Unattended substation inspection method based on image recognition

Abstract

The invention discloses an unattended transformer substation inspection method based on image recognition, and relates to the field of intelligent inspection. Visible light and infrared thermal imaging cameras are installed according to the layout of a transformer substation to respectively capture equipment visual information and monitor temperature distribution; and setting an acquisition frequency adjustment strategy according to equipment electrical parameters, a temperature change rate and the like to dynamically acquire data. A generative adversarial network (GAN) is utilized to assist in fault diagnosis, a deep learning model is also established, and a model for classification judgment is obtained through training after equipment state image collection, preprocessing and data set division. And meanwhile, a cloud edge cooperative computing architecture is constructed to efficiently process image data, a real-time interaction feedback mechanism is designed to remotely display related conditions of equipment in real time, the running state of the equipment can be monitored more comprehensively and accurately, and the inspection efficiency, the operation and maintenance management efficiency and the like are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent inspection, and in particular to an unmanned substation inspection method based on image recognition. Background Art

[0002] In the operating architecture of today's power system, substations are undoubtedly an extremely critical link, shouldering important missions such as power conversion and distribution. The stable operation of their equipment is directly related to the reliability and safety of power supply.

[0003] For a long time, substation inspection has been mainly done manually. Operation and maintenance personnel need to go to the site regularly to perform a series of operations on the equipment, such as appearance inspection, temperature detection, and operation parameter reading. However, with the continuous expansion of the power system, the number of substations is increasing day by day, and the traditional manual inspection mode has gradually exposed many problems. On the one hand, its workload is huge and inefficient. Operation and maintenance personnel need to run around among many substations, which consumes a lot of time and energy. In addition, during manual inspection, it is difficult to accurately detect some subtle changes or potential faults of the equipment in a short period of time, and the risk of missed inspection is high. On the other hand, manual inspection is obviously restricted by factors such as time and environment. In bad weather, inspection work may not be carried out on time, and poor light will also affect the inspection effect. For substations in remote areas, the cost of manual inspection is even higher. At the same time, the vigorous development of computer technology and artificial intelligence technology has promoted significant progress in image recognition technology and has shown great application potential in many fields. In the field of substation inspection, the application of image recognition technology has become a new trend. By installing cameras and other image acquisition devices in substations, image recognition technology can be used to monitor the appearance and operating status of equipment in real time or regularly, effectively making up for the shortcomings of manual inspections. It can continuously monitor equipment without time restrictions, accurately capture subtle changes in equipment with high accuracy and sensitivity, and operation and maintenance personnel can remotely understand the equipment situation without going to the site, greatly improving the convenience and efficiency of inspections. Nowadays, the construction of unmanned substations has become a trend. It uses automated control systems to achieve remote monitoring, operation and management of equipment, reducing the configuration of on-site operation and maintenance personnel. However, this also puts higher requirements on inspection methods. Some existing inspection methods based on image recognition have problems such as poor data processing capabilities, inaccurate fault diagnosis, and lack of real-time interactive feedback mechanisms. Therefore, it is particularly important and urgent to develop an inspection method that can give full play to the advantages of image recognition technology and integrate perfect data processing, accurate fault diagnosis and real-time feedback functions to ensure efficient, accurate and real-time inspection of unmanned substation equipment, thereby ensuring the safe and stable operation of the power system. Summary of the invention

[0004] The purpose of the present invention is to provide an unattended substation inspection method based on image recognition. Visible light and infrared thermal imaging cameras are reasonably installed according to the substation layout. The former captures visual information such as the appearance of the equipment, and the latter monitors the temperature distribution. At the same time, the acquisition frequency adjustment strategy is set according to the electrical parameters of the equipment, the temperature change rate, etc., and the data is dynamically acquired. The generator is trained to generate pseudo-fault images, the discriminator is trained to distinguish the authenticity of the image, and a deep learning model is established for classification and judgment.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An unmanned substation inspection method based on image recognition is characterized by comprising the following steps:

[0007] S1: Install visible light cameras and infrared thermal imaging cameras; these two different types of cameras are installed to monitor the equipment in the substation from different angles. Visible light cameras can obtain visual information such as the appearance of the equipment under normal light, while infrared thermal imaging cameras focus on capturing the temperature distribution of the equipment in order to detect potential faults caused by overheating and other reasons.

[0008] S2: Set the acquisition frequency adjustment strategy; the operating conditions of substation equipment are dynamically changing, and their electrical parameters, temperature, etc. may be different at different times, so it is necessary to flexibly adjust the image acquisition frequency according to the actual situation.

[0009] S3: Generative Adversarial Network (GAN) assisted fault diagnosis: Generative Adversarial Network (GAN) machine learning technology is used to help determine whether there is a fault in the substation equipment. GAN consists of two parts: the generator and the discriminator. The adversarial training between them is used to improve the fault diagnosis ability.

[0010] S4: Establish a deep learning model; Use deep learning technology to establish a model that can accurately classify and judge the status of substation equipment. Use the mean square error loss function to measure the difference between the output of the model and the actual result.

[0011] S5: Build a cloud-edge collaborative computing architecture to more efficiently process and analyze the large amount of image data obtained during substation inspections

[0012] S6: Design a real-time interactive feedback mechanism. By configuring the cloud platform, establish a command transmission link from the cloud platform to the edge end, and realize real-time display of the image, recognition results and processing process of the equipment in the substation. The purpose of this is to enable operation and maintenance personnel to understand the actual situation of the equipment in the substation in real time remotely, including the appearance image of the equipment, the equipment status results obtained through image recognition analysis, and the specific process of edge end processing of these data, so as to facilitate timely detection of problems and take corresponding measures, and improve the operation and maintenance management efficiency of unmanned substation inspections.

[0013] Visible light camera: Installed in the substation, it is used to capture the appearance of the equipment under normal lighting conditions. These images provide intuitive visual information about the equipment, such as damage on the surface of the equipment, foreign matter attached, etc.

[0014] Infrared thermal imaging camera: Also installed in the substation, but focuses on capturing the temperature distribution of the equipment. Infrared thermal imaging technology can detect abnormal temperature distribution caused by overheating of the equipment, thereby revealing potential fault hazards such as poor contact, overload, etc.

[0015] The operating conditions of substation equipment are changing dynamically, so the image acquisition frequency needs to be flexibly adjusted according to the actual situation. This can be achieved by setting an acquisition frequency adjustment strategy to ensure that key information can be captured at critical moments while avoiding unnecessary waste of resources.

[0016] Composition of GAN: GAN consists of two parts: the generator and the discriminator. The generator is responsible for generating simulated data, while the discriminator is responsible for distinguishing between real data and simulated data.

[0017] Application of GAN: In substation inspection, GAN can be used to assist fault diagnosis. By training the GAN model, it can learn the feature differences between normal equipment and faulty equipment. Then, the trained GAN model is used to diagnose new image data to determine whether the equipment is faulty.

[0018] Advantages of GAN: GAN continuously improves its fault diagnosis capabilities through adversarial training. It has powerful feature extraction and classification capabilities and can accurately identify the fault type of the equipment.

[0019] Model function: Establish a deep learning model that can accurately classify and judge the status of substation equipment.

[0020] Loss function: The mean square error loss function is used to measure the difference between the model output and the actual result. By continuously optimizing the loss function, the model output is closer to the actual result.

[0021] Model training: Use a large amount of image data to train the deep learning model so that it can learn the characteristic differences between the normal and fault states of the equipment.

[0022] Architecture composition: The cloud-edge collaborative computing architecture consists of the cloud, edge, and communication network. The cloud is responsible for global data processing and analysis tasks, while the edge is responsible for local data processing and analysis tasks and real-time response.

[0023] Data processing: The large amount of image data obtained from substation inspections is first processed and analyzed on the edge, and then the key information is uploaded to the cloud for further processing and analysis.

[0024] The step S1 of installing the visible light camera and the infrared thermal imaging camera is to determine the installation position of the visible light high-definition camera according to the layout of the substation and the distribution of the equipment to be monitored, and the requirement is to ensure that each key equipment and equipment group can be within the visual range of at least one camera, and the camera's field of view can cover the main components and connection parts of the equipment; the installation of the infrared thermal imaging camera is also based on the substation equipment layout, and the infrared thermal imaging camera is installed at a position corresponding to the visible light camera or at a suitable position selected according to the key temperature monitoring area.

[0025] The step S2 sets the acquisition frequency adjustment strategy to obtain the electrical parameters of the equipment from the substation monitoring system and set them as vector E, including voltage and current parameters, with a dimension of n E , get the temperature change rate r T , defined as a collection frequency adjustment function f freq (E,r T ), calculate f freq Comprehensive index S:

[0026]

[0027] Where E i is the i-th element of the electrical parameter vector E, w E The weight corresponding to the electrical parameter, w T The weight corresponding to the temperature change rate, w I The weight of the corresponding image recognition result;

[0028] Define the range value of the comprehensive indicator S And the corresponding adjusted acquisition frequency when When when When when When

[0029] The step S3 of generating adversarial network (GAN)-assisted fault diagnosis includes the following steps:

[0030] S31: Data preparation and preprocessing: Before conducting GAN-related training, you need to prepare appropriate data and preprocess it to meet the training requirements.

[0031] S32: Generator training: Extract features from the preprocessed normal state image dataset and fault state image dataset, and then use these features as input to train the generator. The main task of the generator is to generate pseudo fault images that are similar to the normal state but have slight fault features. The purpose of this is to enable the discriminator to better learn to distinguish between real fault images and images that are similar but not really faulty.

[0032] S33: Discriminator training: The images in the real fault image dataset and the pseudo fault images generated by the generator are used as input to train the discriminator. The role of the discriminator is to determine whether the input image is a real fault image or a pseudo fault image generated by the generator. By continuously adjusting the parameters of the discriminator, it can make a more accurate judgment.

[0033] S34: GAN overall training; GAN overall training is achieved by alternately training the generator and the discriminator. In each round of training, the discriminator is first fixed, and the focus is on training the generator to minimize its loss function, that is, to make the pseudo-fault images generated by the generator more realistic; then the generator is fixed, and the discriminator is trained to minimize its loss function, that is, to improve the discriminator's ability to distinguish true and false fault images. These two steps are repeated alternately until the predetermined number of training rounds is reached. Through this overall training method, the generative adversarial network can better assist in the fault diagnosis of substation equipment.

[0034] The step S31 data preparation and preprocessing specifically collects the normal state images of the substation equipment and sets them as the data set D norrmal ; and collect the existing fault state images, set as data set D fault ; Normalize, crop, and resize the images. The goal is to have a unified format and data range for all images. The normal state image dataset after preprocessing is set as D′ normal , the fault state image dataset is D′ fault .

[0035] The generator training in step S32 is specifically to train the preprocessed normal state image data set D′ normal , the fault state image dataset is D′ faultFeature extraction is performed and the features are used as input to train the generator to generate pseudo fault images that are similar to the normal state but have slight fault features. During the training process, the goal is to minimize the difference between the image generated by the generator and the real fault image. The loss function used is the mean square error (MSE) loss function, and the loss function L of the generator is obtained. G :

[0036]

[0037] where |D′ fault | represents the fault state image dataset D′ fault The size of , z is the input of the generator as a random noise vector, and G(z) is the generator function.

[0038] The step S33 discriminator training is specifically to train the real fault image dataset D′ fault The image in and the pseudo fault image generated by the generator are used as input to train the discriminator, and the cross entropy loss function is used to obtain the loss function L of the discriminator. D :

[0039]

[0040] Where D(x) is the discriminator function, and its output is a probability value, which indicates the probability that the input image is a real fault image, |D′ fault | represents the fault state image dataset D′ fault The size of the real fault image is set to y true = 1, the label of the pseudo fault image generated by the generator is y false =0.

[0041] The overall training tool in step S34 is to realize the overall training of GAN by alternately training the generator and the discriminator. In each round of training, the discriminator is first fixed, and the generator is trained so that its loss function L G Minimize; then fix the generator and train the discriminator to make its loss function L D Minimize; repeat this process until the predetermined number of training rounds is reached.

[0042] The step S4 of establishing the deep learning model specifically collects various state images of substation equipment as training data sets, set as D train , perform normalization, cropping, and resizing preprocessing operations on these images to obtain the preprocessed training data set D′ train ; Divide the training data set D′ train The training set, validation set, and test set are divided in a ratio of 8:1:1 and are set as D train_subset， D validation_subset and Dtest_subset ; The training set D train_subset The input is put into the deep learning model for training. The output is the classification result of the model for the image. The mean square error loss function is used to measure the difference between the output of the model and the actual result. The formula is:

[0043]

[0044] Where n is the number of samples, y pred,i and true,i Represent the predicted value and true value of the i-th sample respectively;

[0045] During the training process, the parameters of the model are adjusted by the stochastic gradient descent algorithm to minimize the value of the loss function. The calculation formula for updating the model parameters θ at each iteration is as follows:

[0046]

[0047] Where α is the learning rate, is the gradient of the loss function L(x,y) with respect to the parameter θ;

[0048] Repeat the training process and validate the set D vulidation_subset The accuracy and recall rate of the model are monitored. When the model reaches the predetermined performance indicators on the validation set, the training is stopped to obtain the final deep learning model.

[0049] The step S5 of constructing a cloud-edge collaborative computing architecture specifically includes installing and configuring computing hardware, including a processor, memory, and storage device, at a local server location near the camera installation location of the substation, completing preliminary image preprocessing and feature extraction, and installing a deep learning framework and related image processing libraries to perform subsequent computing operations on the edge device; similarly, installing a deep learning framework and related image processing libraries on the cloud server to complete more complex computing tasks; establishing a stable data transmission link from the edge to the cloud based on a network protocol communication method to complete the cloud-edge collaborative computing architecture.

[0050] The step S6 designs a real-time interactive feedback mechanism, specifically configuring a cloud platform, and establishing a command transmission link from the cloud platform to the edge end to display the image of the equipment in the substation, the recognition results, and the processing process of the edge end in real time.

[0051] Its working mechanism is to achieve efficient inspection through multi-faceted collaborative operation. First of all, in data collection, visible light and infrared thermal imaging cameras are reasonably installed according to the layout of the substation. The former captures visual information such as the appearance of the equipment, and the latter monitors the temperature distribution. At the same time, the acquisition frequency adjustment strategy is set according to the electrical parameters of the equipment, the temperature change rate, etc., and the data is dynamically acquired. Then, in terms of fault diagnosis, the generative adversarial network is used to assist, firstly, data preparation and preprocessing are carried out, normal and fault state images are collected and processed in a standardized manner, and then pseudo-fault images are generated by training generators, and the discriminators are trained to distinguish the authenticity of images. After overall training, fault diagnosis can be better assisted. A deep learning model will also be established to collect and preprocess equipment status images, divide the data set, and then train the model. The final model is determined according to the verification set indicators for classification and judgment. At the same time, a cloud-edge collaborative computing architecture is constructed, hardware is configured locally and in the cloud, frameworks and libraries are installed, and transmission links are established to efficiently process data. Finally, a real-time interactive feedback mechanism is designed to display equipment images, recognition results, and processing processes in real time through the cloud platform and edge links, so as to fully guarantee the inspection effect of unmanned substations.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] By installing visible light cameras and infrared thermal imaging cameras at the same time, comprehensive monitoring of the equipment in the substation from different angles is achieved. The visible light camera can clearly capture the appearance of the equipment under normal lighting, component connections and other visual information, providing a basis for the appearance status evaluation of the equipment; while the infrared thermal imaging camera focuses on monitoring the temperature distribution of the equipment, and can promptly detect potential fault hazards caused by overheating and other reasons. Compared with the existing technology that only uses a single type of camera, this collaborative working mode of multiple types of cameras can obtain richer and more comprehensive equipment operation status information, greatly improving the accuracy and reliability of inspections.

[0054] The acquisition frequency adjustment strategy is set according to factors such as the electrical parameters and temperature change rate of the substation equipment, and the image acquisition frequency can be flexibly adjusted according to the dynamic changes in the actual operating conditions of the equipment. When the equipment status changes greatly, the acquisition frequency can be increased in time to obtain more key data; when the equipment status is relatively stable, the acquisition frequency can be reasonably reduced to avoid the waste of resources caused by excessive acquisition. Compared with the method of fixed acquisition frequency that may be used in the prior art, the dynamic adjustment mechanism of the present invention is more adaptable to the complex and changeable operating environment of substation equipment, ensuring that the efficiency and effectiveness of data acquisition are optimized without missing important information.

[0055] A deep learning model is established and the mean square error loss function is used to measure the difference between the model output and the actual result. The collected equipment status images are preprocessed by normalization, cropping, resizing and other preprocessing operations, and a reasonable data set is divided for training. At the same time, the model parameters are continuously optimized according to the validation set indicators until the final deep learning model is obtained for classification judgment.

[0056] Build a cloud-edge collaborative computing architecture, install and configure computing hardware on the local server near the camera installation location of the substation to complete preliminary processing, such as image preprocessing, feature extraction, etc., and install the deep learning framework and related image processing library for subsequent computing operations; at the same time, the cloud server is also configured accordingly to complete more complex computing tasks and establish a stable data transmission link. It can more efficiently process and analyze the large amount of image data obtained from substation inspections.

[0057] A real-time interactive feedback mechanism is designed to configure the cloud platform and establish a command transmission link from the cloud platform to the edge to achieve real-time display of the images, recognition results and processing process of the equipment in the substation. This enables operation and maintenance personnel to remotely understand the actual situation of the equipment in the substation in real time, including the appearance image of the equipment, the equipment status results obtained through image recognition analysis, and the specific process of edge processing of these data. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of an unattended substation inspection method based on image recognition according to the present invention; DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be fully described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0060] like Figure 1 As shown, an unmanned substation inspection method based on image recognition is characterized in that it includes the following steps:

[0061] S1: Install visible light camera and infrared thermal imaging camera;

[0062] S2: Set the acquisition frequency adjustment strategy;

[0063] S3: Generative Adversarial Network (GAN) assisted fault diagnosis;

[0064] S4: Building a deep learning model;

[0065] S5: Build a cloud-edge collaborative computing architecture;

[0066] S6: Design a real-time interactive feedback mechanism.

[0067] The step S1 of installing the visible light camera and the infrared thermal imaging camera is to determine the installation position of the visible light high-definition camera according to the layout of the substation and the distribution of the equipment to be monitored, and the requirement is to ensure that each key equipment and equipment group can be within the visual range of at least one camera, and the camera's field of view can cover the main components and connection parts of the equipment; the installation of the infrared thermal imaging camera is also based on the substation equipment layout, and the infrared thermal imaging camera is installed at a position corresponding to the visible light camera or at a suitable position selected according to the key temperature monitoring area.

[0068] The step S2 sets the acquisition frequency adjustment strategy to obtain the electrical parameters of the equipment from the substation monitoring system and set them as vector E, including voltage and current parameters, with a dimension of n E , get the temperature change rate r T , defined as a collection frequency adjustment function f freq (E,r T ), calculate f freq Comprehensive index S:

[0069]

[0070] Where E i is the i-th element of the electrical parameter vector E, w E The weight corresponding to the electrical parameter, w T The weight corresponding to the temperature change rate, w I The weight of the corresponding image recognition result;

[0071] Define the range value of the comprehensive indicator S And the corresponding adjusted acquisition frequency when When when When when When

[0072] The step S3 of generating adversarial network (GAN)-assisted fault diagnosis includes the following steps:

[0073] S31: Data preparation and preprocessing;

[0074] S32: generator training;

[0075] S33: discriminator training;

[0076] S34: GAN overall training;

[0077] The step S31 data preparation and preprocessing specifically collects the normal state images of the substation equipment and sets them as the data set D norrmal ; and collect the existing fault state images, set as data set D fault ; Normalize, crop, and resize the images. The goal is to have a unified format and data range for all images. The normal state image dataset after preprocessing is set as D′ normal , the fault state image dataset is D′ fault .

[0078] The generator training in step S32 is specifically to train the preprocessed normal state image data set D′ normal , the fault state image dataset is D′ fault Feature extraction is performed and the features are used as input to train the generator to generate pseudo fault images that are similar to the normal state but have slight fault features. During the training process, the goal is to minimize the difference between the image generated by the generator and the real fault image. The loss function used is the mean square error (MSE) loss function, and the loss function L of the generator is obtained. G :

[0079]

[0080] where |D′ fault | represents the fault state image dataset D′ fault The size of , z is the input of the generator as a random noise vector, and G(z) is the generator function.

[0081] The step S33 discriminator training is specifically to train the real fault image dataset D′ fault The image in and the pseudo fault image generated by the generator are used as input to train the discriminator, and the cross entropy loss function is used to obtain the loss function L of the discriminator. D :

[0082]

[0083] Where D(x) is the discriminator function, and its output is a probability value, which indicates the probability that the input image is a real fault image, |D′ fault | represents the fault state image dataset D′ fault The size of the real fault image is set to y true = 1, the label of the pseudo fault image generated by the generator is y false =0.

[0084] The overall training tool in step S34 is to realize the overall training of GAN by alternately training the generator and the discriminator. In each round of training, the discriminator is first fixed, and the generator is trained so that its loss function L GMinimize; then fix the generator and train the discriminator to make its loss function L D Minimize; repeat this process until the predetermined number of training rounds is reached.

[0085] The step S4 of establishing the deep learning model specifically collects various state images of substation equipment as training data sets, set as D train , perform normalization, cropping, and resizing preprocessing operations on these images to obtain the preprocessed training data set D′ train ; Divide the training data set D′ train The training set, validation set, and test set are divided in a ratio of 8:1:1 and are set as D train_subset， D validation_subset and D test_subset ; The training set D train_subset The input is put into the deep learning model for training. The output is the classification result of the model for the image. The mean square error loss function is used to measure the difference between the output of the model and the actual result. The formula is:

[0086]

[0087] Where n is the number of samples, y pred,i and true,i Represent the predicted value and true value of the i-th sample respectively;

[0088] During the training process, the parameters of the model are adjusted by the stochastic gradient descent algorithm to minimize the value of the loss function. The calculation formula for updating the model parameters θ at each iteration is as follows:

[0089]

[0090] Where α is the learning rate, is the gradient of the loss function L(x,y) with respect to the parameter θ;

[0091] Repeat the training process and validate the set D vulidation_subset The accuracy and recall rate of the model are monitored. When the model reaches the predetermined performance indicators on the validation set, the training is stopped to obtain the final deep learning model.

[0092] The step S5 of constructing a cloud-edge collaborative computing architecture specifically includes installing and configuring computing hardware, including a processor, memory, and storage device, at a local server location near the camera installation location of the substation, completing preliminary image preprocessing and feature extraction, and installing a deep learning framework and related image processing libraries to perform subsequent computing operations on the edge device; similarly, installing a deep learning framework and related image processing libraries on the cloud server to complete more complex computing tasks; establishing a stable data transmission link from the edge to the cloud based on a network protocol communication method to complete the cloud-edge collaborative computing architecture.

[0093] The step S6 designs a real-time interactive feedback mechanism, specifically configuring a cloud platform, and establishing a command transmission link from the cloud platform to the edge end to display the image of the equipment in the substation, the recognition results, and the processing process of the edge end in real time.

[0094] The specific implementation is as follows:

[0095] S1: Install visible light cameras and infrared thermal imaging cameras; specifically, determine the installation location of the visible light high-definition camera based on the layout of the substation and the distribution of equipment that needs to be monitored. The requirement is to ensure that each key equipment and equipment group is within the visual range of at least one camera, and the camera's field of view can cover the main components and connection parts of the equipment; the installation of infrared thermal imaging cameras is also based on the substation equipment layout, and infrared thermal imaging cameras are installed at the position corresponding to the visible light camera or at a suitable location selected according to the key temperature monitoring area.

[0096] S2: Set the acquisition frequency adjustment strategy; specifically, obtain the electrical parameters of the equipment from the substation monitoring system and set them as vector E, including voltage and current parameters, with a dimension of n E , get the temperature change rate r T , defined as a collection frequency adjustment function f freq (E,r T ), calculate f freq Comprehensive index S:

[0097]

[0098] Where E i is the i-th element of the electrical parameter vector E, w E The weight corresponding to the electrical parameter, w T The weight corresponding to the temperature change rate, w I The weight of the corresponding image recognition result;

[0099] Define the range value of the comprehensive indicator S And the corresponding adjusted acquisition frequency when When when When when When

[0100] S3: Generative Adversarial Network (GAN) assisted fault diagnosis; including the following steps:

[0101] S31: Data preparation and preprocessing; specifically, collecting the normal state images of substation equipment, set as data set D norrmal ; and collect the existing fault state images, set as data set D fault ; Normalize, crop, and resize the images. The goal is to have a unified format and data range for all images. The normal state image dataset after preprocessing is set as D′ normal , the fault state image dataset is D′ fault .

[0102] S32: Generator training; specifically, the preprocessed normal state image dataset is D′ normal , the fault state image dataset is D′ fault Feature extraction is performed and the features are used as input to train the generator to generate pseudo fault images that are similar to the normal state but have slight fault features. During the training process, the goal is to minimize the difference between the image generated by the generator and the real fault image. The loss function used is the mean square error (MSE) loss function, and the loss function L of the generator is obtained. G :

[0103]

[0104] where |D′ fault | represents the fault state image dataset D′ fault The size of , z is the input of the generator as a random noise vector, and G(z) is the generator function.

[0105] S33: Discriminator training; specifically, the real fault image dataset D′ fault The image in and the pseudo fault image generated by the generator are used as input to train the discriminator, and the cross entropy loss function is used to obtain the loss function L of the discriminator. D :

[0106]

[0107] Where D(x) is the discriminator function, and its output is a probability value, which indicates the probability that the input image is a real fault image, |D′ fault | represents the fault state image dataset D′ faultThe size of the real fault image is set to y true = 1, the label of the pseudo fault image generated by the generator is y false =0.

[0108] S34: GAN overall training; specifically, the overall training of GAN is achieved by alternately training the generator and the discriminator. In each round of training, the discriminator is first fixed, and the generator is trained so that its loss function L G Minimize; then fix the generator and train the discriminator to make its loss function L D Minimize; repeat this process until the predetermined number of training rounds is reached.

[0109] S4: Establish a deep learning model; specifically, collect various state images of substation equipment as training data sets, set as D train , perform normalization, cropping, and resizing preprocessing operations on these images to obtain the preprocessed training data set D′ train ; Divide the training data set D′ train The training set, validation set, and test set are divided in a ratio of 8:1:1 and are set as D train_subset， D validation_subset and D test_subset ; The training set D train_subset The input is put into the deep learning model for training. The output is the classification result of the model for the image. The mean square error loss function is used to measure the difference between the output of the model and the actual result. The formula is:

[0110]

[0111] Where n is the number of samples, y pred,i and true,i Represent the predicted value and true value of the i-th sample respectively;

[0112] During the training process, the parameters of the model are adjusted by the stochastic gradient descent algorithm to minimize the value of the loss function. The calculation formula for updating the model parameters θ at each iteration is as follows:

[0113]

[0114] Where α is the learning rate, is the gradient of the loss function L(x,y) with respect to the parameter θ;

[0115] Repeat the training process and validate the set D vulidation_subset The accuracy and recall rate of the model are monitored. When the model reaches the predetermined performance indicators on the validation set, the training is stopped to obtain the final deep learning model.

[0116] S5: Build a cloud-edge collaborative computing architecture; specifically, install and configure computing hardware, including processors, memory, and storage devices, at a local server location near the camera installation location in the substation to complete preliminary image preprocessing and feature extraction, and install a deep learning framework and related image processing libraries to perform subsequent computing operations on edge devices; similarly, install a deep learning framework and related image processing libraries on the cloud server to complete more complex computing tasks; establish a stable data transmission link from the edge to the cloud based on the network protocol communication method to complete the cloud-edge collaborative computing architecture.

[0117] S6: Design a real-time interactive feedback mechanism. Specifically, configure the cloud platform to establish a command transmission link from the cloud platform to the edge to display the images of the equipment in the substation, the recognition results, and the processing process of the edge in real time.

[0118] According to the layout of the substation and the distribution of equipment, visible light high-definition cameras and infrared thermal imaging cameras are installed scientifically and reasonably. The visible light camera ensures that each key equipment and equipment group is within its visual range, and the viewing angle covers the main components and connection parts of the equipment to capture the appearance changes and abnormal conditions of the equipment. Infrared thermal imaging cameras are installed according to the key areas of temperature monitoring to capture the temperature distribution of the equipment and promptly detect potential faults such as overheating.

[0119] Using optical imaging and infrared thermal imaging technology, the camera captures the image and temperature information of the equipment, providing data support for subsequent image processing and fault diagnosis.

[0120] Obtain the electrical parameters of the equipment from the substation monitoring system, such as voltage, current, etc., and combine them with the temperature change rate to define a comprehensive index. According to the range value of the comprehensive index, dynamically adjust the image acquisition frequency. When the comprehensive index is high, increase the acquisition frequency to monitor the equipment status more frequently; when the comprehensive index is low, reduce the acquisition frequency to reduce resource consumption.

[0121] By comprehensively considering the electrical parameters, temperature change rate and image recognition results of the equipment, the adjustment function is used to calculate the comprehensive index, and the collection frequency is dynamically adjusted according to the index value to achieve more efficient monitoring and fault warning.

[0122] First, the image data of normal state and fault state are collected and preprocessed. Then the generator is trained to generate pseudo fault images that are similar to the normal state but have slight fault features. Then the discriminator is trained to distinguish between real fault images and pseudo fault images. By alternately training the generator and the discriminator, the fault diagnosis ability of GAN is continuously improved.

[0123] Configure the cloud platform and establish a command transmission link from the cloud platform to the edge. Real-time interaction and feedback between operation and maintenance personnel and the system can be achieved by displaying the images, recognition results and processing process of the equipment in the substation in real time.

[0124] By using the real-time interactive feedback mechanism, the status information and processing results of the equipment in the substation are fed back to the operation and maintenance personnel in real time, so that they can understand the equipment status in time and take corresponding measures. At the same time, the operation and maintenance personnel can also send instructions to the edge through the cloud platform to adjust the collection frequency, optimize the model parameters, etc., to achieve dynamic optimization and adjustment of the system.

Claims

1. An unattended substation inspection method based on image recognition, characterized in that: The following steps are included: S1: Install visible light camera and infrared thermal imaging camera; S2: Set the acquisition frequency adjustment strategy; S3: Generative Adversarial Network (GAN) assisted fault diagnosis; S4: Building a deep learning model; S5: Build a cloud-edge collaborative computing architecture; S6: Design a real-time interactive feedback mechanism.

2. The unattended substation inspection method based on image recognition according to claim 1 is characterized in that: The step S1 of installing the visible light camera and the infrared thermal imaging camera is to determine the installation position of the visible light high-definition camera according to the layout of the substation and the distribution of the equipment to be monitored, and the requirement is to ensure that each key equipment and equipment group can be within the visual range of at least one camera, and the camera's field of view can cover the main components and connection parts of the equipment; the installation of the infrared thermal imaging camera is also based on the substation equipment layout, and the infrared thermal imaging camera is installed at a position corresponding to the visible light camera or at a suitable position selected according to the key temperature monitoring area.

3. The unattended substation inspection method based on image recognition according to claim 1 is characterized in that: The step S2 sets the acquisition frequency adjustment strategy to obtain the electrical parameters of the equipment from the substation monitoring system and set them as vector E, including voltage and current parameters, with a dimension of n E , get the temperature change rate r T , defined as a collection frequency adjustment function f freq (E,r T ), calculate f freq Comprehensive index S: Where E i is the i-th element of the electrical parameter vector E, w E The weight corresponding to the electrical parameter, w T The weight corresponding to the temperature change rate, w I The weight of the corresponding image recognition result; Define the range value of the comprehensive indicator S And the corresponding adjusted acquisition frequency when When when When when When 4. The unattended substation inspection method based on image recognition according to claim 1 is characterized in that: The step S3 of generating adversarial network (GAN)-assisted fault diagnosis includes the following steps: S31: Data preparation and preprocessing; S32: Generator training; S33: discriminator training; S34: GAN overall training; The step S31 data preparation and preprocessing specifically collects the normal state images of the substation equipment and sets them as the data set D norrmal ; and collect the existing fault state images, set as data set D fault ; Normalize, crop, and resize the images. The goal is to have a unified format and data range for all images. The normal state image dataset after preprocessing is set as D′ normal , the fault state image dataset is D′ fault .

5. The unmanned substation inspection method based on image recognition according to claim 4 is characterized in that: The generator training in step S32 is specifically to train the preprocessed normal state image data set D′ normal , the fault state image dataset is D′ fault Feature extraction is performed and the features are used as input to train the generator to generate pseudo fault images that are similar to the normal state but have slight fault features. During the training process, the goal is to minimize the difference between the image generated by the generator and the real fault image. The loss function used is the mean square error (MSE) loss function, and the loss function L of the generator is obtained. G : where |D′ fault | represents the fault state image dataset D′ fault The size of , z is the input of the generator as a random noise vector, and G(z) is the generator function.

6. The unattended substation inspection method based on image recognition according to claim 4 is characterized in that: The step S33 discriminator training is specifically to take the real fault image data set D′ fault The image in and the pseudo fault image generated by the generator are used as input to train the discriminator, and the cross entropy loss function is used to obtain the loss function L of the discriminator. D : Where D(x) is the discriminator function, and its output is a probability value, which indicates the probability that the input image is a real fault image, |D′ fault | represents the fault state image dataset D′ fault The size of the real fault image is set to y true = 1, the label of the pseudo fault image generated by the generator is y false =0.

7. The unattended substation inspection method based on image recognition according to claim 4 is characterized in that: The overall training tool in step S34 is to realize the overall training of GAN by alternately training the generator and the discriminator. In each round of training, the discriminator is first fixed, and the generator is trained so that its loss function L G Minimize; then fix the generator and train the discriminator to make its loss function L D Minimize; repeat this process until the predetermined number of training rounds is reached.

8. The unattended substation inspection method based on image recognition according to claim 1 is characterized in that: The step S4 of establishing the deep learning model specifically collects various state images of substation equipment as training data sets, set as D train , perform normalization, cropping, and resizing preprocessing operations on these images to obtain the preprocessed training data set D′ train ; Divide the training data set D′ train The training set, validation set, and test set are divided in a ratio of 8:1:1 and are set as D train_subset， D validation_subset and D test_subset ; The training set D train_subset The input is put into the deep learning model for training. The output is the classification result of the model for the image. The mean square error loss function is used to measure the difference between the output of the model and the actual result. The formula is: Where n is the number of samples, y pred,i and true,i Represent the predicted value and true value of the i-th sample respectively; During the training process, the parameters of the model are adjusted by the stochastic gradient descent algorithm to minimize the value of the loss function. The calculation formula for updating the model parameters θ at each iteration is as follows: Where α is the learning rate, is the gradient of the loss function L(x,y) with respect to the parameter θ; Repeat the training process and validate the set D vulidation_subset The accuracy and recall rate of the model are monitored. When the model reaches the predetermined performance indicators on the validation set, the training is stopped to obtain the final deep learning model.

9. The unattended substation inspection method based on image recognition according to claim 1 is characterized in that: The step S5 of constructing a cloud-edge collaborative computing architecture specifically includes installing and configuring computing hardware, including a processor, memory, and storage device, at a local server location near the camera installation location of the substation, completing preliminary image preprocessing and feature extraction, and installing a deep learning framework and related image processing libraries to perform subsequent computing operations on the edge device; similarly, installing a deep learning framework and related image processing libraries on the cloud server to complete more complex computing tasks; establishing a stable data transmission link from the edge to the cloud based on a network protocol communication method to complete the cloud-edge collaborative computing architecture.

10. The unattended substation inspection method based on image recognition according to claim 1, characterized in that: The step S6 designs a real-time interactive feedback mechanism, specifically configuring a cloud platform, and establishing a command transmission link from the cloud platform to the edge end to display the image of the equipment in the substation, the recognition results, and the processing process of the edge end in real time.

Images