Remote vision monitoring method and system for pumped storage power station

Abstract

The invention discloses a remote vision monitoring method for a pumped storage power station, and the method comprises the steps: S1, collecting an equipment video image and an equipment infrared full image of the pumped storage power station in real time, and constructing a database for storage; s2, selectively carrying out data preprocessing on the video image of the equipment according to external factors so as to enable the video image to meet a high-resolution requirement, and storing the video image into a database; s3, performing real-time fusion on the equipment video image and the equipment infrared full image in the database by adopting a deep learning model based on a GAN architecture; and S4, transmitting the fused image obtained in the step S3 to a display terminal, wherein the fused image is used for displaying the real-time temperature of the selected site in the equipment video image. Two images of equipment operation are obtained through the camera and the infrared full-image temperature measuring device of the remote vision system, the two images are effectively fused by means of the GAN algorithm, the temperature of key components is displayed while the high definition and resolution of the original image are kept, and the identification accuracy and the remote vision practicability are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing and remote viewing systems, and in particular relates to a remote viewing monitoring method and system for a pumped storage power station. Background Art

[0002] Pumped storage power stations are important peak load regulation, frequency regulation and emergency backup power sources in power systems. However, traditional monitoring systems for pumped storage power stations mainly rely on manual monitoring, which has problems such as poor real-time performance, low data processing efficiency, and untimely early warning. The monitoring method of traditional remote viewing systems is relatively simple, and it is difficult to achieve an intuitive integration of appearance video and temperature testing. With the development of machine learning technology, its application in the remote viewing system of pumped storage power stations can realize real-time monitoring of the operating status and temperature of the power station, thereby improving the accuracy and practicality of the remote viewing system.

[0003] Patent CN116761073A discloses a device-oriented substation remote viewing system control method, which generates a panoramic image based on each camera in the substation remote viewing system, obtains the position of each device in the panoramic image and the control parameters of the corresponding camera, and establishes a monitoring database; evaluates the imaging quality of each camera in the monitoring database; obtains the camera with the best imaging quality and the image detected by the camera. However, the image acquisition process of this invention is complicated, and when the external weather and objective environment change, the system adaptability is poor.

[0004] Patent CN115714466A discloses a dual-path confirmation method for command feedback of a hydropower plant SCADA system based on remote viewing intelligent edge computing, which uses a convolutional neural network to identify the operating status of the equipment, and uploads the operating status of the equipment to the SCADA server through a linkage server to implement a dual-path confirmation method for system commands. However, the method and function of this invention are relatively simple, relying only on video images as the only basis for judgment, and the accuracy and reliability need to be improved.

[0005] Patent CN103674263A discloses a temperature measurement device for a remote viewing system of a substation, including substation equipment to be measured, a remote viewing system centralized control center, a remote viewing system main control center and a remote viewing system camera. The invention measures the temperature of specific substation equipment within the visible range of the camera in real time by installing an equipment infrared full-image temperature measurement device above the camera of the remote viewing system. Although the equipment infrared full image can achieve real-time temperature measurement, the boundaries of objects in the infrared image will become blurred, resulting in a decrease in the resolution accuracy, and the practicality needs to be improved. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a remote viewing monitoring method and system for a pumped storage power station, which obtains two images of equipment operation through a remote viewing system camera and an infrared full-image temperature measuring device, and effectively fuses the two images with the help of a GAN algorithm, retaining the high definition and resolution of the original image while displaying the temperature of key components, greatly improving the recognition accuracy and practicality of remote viewing.

[0007] The main technical solutions adopted in the present invention are:

[0008] A remote monitoring method for a pumped storage power station, comprising:

[0009] S1: Real-time acquisition of equipment video images and equipment infrared full images of pumped storage power stations, and building a database for storage;

[0010] S2: Preprocess the device video image selectively according to external factors to meet the high-resolution requirements and store it in the database;

[0011] S3: A deep learning model based on the GAN architecture is used to fuse the device video image and the device infrared full image in the database in real time, and the temperature value obtained from the device infrared full image is added to the device video image to obtain a fused image;

[0012] S4: The fused image obtained in step S3 is transmitted to the display terminal for displaying the real-time temperature of the selected location in the device video image.

[0013] Preferably, the data preprocessing method in S2 is as follows: an unsupervised image-to-image conversion algorithm is used to perform contour repair on the device video image in the extreme environment, clear pictures of the device in the database and the device video image in the extreme environment are used as input for contour repair, and a high-resolution image in the extreme environment with a clear device contour is output.

[0014] Preferably, in S3, the specific training optimization process of the deep learning model based on the GAN architecture is as follows:

[0015] Define a generator network and a discriminator network, and integrate the attention module and pyramid structure in the generator network;

[0016] The device video image and the device infrared full image at the same time and position are used as the input of the generator network, and the fused image is obtained according to the total loss function of the generator in the generator network;

[0017] The fused image and the corresponding real image are used as the input of the discriminator network, which makes a judgment and outputs the fused image or optimizes and updates the parameters of the generator network according to the judgment result.

[0018] Preferably, the generator total loss function L total The expression is:

[0019] L total =L adv +λ1L perceptual +λ2L pixel +λ3L temp (1);

[0020] Among them, L adv To combat the loss, L perceptual is the perceptual loss, L pixel is the pixel-level loss, L temp is the temperature consistency loss, λ1, λ2, λ3 are weight coefficients;

[0021] Adversarial loss L adv The expression is:

[0022] Ladv=Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1-D(G(z)))];

[0023] Where D is the discriminator network, G is the generator network, x is the real image, z is random noise, D(x) is the real image in the discriminator, G(z) is the image generated by the generator, Ex~pdata(x)[] represents the average logarithmic probability that the real image passes through the discriminator D and the discriminator predicts it to be real; Ez~pz(z)[] represents the average logarithmic probability that the generated image passes through the discriminator D and the discriminator predicts it to be false;

[0024] Perceptual loss L perceptual The expression is:

[0025] Lperceptual=||φ(x)-φ(G(z))||;

[0026] Where φ is the pre-trained perceptual network; φ(x) is the feature of the real image x after being extracted by φ; φ(G(z)) is the feature of the generated image G(z) after being extracted by φ;

[0027] Pixel level loss L pixel The expression is:

[0028] Lpixel = ||G(z)-x||;

[0029] The expression for temperature consistency loss is:

[0030] Ltemp=||T(G(z))-T(x)||;

[0031] Where T is the temperature extractor; T(x) is the temperature information extracted from the generated image G(z); and T(G(z)) is the temperature information extracted from the real image x.

[0032] Preferably, in S3, an Adam optimizer is used to minimize the total loss function L of the generator network. total , use the prepared dataset to train the model, continuously optimize the generator and discriminator through adversarial learning, and at the same time, regularly verify the model performance and adjust the hyperparameters to improve the model performance.

[0033] Preferably, in S3, the discriminator network uses multiple discriminators for discrimination, and each discriminator evaluates the authenticity of the fused image at different scales.

[0034] A remote monitoring system for a pumped storage power station, comprising:

[0035] Image acquisition module, including:

[0036] Visible light camera, used to collect real-time video images of devices within the visible spectrum;

[0037] and infrared full-image temperature testing device, used to collect infrared full images of equipment within the infrared spectrum in real time;

[0038] Image preprocessing module, used to process the video images of equipment in extreme environments;

[0039] A data storage module, used to store the image data collected by the image acquisition module and the image data processed by the image preprocessing module;

[0040] The image fusion and display module uses a deep learning model based on the GAN architecture to fuse the infrared temperature information of the device's full infrared image at the same time and location into the device's video image, and displays the real-time temperature of each point in the video image on the display terminal.

[0041] Preferably, in the image fusion and display module, the GAN architecture includes:

[0042] A generator module for generating a fused image with temperature values ​​from a high-resolution device video image;

[0043] A discriminator module, used to distinguish between real images and fused images;

[0044] An attention module to strengthen the model's focus on key areas in the video image;

[0045] A pyramid structure is used to capture image features at different scales.

[0046] Beneficial effects: The present invention provides a remote monitoring method and system for a pumped storage power station, which has the following advantages:

[0047] (1) The present invention adopts a deep learning model based on the GAN architecture to fuse the visible light image of the power station equipment with its corresponding infrared full image, and displays the temperature of key components on the equipment video image in real time, thereby realizing real-time and accurate monitoring of key equipment, greatly reducing the frequency of manual inspections, and improving operation and maintenance efficiency and safety.

[0048] (2) The present invention pre-processes the video images of equipment in extreme environments, making the monitoring system perform particularly well under severe weather conditions, effectively avoiding shutdown accidents caused by equipment overheating, short circuit and other problems in extreme environments, and significantly improving the overall operational efficiency and economic benefits of the power station.

[0049] (3) The present invention realizes the long-term accumulation of historical data by constructing a database, which provides a valuable basis for the equipment health management and predictive maintenance of the power plant, and helps to further optimize the maintenance strategy of the power plant and extend the life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The following is a schematic diagram of the system construction of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application are clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0052] Example 1

[0053] A remote monitoring method for a pumped storage power station, comprising:

[0054] S1: Real-time acquisition of equipment video images and equipment infrared full images of the pumped storage power station, and building a database for storage; in this embodiment 1, the real-time acquired data includes but is not limited to parameters such as water temperature, turbine external temperature, transformer temperature, and ambient temperature. The database includes high-resolution images of the equipment in various weather conditions such as sunny days and rainy days.

[0055] S2: Preprocess the device video image selectively according to external factors to meet the high-resolution requirements, and store the processed data in the database;

[0056] In this embodiment 1, the specific method of data preprocessing is as follows: an unsupervised image-to-image conversion algorithm is used to perform contour repair on the device video image in extreme environments (such as heavy rain and other weather), and the clear pictures of the device in the database and the device video image in extreme environments (such as heavy rain and other weather) are used as input for contour repair, and a high-resolution image in extreme environments (such as heavy rain and other weather) with clear device contours is output. In the present invention, high-resolution clear images in good weather are used to predict images in extreme weather, so the positions of the two images need to be guaranteed to be the same and are input together as an "image pair" during the training process.

[0057] In the present invention, the use of an unsupervised image-to-image conversion algorithm for image conversion belongs to the prior art, and therefore is not described in detail.

[0058] S3: A deep learning model based on the GAN architecture is used to fuse the device video image and the device infrared full image in the database in real time, and the temperature value obtained from the device infrared full image is added to the device video image to obtain a fused image.

[0059] In this embodiment 1, the deep learning model is a GAN architecture that uses an attention mechanism and / or a pyramid structure to enhance the model's learning ability for local and global features. The specific training optimization process of the deep learning model based on the GAN architecture is as follows:

[0060] Define a generator network and a discriminator network, and integrate the attention module and pyramid structure in the generator network;

[0061] The device video image and the device infrared full image at the same time and position are used as the input of the generator network, and the fused image is obtained according to the total loss function of the generator in the generator network;

[0062] The fused image and the corresponding real image are used as inputs of the discriminator network, which discriminates and outputs a fused image (a high-resolution image with temperature) or optimizes and updates the parameters of the generator network according to the discrimination result. In the present invention, the real image is obtained by artificially fusing the device video image and the device infrared full image.

[0063] In this embodiment 1, the specific construction of the generator network is as follows:

[0064] Infrastructure: It adopts U-Net architecture with an encoder-decoder structure, which can retain more spatial information to achieve high-precision image generation.

[0065] Attention module: An attention mechanism is embedded in the generator network so that the model can focus on processing information in key areas.

[0066] Pyramid structure: Through multi-scale feature extraction layers, the model's ability to handle objects of different sizes is enhanced.

[0067] The discriminator network uses multiple discriminators for multi-scale discrimination, and each discriminator evaluates the authenticity of the generated image at different scales to improve the quality assessment of the generated image.

[0068] In this embodiment 1, the total loss function of the generator L total The expression is:

[0069] L total =L adv +λ1L perceptual +λ2L pixel +λ3L temp ;

[0070] Among them, L adv To combat the loss, L perceptual is the perceptual loss, L pixel is the pixel-level loss, L temp is the temperature consistency loss, λ1, λ2, λ3 are weight coefficients.

[0071] In this embodiment 1, the resistance loss L adv : The standard minimum-maximum game form is adopted, that is, minimizing the loss of the generator and maximizing the loss of the discriminator. The specific expression is:

[0072] L adv =Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1-D(G(z)))];

[0073] Where D is the discriminator network, G is the generator network, x is the real image, z is random noise, D(x) is the real image in the discriminator, G(z) is the image generated by the generator, Ex~pdata(x)[] represents the average logarithmic probability that the discriminator predicts the real image after passing through the discriminator D; Ez~pz(z)[] represents the average logarithmic probability that the generated image is predicted to be false after passing through the discriminator D.

[0074] In this embodiment 1, the perceptual loss L perceptual The feature maps of pre-trained networks such as VGG are introduced to maintain the high quality of generated images. The specific expression is:

[0075] L perceptua l = ||φ(x)-φ(G(z))||;

[0076] Among them, φ is the pre-trained perception network; φ(x) is the feature of the real image x after being extracted by φ; φ(G(z)) is the feature of the generated image G(z) after being extracted by φ.

[0077] In this embodiment 1, the pixel level loss L pixel : Calculate the pixel difference between the fused image and the real image to maintain the local details of the image. The specific expression is:

[0078] L pixel =||G(z)-x||;

[0079] In this embodiment 1, the temperature consistency loss L temp : Ensure that the temperature information on the generated image is consistent with the real temperature image. The specific expression is:

[0080] L temp =||T(G(z))-T(x)||;

[0081] Where T is the temperature extractor; T(x) is the temperature information extracted from the generated image G(z); and T(G(z)) is the temperature information extracted from the real image x.

[0082] In this embodiment 1, the Adam optimizer is used to minimize the total loss function L of the generator network total , use the training data set to train the model (using high-resolution visible light images and corresponding infrared full images (image pairs) at the same time and position as training data), continuously optimize the generator and discriminator through adversarial learning, and at the same time, regularly verify the model performance and adjust the hyperparameters (such as learning rate, weight decay, etc.) to improve the model performance.

[0083] The optimizer used in the present invention can be but not limited to the Adam optimizer. The existing optimizer can minimize the total loss function L of the generator network. total All optimizers are applicable.

[0084] S4: The fused image obtained in step S3 is transmitted to the display terminal for displaying the real-time temperature of the selected location in the device video image.

[0085] Based on the above-mentioned remote monitoring method of pumped storage power station, Figure 1 As shown, this embodiment 1 constructs a pumped storage power station remote monitoring system to monitor the operation status of the pumped storage power station in real time, specifically including:

[0086] Image acquisition module, including:

[0087] The visible light camera is used to collect video images of the device in the visible spectrum in real time; the visible light camera in this embodiment 1 has high resolution, wide angle and night vision functions, and is used to capture video images within the visible spectrum.

[0088] and an infrared full-image temperature testing device (e.g., an infrared thermal imager) for real-time acquisition of an infrared full image of the equipment within the infrared spectrum;

[0089] Image preprocessing module, used to process the video images of equipment in extreme environments (e.g. image sharpening, contour repair, etc.);

[0090] A data storage module, used to store the image data collected by the image acquisition module and the image data processed by the image preprocessing module;

[0091] The image fusion and display module uses a deep learning model based on the GAN architecture to fuse the infrared temperature information of the device's full infrared image at the same time and location into the device's video image, and displays the real-time temperature of each point in the video image on the display terminal.

[0092] In this embodiment 1, the GAN architecture includes:

[0093] A generator module is used to generate a fused image with temperature values ​​from a high-resolution device video image; in this embodiment 1, the generator module adopts a U-Net architecture, which has an encoder-decoder structure and can retain more spatial information.

[0094] A discriminator module to distinguish between real images and fused images; an attention module to strengthen the model's attention to key areas in the image;

[0095] A pyramid structure is used to capture image features at different scales.

[0096] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A remote monitoring method for a pumped storage power station, characterized in that: include: S1: Real-time acquisition of equipment video images and equipment infrared full images of pumped storage power stations, and building a database for storage; S2: Preprocess the device video image selectively according to external factors to meet the high-resolution requirements and store it in the database; S3: A deep learning model based on the GAN architecture is used to fuse the device video image and the device infrared full image in the database in real time, and the temperature value obtained from the device infrared full image is added to the device video image to obtain a fused image; S4: The fused image obtained in step S3 is transmitted to a display terminal for displaying the real-time temperature of the selected location in the device video image.

2. The remote monitoring method for a pumped storage power station according to claim 1, characterized in that: The data preprocessing method in S2 is as follows: an unsupervised image-to-image conversion algorithm is used to perform contour repair on the video image of the device in the extreme environment, clear pictures of the device in the database and the video image of the device in the extreme environment are used as input for contour repair, and a high-resolution image with a clear device contour in the extreme environment is output.

3. The remote monitoring method for a pumped storage power station according to claim 1, characterized in that: In S3, the specific training optimization process of the deep learning model based on the GAN architecture is as follows: Define a generator network and a discriminator network, and integrate the attention module and pyramid structure in the generator network; The device video image and the device infrared full image at the same time and position are used as the input of the generator network, and the fused image is obtained according to the total loss function of the generator in the generator network; The fused image and the corresponding real image are used as the input of the discriminator network, which makes a judgment and outputs the fused image or optimizes and updates the parameters of the generator network according to the judgment result.

4. The remote monitoring method for a pumped storage power station according to claim 3, characterized in that: The total loss function of the generator L total The expression is: L total =L adv +λ1L perceptual +λ2L pixel +λ3L temp ; Among them, L adv To combat the loss, L perceptual is the perceptual loss, L pixel is the pixel-level loss, L temp is the temperature consistency loss, λ1, λ2, λ3 are weight coefficients; Adversarial loss L adv The expression is: Ladv=Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1-D(G(z)))]; Where D is the discriminator network, G is the generator network, x is the real image, z is random noise, D(x) is the real image in the discriminator, G(z) is the image generated by the generator, Ex~pdata(x)[] represents the average logarithmic probability that the real image passes through the discriminator D and the discriminator predicts it to be real; Ez~pz(z)[] represents the average logarithmic probability that the generated image passes through the discriminator D and the discriminator predicts it to be false; Perceptual loss L perceptual The expression is: Lperceptual=||φ(x)-φ(G(z))||; Where φ is the pre-trained perceptual network; φ(x) is the feature of the real image x after being extracted by φ; φ(G(z)) is the feature of the generated image G(z) after being extracted by φ; Pixel level loss L pixel The expression is: Lpixel = ||G(z)-x||; The expression for temperature consistency loss is: Ltemp=||T(G(z))-T(x)||; Where T is the temperature extractor; T(x) is the temperature information extracted from the generated image G(z); and T(G(z)) is the temperature information extracted from the real image x.

5. The remote monitoring method for a pumped storage power station according to claim 3, characterized in that: In S3, the Adam optimizer is used to minimize the total loss function L of the generator network. total , use the prepared dataset to train the model, continuously optimize the generator and discriminator through adversarial learning, and at the same time, regularly verify the model performance and adjust the hyperparameters to improve the model performance.

6. The remote monitoring method for a pumped storage power station according to claim 3, characterized in that: In S3, the discriminator network uses multiple discriminators for discrimination, and each discriminator evaluates the authenticity of the fused image at different scales.

7. A remote monitoring system for a pumped storage power station, which uses the remote monitoring method for a pumped storage power station according to any one of claims 1 to 6 to monitor the operation status of the pumped storage power station, characterized in that: include: Image acquisition module, including: Visible light camera, used to collect real-time video images of devices within the visible spectrum; and infrared full-image temperature testing device, used to collect infrared full images of equipment within the infrared spectrum in real time; Image preprocessing module, used to process the video images of equipment in extreme environments; A data storage module, used to store the image data collected by the image acquisition module and the image data processed by the image preprocessing module; The image fusion and display module uses a deep learning model based on the GAN architecture to fuse the infrared temperature information of the device's full infrared image at the same time and location into the device's video image, and displays the real-time temperature of each point in the video image on the display terminal.

8. A pumped storage power station remote viewing system according to claim 1, characterized in that: In the image fusion and display module, the GAN architecture includes: A generator module for generating a fused image with temperature values ​​from a high-resolution device video image; A discriminator module, used to distinguish between real images and fused images; An attention module to strengthen the model's focus on key areas in the video image; A pyramid structure is used to capture image features at different scales.

Images