A steam leakage detection method for a waste heat recovery system

Abstract

The present invention relates to the field of steam leakage detection, and particularly to a steam leakage detection method for a waste heat recovery system, including: obtaining a visible light image, an infrared image, and a sampling result label at any sampling moment in history; calculating the illumination intensity of the visible light image, clustering a plurality of visible light images according to the illumination intensity to obtain a plurality of clustering clusters, and training a first model and a second model for the same clustering cluster; determining the clustering cluster to which the real-time acquired visible light image belongs, obtaining a first leakage probability and a second leakage probability, and using the result of weighted summation of the first leakage probability and the second leakage probability as the total leakage probability to complete leakage detection. Through the technical solution of the present invention, the accuracy of the steam leakage detection result can be improved, and the energy utilization efficiency of waste heat recovery can be improved.

Description

Technical Field

[0001] The present invention relates to the field of steam leakage detection. More specifically, the present invention relates to a method for detecting steam leakage in a waste heat recovery system. Background Art

[0002] As a key technology for improving energy utilization efficiency, the waste heat recovery system has been fully applied in industrial production. The waste heat recovery system can recover and convert the waste heat or excess heat generated in the industrial production process into useful energy, thereby improving the energy utilization efficiency, reducing energy consumption, and reducing greenhouse gas emissions. Waste heat recovery can not only effectively save costs, but also reduce environmental pollution, with significant economic and environmental benefits.

[0003] In many high-temperature and high-pressure industrial applications, steam, as an important heat transfer medium, is widely used in waste heat recovery systems. The problem of steam leakage is a technical problem that cannot be ignored in such systems. Steam leakage will not only cause energy waste and reduce the waste heat recovery efficiency, but also pose a potential threat to system safety.

[0004] The existing Chinese patent application document with the publication number CN109447011A discloses a method for real-time monitoring of steam pipeline leakage by infrared, including: detecting whether there is a steam leakage area in the monitoring area by using the current frame infrared thermal image and temperature parameters; and judging whether there is a steam leakage area at the position related to the steam leakage area detected in the next frame infrared thermal image.

[0005] However, infrared Figure 1 Generally, the detection effect is good at night. During the day, with strong sunlight and high ambient temperature, the background thermal signal may be confused with the heat source of steam leakage, thus affecting the clarity and detection accuracy of the infrared image. Moreover, the temperature difference during the day is small. When the infrared camera captures the steam leakage area, it may face insufficient contrast with the environmental heat source, resulting in inaccurate steam leakage detection results. Summary of the Invention

[0006] To solve the technical problem of inaccurate steam leakage detection results, the present invention provides a steam leakage detection method for a waste heat recovery system. The method includes: obtaining a visible light image, an infrared image, and a sampling result label at any sampling moment in history; calculating the illumination intensity of the visible light image, clustering a plurality of visible light images according to the illumination intensity to obtain a plurality of clustering clusters, for the same clustering cluster, training a first model according to the visible light image and the sampling result label, training a second model according to the infrared image and the sampling result label, and traversing to obtain the first model and the second model corresponding to each clustering cluster; determining the clustering cluster to which the real-time collected visible light image belongs, inputting the real-time collected visible light image into the first model of the belonging clustering cluster to obtain a first leakage probability, inputting the real-time collected infrared image into the second model of the belonging clustering cluster to obtain a second leakage probability, and taking the result of weighted summation of the first leakage probability and the second leakage probability as the total leakage probability to complete leakage detection.

[0007] Effectively utilizes the complementary characteristics of visible light images and infrared images, trains specialized leakage detection models separately under different illumination conditions, and ensures the stability and accuracy of detection effects in different environments such as day and night. By combining the leakage probabilities of visible light and infrared images and obtaining the total leakage probability through weighted summation, more accurate leakage detection can be achieved.

[0008] Preferably, calculating the illumination intensity includes: converting the visible light image into a grayscale image, constructing a grayscale histogram with the grayscale value range as the abscissa and the number of pixel points as the ordinate; taking the grayscale mean value of the grayscale value range corresponding to the peak in the grayscale histogram as the illumination intensity.

[0009] Can effectively identify the illumination intensity fluctuations in the image, especially quantifying the illumination situation through the mean value of the peak grayscale value, providing accurate illumination information for subsequent clustering analysis and model training. This method of calculating illumination intensity based on the grayscale histogram can classify and optimize images according to different illumination conditions, improving the adaptability of the leakage detection system to different environmental conditions in practical applications.

[0010] Preferably, calculating the illumination intensity includes: converting the visible light image into a grayscale image, and taking the grayscale mean value of all pixel points in the grayscale image as the illumination intensity.

[0011] The grayscale mean value, as a representative of the illumination intensity, can effectively reflect the brightness distribution of the image and provides a unified standard for measuring the illumination conditions of the image.

[0012] Preferably, the first model is a convolutional neural network, which extracts image features from the visible light image and classifies the image features to output the sampling result label.

[0013] Preferably, the training process of the first model includes: using all visible light images in the same clustering cluster in history as input information, and using the true value of the sampling result label as the network label to obtain a set of training data; inputting the training data into the first model to obtain an output result; based on the output result and the network label, using the cross-entropy loss function to calculate the loss value of the first model, backpropagating the error signal according to the loss value, and updating the model parameters of the first model to make the loss value smaller; iteratively updating the model parameters of the first model, and stopping the update when the first model reaches the set maximum number of training times or the loss value is less than the set loss value to obtain the trained first model.

[0014] Preferably, it further includes: for the same clustering cluster, denoising each visible light image within the clustering cluster respectively.

[0015] Preferably, the denoising of each visible light image within the clustering cluster includes: using wavelet transform on any visible light image to obtain a low-frequency image and detail images, and the detail images include a horizontal detail image, a vertical detail image, and a diagonal detail image; respectively obtaining a first image after removing the horizontal detail image, a second image after removing the vertical detail image, and a third image after removing the diagonal detail image; using UCIQE to calculate the quality of the first image, the quality of the second image, and the quality of the third image respectively, calculating the noise degree according to the quality, and calculating the importance of each detail image according to the noise degree and the steam leakage information amount; deleting the detail images with the importance less than the preset threshold, and obtaining the denoised visible light image through inverse wavelet transform of the remaining detail images and the low-frequency image.

[0016] By analyzing the quality and noise degree of the detail images, and combining with the steam leakage information amount to evaluate the importance of each detail image, so as to targetedly remove the unimportant noise details. This method can retain the key information of the image, while removing interference, improving the quality and clarity of the image, and further enhancing the accuracy and robustness of the subsequent leakage detection model.

[0017] Preferably, calculating the noise degree includes: for any detail image, calculating the ratio of the quality of the detail image to the maximum quality value, and using the normalized ratio as the noise degree of the detail image. The quality of the detail image is positively correlated with the noise degree, and the maximum quality value is negatively correlated with the noise degree.

[0018] The positive correlation between the quality and the noise degree enables the detail images with larger noise to be effectively identified and processed, while the negative correlation between the maximum quality and the noise degree ensures that while retaining the important details of the image, the noise can be accurately removed.

[0019] Preferably, the importance satisfies the relational expression:

[0020] , represents the importance of the detail image of, Indicates the detail map of the noise level, Indicates the steam leakage information of the detail map removal where the steam leakage information is Indicates a constant. The greater the noise level, the less important the detail map, because higher noise may mean distortion of image details, resulting in a reduced contribution in restoring the original information. At the same time, the greater the steam leakage information after removing the detail map, it indicates that the detail map carries more information in the image, resulting in a relatively higher importance.

[0021] Preferably, the steam leakage information includes:

[0022] Taking the difference between the first image and the visible light image as the steam leakage information of the horizontal detail map, taking the difference between the second image and the visible light image as the steam leakage information of the vertical detail map, and taking the difference between the third image and the visible light image as the steam leakage information of the diagonal detail map.

[0023] Advantages of the present invention:

[0024] By combining the visible light image and the infrared image and using the methods of clustering analysis and model training, the present invention realizes the accurate detection of steam leakage in the waste heat recovery system. By calculating the illumination intensity of the visible light image and performing clustering, similar images can be grouped into the same category, thereby improving the accuracy of the model. Further, by extracting features from the visible light image and the infrared image through a convolutional neural network and combining the characteristics of the clustering clusters, the probability of steam leakage can be accurately predicted. The denoising process further improves the quality of the image, reduces interference, and improves the robustness of the model. Finally, by comprehensively considering the detection results of the visible light image and the infrared image, more accurate and stable leakage detection can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flowchart of a method for detecting steam leakage in a waste heat recovery system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0027] The following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings.

[0028] Referring to Figure 1 , a method for detecting steam leakage in a waste heat recovery system includes steps S1 - S3, specifically as follows:

[0029] S1: Obtain the visible light image, infrared image, and sampling result label at any sampling moment in history.

[0030] In one embodiment, in the waste heat recovery system, visible light images and infrared image data are collected to monitor and judge the operating state of the system. Each sampling moment is labeled with a sampling result label, and the sampling result label includes steam leakage and normal, so as to perform fault diagnosis and analysis in the later stage.

[0031] Specifically, at each moment during the system operation, a real-time visible light image and infrared image are obtained by installing a camera and an infrared sensor respectively. These images are stored in real time and labeled in combination with the operation data at the sampling moment. The sampling result label is determined according to the image content. For example, if there are obvious signs of steam leakage in the image, the label is marked as steam leakage, otherwise it is normal.

[0032] S2: Calculate the illumination intensity of the visible light image, cluster several visible light images according to the illumination intensity to obtain several clustering clusters. For the same clustering cluster, train the first model according to the visible light image and the sampling result label, and train the second model according to the infrared image and the sampling result label, and traverse to obtain the first model and the second model corresponding to each clustering cluster.

[0033] It should be noted that the lighting conditions are different during the day and at night, which leads to differences in image quality. Specifically, during the day with good lighting conditions, the visible light image can provide a relatively clear picture, making the recognition effect of steam leakage better. However, at night when the illumination intensity is low, the quality of the visible light image decreases, resulting in a poor recognition effect. Relatively speaking, the infrared image performs better at night. Due to the large temperature difference at night and fewer background heat sources, the detection of heat sources such as steam leakage is more accurate; but during the day, due to the interference of staff activities and other environmental factors, the infrared image may be affected, resulting in a poor recognition effect.

[0034] To cope with these different lighting conditions, by analyzing the illumination intensity in the image, it can be automatically determined whether the image is under low illumination intensity conditions, so as to adopt different processing methods during image classification. For example, if the illumination intensity is low, more infrared image-dependent detection methods may be used; while during the day with high illumination intensity, the visible light image is preferentially used for analysis.

[0035] In one embodiment, convert the visible light image into a grayscale image, construct a grayscale histogram with the grayscale value range as the abscissa and the number of pixel points as the ordinate; take the grayscale mean value of the grayscale value range corresponding to the peak in the grayscale histogram as the illumination intensity.

[0036] Cluster a number of visible light images according to the light intensity to obtain a number of clustering clusters. The clustering method can use Kmeans clustering. After clustering, a number of clustering clusters are obtained, and each clustering cluster represents visible light images of a certain light intensity.

[0037] For the same clustering cluster, train the first model according to the visible light image and the sampling result label. The first model is a convolutional neural network. The convolutional neural network extracts image features from the visible light image and classifies the image features to output the sampling result label.

[0038] The training process of the first model includes:

[0039] Take all the visible light images in the same clustering cluster in history as input information, and take the true value of the sampling result label as the network label to obtain a set of training data; input the training data into the first model to obtain an output result; based on the output result and the network label, use the cross-entropy loss function to calculate the loss value of the first model, backpropagate the error signal according to the loss value, and update the model parameters of the first model to make the loss value smaller; iteratively update the model parameters of the first model. When the first model reaches the set maximum number of training times or the loss value is less than the set loss value, stop updating to obtain the trained first model.

[0040] Exemplarily, when the number of training times reaches 200 or the loss value is less than 0.0001, stop updating.

[0041] The second model is also a convolutional neural network, and the training method and network structure of the second model are the same as those of the first model.

[0042] The difference is that when the first model is trained, it is trained by all the visible light images within the same clustering cluster, and after training, it only accepts the input of real-time visible light images; when the second model is trained, it is trained by the infrared images corresponding to all the visible light images within the same clustering cluster, and after training, it only accepts the input of the infrared images collected simultaneously with the real-time visible light images.

[0043] It should be noted that one clustering cluster corresponds to one first model and one second model.

[0044] In one embodiment, calculating the light intensity includes: converting the visible light image into a grayscale image, and taking the grayscale mean value of all pixel points in the grayscale image as the light intensity.

[0045] In one embodiment, it further includes: for the same clustering cluster, denoise each visible light image within the clustering cluster respectively.

[0046] For any visible light image, wavelet transform is used to obtain a low-frequency image and detail images. The detail images include a horizontal detail image, a vertical detail image, and a diagonal detail image. Respectively obtain a first image with the horizontal detail image removed, a second image with the vertical detail image removed, and a third image with the diagonal detail image removed.

[0047] It should be noted that the noise in the visible light image is mainly distributed in the high-frequency detail images, so there is no need to analyze the image after removing the low-frequency image.

[0048] Use UCIQE to calculate the quality of the first image, the quality of the second image, and the quality of the third image respectively. Calculate the noise level according to the quality. The noise level includes:

[0049] For any detail image, calculate the ratio of the quality of the detail image to the maximum quality value, and use the normalized ratio as the noise level of the detail image. The quality of the detail image is positively correlated with the noise level, and the maximum quality value is negatively correlated with the noise level.

[0050] Use the relational expression to represent the noise level as:

[0051] , represents the noise level of the detail image , represents the quality of the image after removing the detail image , represents the maximum quality value, represents the normalization function.

[0052] Take the difference between the first image and the visible light image as the steam leakage information amount of the horizontal detail image, take the difference between the second image and the visible light image as the steam leakage information amount of the vertical detail image, and take the difference between the third image and the visible light image as the steam leakage information amount of the diagonal detail image.

[0053] Calculate the importance of each detail image according to the noise level and the steam leakage information amount. The importance satisfies the relational expression:

[0054] , represents the importance of the detail image , represents the detail image , represents the steam leakage information amount of the image after removing the detail image .

[0055] Delete the detail image with the smallest preset threshold of importance, and obtain the denoised visible light image by inverse wavelet transform of the remaining detail images and the low-frequency image.

[0056] In one embodiment, filter denoising is performed on the visible light image and the infrared image respectively.

[0057] S3. Determine the cluster to which the real-time collected visible light image belongs, input the real-time collected visible light image into the first model of the belonging cluster to obtain the first leakage probability, input the real-time collected infrared image into the second model of the belonging cluster to obtain the second leakage probability, and use the result of weighted summation of the first leakage probability and the second leakage probability as the total leakage probability to complete leakage detection.

[0058] In one embodiment, compare the total leakage probability with the leakage threshold. When the real-time total leakage probability is greater than the leakage threshold, it is considered that steam leakage occurs in the waste heat recovery system at this time, and an alarm signal is generated and sent to remind the staff to perform maintenance in time.

[0059] It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several deformations and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of the patent of this application shall be subject to the appended claims.

Claims

1. A steam leakage detection method for a waste heat recovery system, characterized in that, Including: Obtain the visible light image, infrared image, and sampling result label at any sampling moment in history; Calculate the illumination intensity of the visible light image, and cluster several visible light images according to the illumination intensity to obtain several clustering clusters; For the same clustering cluster, denoise each visible light image within the clustering cluster, including: obtaining a low-frequency image and a detail image for any visible light image using wavelet transform, where the detail image includes a horizontal detail image, a vertical detail image, and a diagonal detail image; respectively obtaining a first image with the horizontal detail image removed, a second image with the vertical detail image removed, and a third image with the diagonal detail image removed; calculating the quality of the first image, the second image, and the third image respectively using UCIQE; calculating the noise level according to the quality, satisfying the relation: , represents the noise level of the detail image , represents the quality of the image with the detail image removed , represents the maximum quality represents the normalization function; and calculating the importance of each detail image according to the noise level and the steam leakage information amount, satisfying the relation: , represents the importance of the detail image , represents the steam leakage information amount of the image with the detail image removed ; delete the detail image with the importance less than the preset threshold, and obtain the denoised visible light image through inverse wavelet transform of the remaining detail images and the low-frequency image; The steam leakage information includes: taking the difference between the first image and the visible light image as the steam leakage information amount of the horizontal detail image, taking the difference between the second image and the visible light image as the steam leakage information amount of the vertical detail image, and taking the difference between the third image and the visible light image as the steam leakage information amount of the diagonal detail image; For the same clustering cluster, train the first model according to the visible light image and the sampling result label, train the second model according to the infrared image and the sampling result label, and traverse to obtain the first model and the second model corresponding to each clustering cluster; Determine the clustering cluster to which the real-time collected visible light image belongs, input the real-time collected visible light image into the first model of the belonging clustering cluster to obtain the first leakage probability, input the real-time collected infrared image into the second model of the belonging clustering cluster to obtain the second leakage probability, and take the result of weighted summation of the first leakage probability and the second leakage probability as the total leakage probability to complete leakage detection.

2. The steam leakage detection method for a waste heat recovery system according to claim 1, wherein Calculating the illumination intensity includes: Convert the visible light image into a grayscale image, and construct a grayscale histogram with the grayscale value range as the abscissa and the number of pixel points as the ordinate; Take the grayscale mean value of the grayscale value range corresponding to the peak in the grayscale histogram as the illumination intensity.

3. The steam leakage detection method for a waste heat recovery system according to claim 1, characterized in that, Calculating the illumination intensity includes: Convert the visible light image into a grayscale image, and take the grayscale mean value of all pixel points in the grayscale image as the illumination intensity.

4. A steam leakage detection method for a waste heat recovery system according to claim 1, characterized in that, The first model is a convolutional neural network. The convolutional neural network extracts image features from the visible light image and classifies the image features to output the sampling result label.

5. A steam leakage detection method for a waste heat recovery system according to claim 1, characterized in that, The training process of the first model includes: Take all visible light images in the same clustering cluster in history as input information, and take the true value of the sampling result label as the network label to obtain a set of training data; Input the training data into the first model to obtain an output result; Based on the output result and the network label, use the cross-entropy loss function to calculate the loss value of the first model, backpropagate the error signal according to the loss value, and update the model parameters of the first model to make the loss value smaller; Iteratively update the model parameters of the first model. When the first model reaches the set maximum number of training times or the loss value is less than the set loss value, stop updating to obtain the trained first model.

6. The steam leakage detection method for a waste heat recovery system according to claim 1, characterized in that, Calculating the noise degree includes: For any detail image, calculate the ratio of the quality of the detail image to the maximum quality value, and take the normalized ratio as the noise degree of the detail image. The quality of the detail image is positively correlated with the noise degree, and the maximum quality value is negatively correlated with the noise degree.

Images