Image classification method based on semantic knowledge base in smart power grid

Abstract

The invention relates to the technical field of power grid image classification, in particular to an image classification method based on a semantic knowledge base in a smart power grid, and the method comprises the following operation steps: S1, data collection: 1.1, collecting image data from various devices (such as a monitoring camera and unmanned aerial vehicle inspection equipment) of the smart power grid; the images may include the appearance of the power equipment, the operation state (such as the oil temperature indication of a transformer and the on-off state of a switch), and the integrity of the line. 1.2, preliminarily preprocessing the acquired images, including unifying the format of the images (for example, converting the images into a JPEG (Joint Photographic Experts Group) format), adjusting the size (enabling the images to adapt to a subsequent processing algorithm, for example, uniformly adjusting the images into 512 * 512 pixels) and removing noise (removing salt and pepper noise in the images by using a filtering technology, for example, Gaussian filtering); according to the method, the feature extraction process can be guided through the semantic knowledge base, so that the model focuses on key features related to semantics, and the classification precision can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid image classification, and in particular to an image classification method based on a semantic knowledge base in a smart grid. Background Art

[0002] With the development of the national economy and the continuous improvement of people's living standards, the scale of power grid investment has continued to expand, the network structure has become more and more complex, and the reliable operation of power grid lines and power grid equipment has directly affected the production safety and social benefits of power companies and the stability of the power grid. In order to ensure the normal operation of power facilities, the power department must formulate an inspection plan every month and assign personnel to inspect power supply facilities including poles, wires, transformers, capacitors, etc., so as to timely discover equipment defects and safety hazards, and summarize and regularly analyze and count the equipment's operating conditions and defect information. However, smart grids contain a large number of complex power equipment, such as transformers, circuit breakers, insulators, etc. The normal operation of these equipment is crucial to the stable supply of the power system. The traditional manual inspection and classification method is inefficient, labor-intensive, and prone to missed inspections. Therefore, we propose an image classification method based on semantic knowledge base in smart grids to solve the above-mentioned problems. Summary of the invention

[0003] The purpose of the present invention is to solve the shortcomings of the background technology and propose an image classification method based on a semantic knowledge base in a smart grid.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is: an image classification method based on a semantic knowledge base in a smart grid, comprising the following steps:

[0005] S1. Data collection:

[0006] 1.1. Collect image data from various devices in the smart grid (such as surveillance cameras and drone inspection equipment); these images may include the appearance of power equipment, operating status (such as oil temperature indication of transformers, opening and closing status of switches), and line integrity;

[0007] 1.2. Perform preliminary preprocessing on the collected images, including unifying the image format (such as converting them to JPEG format), resizing (to adapt them to subsequent processing algorithms, such as uniformly adjusting them to 512×512 pixels), and removing noise (using filtering technology, such as Gaussian filtering to remove salt and pepper noise in the image);

[0008] S2. Construction of semantic knowledge base:

[0009] 2.1. Definition of terms: collect professional terms in the field of smart grid, such as the names of various power equipment ("transformer", "insulator", "switch"), equipment status ("normal operation", "fault", "overheat"), and related operation and environment concepts ("under maintenance", "lightning strike");

[0010] 2.2. Establishing semantic relationships: define the relationships between terms. For example, the relationship between "transformer" and "overheating" may be "transformer may have overheating fault"; the relationship between "insulator" and "breakage" is "insulator breakage is a fault state". These relationships can be determined through domain expert knowledge, historical fault records and relevant standard documents;

[0011] 2.3. Knowledge representation: Use appropriate knowledge representation methods, such as ontology languages ​​(such as OWL-WebOntologyLanguage) to build a semantic knowledge base. Taking transformers as an example, the categories of transformers (oil-immersed transformers, dry-type transformers), their attributes (rated power, voltage level), and their associations with fault states (such as short circuit, overload) can be defined in the ontology.

[0012] S3. Feature extraction:

[0013] 3.1. Use deep learning models (such as convolutional neural networks (CNNs)) to extract image features; use classic VGG-16 and ResNet models; input the preprocessed images into these models, and the models will automatically extract the texture and shape features of the images;

[0014] 3.2. For smart grid images, we focus on the specific features of power equipment. Taking substation equipment images as an example, we may focus on extracting the oil level gauge position of the transformer, the contact shape of the switch, and the number of insulator strings. These feature vectors will serve as an important basis for subsequent classification.

[0015] S4. Classification Model

[0016] 4.1. Semantic-based classification rule construction: Classification rules are constructed by combining the knowledge in the semantic knowledge base and the extracted image features. If the image features show that the temperature of a specific area of ​​a device is too high, and the semantic knowledge base contains the knowledge that "overtemperature of the device may cause failure", then the image can be classified as an image of a device that may have a failure.

[0017] 4.2. Classification algorithm selection: You can use traditional classification algorithms such as support vector machine (SVM) and decision tree, or you can use classifiers in deep learning (such as fully connected neural network); input the extracted image features into the classification model, and the model classifies the image according to pre-built classification rules, for example, into "normal equipment", "faulty equipment-overheating", and "faulty equipment-mechanical damage" categories.

[0018] Preferably, the updating and maintenance of the semantic knowledge base in step S2 includes the following specific steps:

[0019] (1) Regularly update the semantic knowledge base to adapt to the development of smart grid technology and the introduction of new equipment. When a new smart meter model is put into use, its related appearance features and normal / fault status knowledge need to be added to the semantic knowledge base.

[0020] (2) Use data mining technology to extract useful semantic information from new fault cases and maintenance records, and correct and improve the semantic relationships in the knowledge base.

[0021] Preferably, the feature extraction optimization in step S3 includes the following specific steps:

[0022] (1) Optimize feature extraction based on the content of the semantic knowledge base. If the semantic knowledge base emphasizes the key impact of the state of a specific component of a certain power equipment on the overall operation of the equipment, then the image features of the component can be extracted and enhanced during the feature extraction process.

[0023] (2) Using a multimodal feature extraction method, in addition to the visual features of the image, it can also combine the features of other sensor data (such as temperature sensors and vibration sensors); for the image classification of a motor device, in addition to extracting the image features of the motor's appearance, it can also combine the temperature characteristics of the motor during operation to more accurately determine whether the motor is in a normal state.

[0024] Preferably, the classification model training and evaluation in step S4 includes the following specific steps:

[0025] (1) Use the labeled smart grid image dataset to train the classification model; the labeling information can be determined based on the actual equipment status inspection results, such as marking as "normal transformer" or "faulty switch (mechanical failure)" after inspection by professionals;

[0026] (2) Use the cross-validation method to evaluate the performance of the classification model; divide the data set into training set, validation set and test set, use the validation set to adjust the model parameters during the training process, and finally use the test set to evaluate the model's accuracy and recall performance indicators; if the model performance does not meet the requirements, you can adjust the model structure (such as increasing the number of neural network layers) or optimize the training algorithm (such as adjusting the learning rate).

[0027] Preferably, the application and feedback of the classification results in step S4 includes the following specific steps:

[0028] (1) Apply the classified image results to the operation and management of the smart grid; for images classified as faulty equipment, promptly notify maintenance personnel to conduct repairs; for images of normal equipment, they can be saved as records of normal equipment operation;

[0029] (2) Further improve the classification method based on feedback from actual applications. When maintenance personnel find misjudgments in the classification results, they analyze whether it is an error in the semantic knowledge base, inaccurate feature extraction, or a problem with the classification model, and then make targeted improvements.

[0030] Preferably, the convolution operation in the convolutional neural network model in step S3 can be expressed by the following formula:

[0031] For an input feature map X (size H in ×W in ×C in , where H in is the height, W in is the width, C in is the number of channels), using a convolution kernel (also called a filter, with a size of h×w×C in , h and w are the height and width of the convolution kernel respectively) to perform convolution operation and output a feature map (size is H out ×W out ×C out .

[0032]

[0033] Among them, i and j represent the position coordinates (height and width) of the output feature map, k represents the output channel index, K is the weight parameter of the convolution kernel, and b k is the offset term of the corresponding channel; H out , W out , C out The calculation of is usually as follows:

[0034]

[0035] Among them, padding represents the number of pixels filled at the edge of the input feature map, and stride represents the step size of the convolution kernel sliding.

[0036] Preferably, in the model establishment in the S3 step, Sigmoid function, Tanh function, ReLU function, LeakyReLU function and ELU function can be used for activation.

[0037] Preferably, the activation step includes the following steps:

[0038] (1) Neuron input calculation: First, for each layer in the neural network (assuming it is a fully connected layer), the neuron input is obtained by multiplying the output of the previous layer by the connection weight and summing them up;

[0039] (2) Application of activation function: After calculating the input of the neuron, use it as the input of the activation function to obtain the output of the neuron (the value after activation);

[0040] (3) Calculate layer by layer: Following the above steps, starting from the input layer, calculate the activated output of each layer of neurons in turn until the final output of the network is obtained. When a simple multi-layer perceptron (MLP) is used for image classification, the input layer receives the feature vector of the image (after preprocessing and feature extraction), and after the activation calculation of the hidden layer, the probability distribution of the classification is finally obtained at the output layer (if the Softmax activation function is used in the output layer, it is used for multi-classification tasks);

[0041] (4) Activation steps in the back propagation phase:

[0042] (4.1) Calculate the gradient of the activation function. During the back-propagation process, the derivative of the activation function is calculated to update the weights.

[0043] (4.2) Calculate the gradient of the loss function with respect to the weights. Use the chain rule to combine the gradient of the activation function and the gradient of the loss function with respect to the output to calculate the gradient of the loss function with respect to the weights.

[0044] (4.3) Weight update: Based on the calculated gradient, an optimization algorithm (such as stochastic gradient descent, SGD) is used to update the weights.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention integrates professional knowledge in the field of smart grid into the image classification process by constructing a semantic knowledge base. When identifying power equipment faults, the semantic knowledge base contains semantic descriptions of the normal state and various fault states of the equipment. Combined with image feature extraction, the classification model can more accurately judge the equipment status. When the image features show that the temperature of a certain part of the equipment is abnormally high, the knowledge about equipment overheating faults in the semantic knowledge base can assist the model to make a more accurate "faulty equipment-overheating" classification, rather than relying solely on the visual features of the image for guessing, thereby reducing misjudgment and improving classification accuracy.

[0047] 2. The semantic knowledge base can guide the feature extraction process and make the model focus on key features related to semantics. Taking the switch equipment in the substation as an example, the knowledge base emphasizes that the state of the switch contacts is crucial to the operation of the equipment. In the feature extraction stage, the image features of the switch contact part can be extracted, such as the shape and spacing of the contacts. In this way, the extracted features are more representative and discriminative, which helps to improve the classification accuracy.

[0048] 3. Unlike traditional black-box image classification models, methods based on semantic knowledge bases can provide a certain degree of interpretability. At the same time, the semantic knowledge base can be easily updated to adapt to these changes. For example, when a new type of smart meter is introduced, its appearance characteristics, normal operating parameters, possible failure modes and other knowledge can be added to the semantic knowledge base. This updated knowledge can immediately guide the model to classify new device images, allowing the model to quickly adapt to new application scenarios without the need to retrain the entire model.

[0049] 4. In the smart grid system, multiple types of image data may be involved, such as substation images of different voltage levels, inspection images of different lines, etc. The classification method based on the semantic knowledge base can be easily extended to these new image types. As long as the device knowledge and semantic relationships corresponding to the new type of image are added to the knowledge base, it can be expanded on the basis of the existing model framework to realize the classification of multiple images, thereby improving the overall scalability of the system.

[0050] 5. In actual use, the semantic knowledge base can assist in mining more valuable information from image data. By analyzing a large amount of classified image data and combining the knowledge in the semantic knowledge base, some potential equipment failure patterns or operating trends can be discovered. By analyzing the classification results of faulty equipment images over a period of time and the corresponding semantic descriptions, it may be found that some equipment is more prone to certain failures in specific seasons or operating conditions, which provides a basis for preventive maintenance of smart grids and improves the utilization value of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention is a schematic diagram of a classification process of an image classification method based on a semantic knowledge base in a smart grid. DETAILED DESCRIPTION

[0052] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0053] like Figure 1 The image classification method based on a semantic knowledge base in a smart grid shown in the figure includes the following steps:

[0054] S1. Data collection:

[0055] 1.1. Collect image data from various devices in the smart grid (such as surveillance cameras and drone inspection equipment); these images may include the appearance of power equipment, operating status (such as oil temperature indication of transformers, opening and closing status of switches), and line integrity;

[0056] 1.2. Perform preliminary preprocessing on the collected images, including unifying the image format (such as converting them to JPEG format), resizing (to adapt them to subsequent processing algorithms, such as uniformly adjusting them to 512×512 pixels), and removing noise (using filtering technology, such as Gaussian filtering to remove salt and pepper noise in the image);

[0057] S2. Construction of semantic knowledge base:

[0058] 2.1. Definition of terms: collect professional terms in the field of smart grid, such as the names of various power equipment ("transformer", "insulator", "switch"), equipment status ("normal operation", "fault", "overheat"), and related operation and environment concepts ("under maintenance", "lightning strike");

[0059] 2.2. Establishing semantic relationships: define the relationships between terms. For example, the relationship between "transformer" and "overheating" may be "transformer may have overheating fault"; the relationship between "insulator" and "breakage" is "insulator breakage is a fault state". These relationships can be determined through domain expert knowledge, historical fault records and relevant standard documents;

[0060] 2.3. Knowledge representation: Use appropriate knowledge representation methods, such as ontology languages ​​(such as OWL-WebOntologyLanguage) to build a semantic knowledge base. Taking transformers as an example, the categories of transformers (oil-immersed transformers, dry-type transformers), their attributes (rated power, voltage level), and their associations with fault states (such as short circuit, overload) can be defined in the ontology.

[0061] S3. Feature extraction:

[0062] 3.1. Use deep learning models (such as convolutional neural networks (CNNs)) to extract image features; use classic VGG-16 and ResNet models; input the preprocessed images into these models, and the models will automatically extract the texture and shape features of the images;

[0063] 3.2. For smart grid images, we focus on the specific features of power equipment. Taking substation equipment images as an example, we may focus on extracting the oil level gauge position of the transformer, the contact shape of the switch, and the number of insulator strings. These feature vectors will serve as an important basis for subsequent classification.

[0064] S4. Classification Model

[0065] 4.1. Semantic-based classification rule construction: Classification rules are constructed by combining the knowledge in the semantic knowledge base and the extracted image features. If the image features show that the temperature of a specific area of ​​a device is too high, and the semantic knowledge base contains the knowledge that "overtemperature of the device may cause failure", then the image can be classified as an image of a device that may have a failure.

[0066] 4.2. Classification algorithm selection: You can use traditional classification algorithms such as support vector machine (SVM) and decision tree, or you can use classifiers in deep learning (such as fully connected neural network); input the extracted image features into the classification model, and the model classifies the image according to pre-built classification rules, for example, into "normal equipment", "faulty equipment-overheating", and "faulty equipment-mechanical damage" categories.

[0067] The updating and maintenance of the semantic knowledge base in step S2 includes the following specific steps:

[0068] (1) Regularly update the semantic knowledge base to adapt to the development of smart grid technology and the introduction of new equipment. When a new smart meter model is put into use, its related appearance features and normal / fault status knowledge need to be added to the semantic knowledge base.

[0069] (2) Use data mining technology to extract useful semantic information from new fault cases and maintenance records, and modify and improve the semantic relationships in the knowledge base;

[0070] Furthermore, in the specific implementation, by constructing a semantic knowledge base, professional knowledge in the field of smart grid is integrated into the image classification process. When identifying power equipment failures, the semantic knowledge base contains semantic descriptions of the normal state and various fault states of the equipment. Combined with image feature extraction, the classification model can more accurately judge the equipment status. When the image features show that the temperature of a certain part of the equipment is abnormally high, the knowledge about equipment overheating failure in the semantic knowledge base can assist the model to make a more accurate "faulty equipment-overheating" classification, rather than relying solely on the visual features of the image to make guesses, thereby reducing misjudgments and improving classification accuracy.

[0071] The feature extraction optimization in step S3 includes the following specific steps:

[0072] (1) Optimize feature extraction based on the content of the semantic knowledge base. If the semantic knowledge base emphasizes the key impact of the state of a specific component of a certain power equipment on the overall operation of the equipment, then the image features of the component can be extracted and enhanced during the feature extraction process.

[0073] (2) Using multimodal feature extraction methods, in addition to the visual features of the image, the features of other sensor data (such as temperature sensors and vibration sensors) can also be combined; for the image classification of a motor device, in addition to extracting the image features of the motor appearance, the temperature features of the motor during operation can also be combined to more accurately determine whether the motor is in a normal state;

[0074] Furthermore, in specific implementation, the semantic knowledge base can guide the feature extraction process, so that the model can focus on key features related to semantics. Taking the switch equipment in the substation as an example, the knowledge base emphasizes that the state of the switch contacts is crucial to the operation of the equipment. In the feature extraction stage, the image features of the switch contact part can be extracted, such as the shape and spacing of the contacts. In this way, the extracted features are more representative and distinctive, which helps to improve the classification accuracy.

[0075] The classification model training and evaluation in step S4 includes the following specific steps:

[0076] (1) Use the labeled smart grid image dataset to train the classification model; the labeling information can be determined based on the actual equipment status inspection results, such as marking as "normal transformer" or "faulty switch (mechanical failure)" after inspection by professionals;

[0077] (2) Use the cross-validation method to evaluate the performance of the classification model; divide the data set into training set, validation set and test set, use the validation set to adjust the model parameters during the training process, and finally use the test set to evaluate the model's accuracy and recall performance indicators; if the model performance does not meet the requirements, you can adjust the model structure (such as increasing the number of neural network layers) or optimize the training algorithm (such as adjusting the learning rate);

[0078] The application and feedback of the classification results in step S4 include the following specific steps:

[0079] (1) Apply the classified image results to the operation and management of the smart grid; for images classified as faulty equipment, promptly notify maintenance personnel to conduct repairs; for images of normal equipment, they can be saved as records of normal equipment operation;

[0080] (2) Further improve the classification method based on feedback from actual applications. When maintenance personnel find misjudgments in the classification results, they analyze whether it is an error in the semantic knowledge base, inaccurate feature extraction, or a problem with the classification model, and then make targeted improvements.

[0081] Furthermore, in specific implementation, the classification model learns the general feature representation of the image through a large amount of training data. In the smart grid environment, the model can adapt to changes in different lighting conditions, shooting angles, equipment models, etc. The model can learn general features such as the basic shape of the transformer and the location of main components, thereby accurately classifying. This generalization capability enables the model to play a role in different substation scenarios or inspection routes without the need for retraining for each specific scenario. At the same time, the semantic knowledge base further enhances the generalization ability of the model. When encountering new equipment types or states, semantic knowledge can help the model make analogies and reasoning.

[0082] The convolution operation in the convolutional neural network model in step S3 can be expressed by the following formula:

[0083] For an input feature map X (size H in ×W in ×C in , where H in is the height, W in is the width, C in is the number of channels), using a convolution kernel (also called a filter, with a size of h×w×C in , h and w are the height and width of the convolution kernel respectively) to perform convolution operation and output a feature map (size is H out ×W out ×C out .

[0084]

[0085] Among them, i and j represent the position coordinates (height and width) of the output feature map, k represents the output channel index, K is the weight parameter of the convolution kernel, and b k is the offset term of the corresponding channel; H out , W out , C out The calculation of is usually as follows:

[0086]

[0087] Among them, padding represents the number of pixels filled at the edge of the input feature map, and stride represents the step size of the convolution kernel sliding;

[0088] Furthermore, in specific implementation, the convolutional neural network (CNN) has multiple convolutional layers and pooling layers, which can automatically learn different levels of features from the original image data. In the classification of smart grid images, CNN can first learn low-level features of the image, such as edges, lines and simple textures, and can identify the edges of the transformer casing, the outline of the insulator, etc. As the number of network layers increases, it can learn more advanced semantic features, such as the complex structural features of the internal windings of the transformer or the state features of the switch device (such as the degree of opening and closing). There is no need to manually design feature extractors, which greatly reduces the workload and can mine complex features that are difficult for humans to define and extract. In actual use, smart grid images contain a variety of complex power equipment and scenes. The automatic feature extraction capability of CNN enables it to adapt to this complexity. CNN can filter out features related to device classification through its multi-layer structure, ignore irrelevant information such as background, and accurately focus on the features of each device, such as the shape, color, texture of the device and the relative position relationship between them, so as to effectively classify different devices.

[0089] Among them, in the model establishment in the S3 step, Sigmoid function, Tanh function, ReLU function, LeakyReLU function and ELU function can be used for activation; the activation step includes the following operation steps:

[0090] (1) Neuron input calculation: First, for each layer in the neural network (assuming it is a fully connected layer), the neuron input is obtained by multiplying the output of the previous layer by the connection weight and summing them up;

[0091] (2) Application of activation function: After calculating the input of the neuron, use it as the input of the activation function to obtain the output of the neuron (the value after activation);

[0092] (3) Calculate layer by layer: Following the above steps, starting from the input layer, calculate the activated output of each layer of neurons in turn until the final output of the network is obtained. When a simple multi-layer perceptron (MLP) is used for image classification, the input layer receives the feature vector of the image (after preprocessing and feature extraction), and after the activation calculation of the hidden layer, the probability distribution of the classification is finally obtained at the output layer (if the Softmax activation function is used in the output layer, it is used for multi-classification tasks);

[0093] (4) Activation steps in the back propagation phase:

[0094] (4.1) Calculate the gradient of the activation function. During the back-propagation process, the derivative of the activation function is calculated to update the weights.

[0095] (4.2) Calculate the gradient of the loss function with respect to the weights. Use the chain rule to combine the gradient of the activation function and the gradient of the loss function with respect to the output to calculate the gradient of the loss function with respect to the weights.

[0096] (4.3) Weight update: Based on the calculated gradient, an optimization algorithm (such as stochastic gradient descent, SGD) is used to update the weight.

[0097] Furthermore, in specific implementation, the introduction of activation functions can enable neural networks to approximate any complex nonlinear function. Taking the power load prediction in smart grids as an example, there is a complex nonlinear relationship between power load and multiple factors such as time, weather, and economic activities. By using activation functions in neural networks, the model can better fit this nonlinear relationship, thereby more accurately predicting the power load. Image data in smart grid image classification and time series data in power systems (such as voltage and current fluctuations) often have complex distributions. The nonlinear characteristics of activation functions enable neural networks to learn nonlinear patterns in data. For example, when processing image data for smart grid equipment fault detection, the distribution of equipment fault states and normal states in the image feature space may be highly nonlinear. Activation functions help the model build complex decision boundaries to distinguish devices in different states and improve classification accuracy.

[0098] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. An image classification method based on a semantic knowledge base in a smart grid, characterized by comprising the following steps: S1. Data collection: 1.

1. Collect image data from various devices in the smart grid (such as surveillance cameras and drone inspection equipment); these images may include the appearance of power equipment, operating status (such as oil temperature indication of transformers, opening and closing status of switches), and line integrity; 1.

2. Perform preliminary preprocessing on the collected images, including unifying the image format (such as converting them to JPEG format), resizing (to adapt them to subsequent processing algorithms, such as uniformly adjusting them to 512×512 pixels), and removing noise (using filtering technology, such as Gaussian filtering to remove salt and pepper noise in the image); S2. Construction of semantic knowledge base: 2.

1. Definition of terms: collect professional terms in the field of smart grid, such as the names of various power equipment ("transformer", "insulator", "switch"), equipment status ("normal operation", "fault", "overheat"), and related operation and environment concepts ("under maintenance", "lightning strike"); 2.

2. Establishing semantic relationships: define the relationships between terms. For example, "transformer" and "overheating" may have a relationship that "transformer may have an overheating fault"; "insulator" and "breakage" may have a relationship that "insulator breakage is a fault state". These relationships can be determined through domain expert knowledge, historical fault records, and relevant standard documents; 2.

3. Knowledge representation: Use appropriate knowledge representation methods, such as ontology languages ​​(such as OWL-WebOntologyLanguage) to build a semantic knowledge base. Taking transformers as an example, the categories of transformers (oil-immersed transformers, dry-type transformers), their attributes (rated power, voltage level), and their associations with fault states (such as short circuit, overload) can be defined in the ontology. S3. Feature extraction: 3.

1. Use deep learning models (such as convolutional neural networks (CNNs)) to extract image features; use classic VGG-16 and ResNet models; input the preprocessed images into these models, and the models will automatically extract the texture and shape features of the images; 3.

2. For smart grid images, focus on the specific features of power equipment; Taking the substation equipment image as an example, the focus may be on extracting the oil level gauge position of the transformer, the contact shape of the switch, and the number of insulator strings; These feature vectors will serve as an important basis for subsequent classification; S4. Classification Model 4.

1. Semantic-based classification rule construction: Classification rules are constructed by combining the knowledge in the semantic knowledge base and the extracted image features. If the image features show that the temperature of a specific area of ​​a device is too high, and the semantic knowledge base contains the knowledge that "overtemperature of the device may cause failure", then the image can be classified as an image of a device that may have a failure. 4.

2. Classification algorithm selection: You can use traditional classification algorithms such as support vector machine (SVM) and decision tree, or you can use classifiers in deep learning (such as fully connected neural network); input the extracted image features into the classification model, and the model classifies the images according to pre-built classification rules, for example, into "normal equipment", "faulty equipment-overheating", and "faulty equipment-mechanical damage" categories.

2. The image classification method based on semantic knowledge base in smart grid according to claim 1, characterized in that: The updating and maintenance of the semantic knowledge base in step S2 includes the following specific steps: (1) Regularly update the semantic knowledge base to adapt to the development of smart grid technology and the introduction of new equipment. When a new smart meter model is put into use, its related appearance features and normal / fault status knowledge need to be added to the semantic knowledge base. (2) Use data mining technology to extract useful semantic information from new fault cases and maintenance records, and correct and improve the semantic relationships in the knowledge base.

3. The image classification method based on semantic knowledge base in smart grid according to claim 1, characterized in that: The feature extraction optimization in step S3 includes the following specific steps: (1) Optimize feature extraction based on the content of the semantic knowledge base. If the semantic knowledge base emphasizes the key impact of the state of a specific component of a certain power equipment on the overall operation of the equipment, then the image features of the component can be extracted and enhanced during the feature extraction process. (2) Using a multimodal feature extraction method, in addition to the visual features of the image, it can also combine the features of other sensor data (such as temperature sensors and vibration sensors); for the image classification of a motor device, in addition to extracting the image features of the motor's appearance, it can also combine the temperature characteristics of the motor during operation to more accurately determine whether the motor is in a normal state.

4. The image classification method based on semantic knowledge base in smart grid according to claim 1, characterized in that: The classification model training and evaluation in step S4 includes the following specific steps: (1) Use the labeled smart grid image dataset to train the classification model; the labeling information can be determined based on the actual equipment status inspection results, such as "normal transformer" and "faulty switch (mechanical failure)" after inspection by professionals; (2) Use the cross-validation method to evaluate the performance of the classification model; divide the data set into training set, validation set and test set, use the validation set to adjust the model parameters during the training process, and finally use the test set to evaluate the model's accuracy and recall performance indicators; if the model performance does not meet the requirements, you can adjust the model structure (such as increasing the number of neural network layers) or optimize the training algorithm (such as adjusting the learning rate).

5. The image classification method based on semantic knowledge base in smart grid according to claim 1, characterized in that: The application and feedback of the classification results in step S4 include the following specific steps: (1) Apply the classified image results to the operation and management of the smart grid; for images classified as faulty equipment, promptly notify maintenance personnel to conduct repairs; for images of normal equipment, they can be saved as records of normal equipment operation; (2) Further improve the classification method based on feedback from actual applications; When maintenance personnel discover misjudgments in classification results, they analyze whether it is an error in the semantic knowledge base, inaccurate feature extraction, or a problem with the classification model, and then make targeted improvements.

6. The image classification method based on semantic knowledge base in smart grid according to claim 1, characterized in that: The convolution operation in the convolutional neural network model in step S3 can be expressed by the following formula: For an input feature map X (size H in ×W in ×C in , where H in is the height, W in is the width, C in is the number of channels), using a convolution kernel (also called a filter, with a size of h×w×C in , h and w are the height and width of the convolution kernel respectively) to perform convolution operation and output a feature map (size is H out ×W out ×C out . Among them, i and j represent the position coordinates (height and width) of the output feature map, k represents the output channel index, K is the weight parameter of the convolution kernel, and b k is the offset term of the corresponding channel; H out , W out , C out The calculation of is usually as follows: Among them, padding represents the number of pixels filled at the edge of the input feature map, and stride represents the step size of the convolution kernel sliding.

7. The image classification method based on semantic knowledge base in smart grid according to claim 1, characterized in that: In the model building in the S3 step, Sigmoid function, Tanh function, ReLU function, LeakyReLU function and ELU function can be used for activation.

8. The image classification method based on semantic knowledge base in smart grid according to claim 7, characterized in that: The activation step includes the following steps: (1) Neuron input calculation: First, for each layer in the neural network (assuming it is a fully connected layer), the neuron input is obtained by multiplying the output of the previous layer by the connection weight and summing them up; (2) Application of activation function: After calculating the input of the neuron, use it as the input of the activation function to obtain the output of the neuron (the value after activation); (3) Calculate layer by layer: Following the above steps, starting from the input layer, calculate the activated output of each layer of neurons in turn until the final output of the network is obtained. When a simple multi-layer perceptron (MLP) is used for image classification, the input layer receives the feature vector of the image (after preprocessing and feature extraction), and after the activation calculation of the hidden layer, the probability distribution of the classification is finally obtained at the output layer (if the Softmax activation function is used in the output layer, it is used for multi-classification tasks); (4) Activation steps in the back propagation phase: (4.1) Calculate the gradient of the activation function. During the back-propagation process, the derivative of the activation function is calculated to update the weights. (4.2) Calculate the gradient of the loss function with respect to the weights. Use the chain rule to combine the gradient of the activation function and the gradient of the loss function with respect to the output to calculate the gradient of the loss function with respect to the weights. (4.3) Weight update: Based on the calculated gradient, an optimization algorithm (such as stochastic gradient descent, SGD) is used to update the weights.

Images