-level edges, texture information, and semantic information.

5. The grid defective device identification algorithm based on redundant information decoupling according to claim 1 is characterized in that: In step S2, in the process of decoupling redundant information of the extracted features, features directly related to the equipment defect category are extracted and enhanced to decouple them from background information.

Images