A boiler tube wall thinning detection method based on multimodal sensor fusion

Abstract

The present invention provides a method for detecting boiler tube wall thinning based on multimodal sensor fusion. Through the simultaneous acquisition and deep fusion of multi-source data from millimeter-wave radar, infrared thermal imager, lidar and visible light camera, combined with a dynamic Bayesian optimization scanning strategy and a four-channel convolutional neural network classification algorithm, it achieves high-precision identification and full-coverage monitoring of boiler tube wall thinning defects, solves the problem of missed detection of traditional single sensors under complex working conditions, eliminates the need for shutdown or manual intervention, improves detection efficiency, and provides safe, reliable, and intelligent decision-making non-destructive testing technology support for high-risk scenarios such as thermal power generation.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal power plant equipment maintenance, and in particular to a boiler tube wall thinning detection method based on multi-modal sensor fusion. Background Art

[0002] The heating surface pipes of boilers, especially those in thermal power plants, operate under high temperature and pressure for long periods of time, gradually reducing their wall thickness due to corrosion and wear. This "wall thinning" phenomenon is one of the main causes of boiler tube bursts and poses a significant risk to boiler safety. To prevent tube wall failure due to thinning, regular wall thickness testing of heating surface pipes, such as boiler water walls, is essential. However, boiler interiors are confined and complex, with numerous and widely distributed pipes. Thinning areas are often irregular and hidden, making early thinning difficult to detect due to minor thickness changes.

[0003] Currently, inspection mainly relies on manual labor and single sensing methods, but manual inspection has limited coverage and low efficiency. There are also safety risks (operation must be carried out at high altitude or in a confined space). It can only detect severe thinning or deformation visible to the naked eye and cannot make reliable judgments on early-stage minor thinning.

[0004] Ultrasonic wall thickness measurement is a common nondestructive testing method. While ultrasonic measurement offers high accuracy, it requires point-by-point contact scanning, making 100% coverage difficult. For example, the commonly used spot check UT method can only obtain thickness data at a limited number of measurement points, failing to fully cover the entire water-cooled wall tube surface and easily missing localized thinning defects. Furthermore, ultrasonic testing requires the surface to be clean (often requiring sandblasting to remove oxide scale) and a coupling agent. The actual operation is cumbersome and time-consuming, and can only be performed after the furnace has been shut down and cooled, making it unsuitable for operational status monitoring.

[0005] Radiographic testing can be used to identify internal defects and uneven thickness in pipe walls, but the equipment is expensive and requires radiation protection, so it is generally only used in small, critical areas. Magnetic particle and eddy current testing primarily detect surface cracks and material variations, but their direct detection of thickness reduction is limited or requires complex calibration.

[0006] Single visual methods (conventional video) can be difficult to detect thinning anomalies in low light conditions or when dust accumulates on the pipe wall. Existing technologies combine multiple methods to improve detection reliability and comprehensiveness (for example, visually locating suspicious areas followed by ultrasonic retesting). However, these methods still rely primarily on manual operation and post-analysis, lacking real-time and intelligent capabilities.

[0007] In summary, the existing pipe wall thinning detection has problems such as low degree of automation, incomplete coverage, and high rate of missed detection of early defects, which makes it difficult to meet the maintenance needs of large boiler equipment.

[0008] Based on this, this application proposes a boiler tube wall thinning detection method based on multimodal sensor fusion, which replaces manual work with multi-sensor information fusion and autonomous detection to achieve high-precision, early, and full-coverage automatic detection of boiler tube wall thinning. Summary of the Invention

[0009] In a first aspect of the present disclosure, a method for detecting boiler tube wall thinning based on multimodal sensor fusion is provided, comprising the following steps:

[0010] Millimeter-wave radar, infrared thermal imager, lidar and visible light camera are used to synchronously collect millimeter-wave echo signals, infrared thermal images, 3D point clouds and visible light images of the boiler tube wall;

[0011] Performing time-frequency analysis on the millimeter wave echo signal to extract the echo time difference and phase difference between the outer surface and the inner surface of the boiler tube wall, calculating the thickness of the boiler tube wall in combination with the dielectric properties of the boiler tube wall, and marking areas with abnormal thickness;

[0012] Performing noise reduction and threshold segmentation on the infrared thermal image, extracting the temperature abnormality area, and matching it with the thickness abnormality area, marking the successfully matched overlapping area as the thinning area;

[0013] Controlling the millimeter-wave radar to scan the thinned area to obtain a two-dimensional image, and simultaneously performing coordinate registration on the two-dimensional image and the three-dimensional point cloud, and calculating the thickness value of the thinned area along the normal direction of the point cloud;

[0014] The thickness value of the thinned area is input into a convolutional neural network, and the classification result and confidence of the thinned area are output. The three-dimensional thickness thermal map of the boiler tube wall is generated in combination with the visible light image, and the position and thickness of the thinned area are marked.

[0015] In combination with the first aspect, the operating frequency band of the millimeter wave radar is greater than the Ka band, and an oblique incidence method is adopted.

[0016] In combination with the first aspect, the echo time difference and phase difference between the outer surface and the inner surface of the boiler tube wall are extracted, and the thickness of the boiler tube wall is calculated in combination with the dielectric properties of the boiler tube wall by the following formula:

[0017] ,

[0018] in, is the speed of light, is the echo time difference, is the dielectric constant of the boiler tube wall.

[0019] In combination with the first aspect, performing threshold segmentation on the infrared thermal image includes:

[0020] By traversing the grayscale threshold, the inter-class variance between the thinned area and the normal area is maximized, and the threshold at this time is used as the segmentation threshold. The calculation formula is:

[0021] ,

[0022] in, is the between-class variance, is the proportion of pixels divided by the normal area threshold, is the proportion of pixels in the thinning area threshold division, is the average grayscale of the normal area, is the average grayscale of the thinning area, is the global average grayscale, and The maximum threshold, the threshold at this time as the segmentation threshold.

[0023] In combination with the first aspect, extracting the temperature abnormality area from the infrared thermal image further includes inverting the temperature field of the infrared thermal image into a thickness distribution and establishing a temperature-to-thickness mapping.

[0024] In combination with the first aspect, the matching with the thickness abnormality area and marking the successfully matched overlapping area as a thinning area includes: mapping the coordinates of the thickness abnormality area to the coordinate system of the infrared thermal image, calculating the overlapping area ratio of the two, and if the ratio exceeds 70%, it is determined that the match is successful.

[0025] In combination with the first aspect, controlling the millimeter-wave radar to scan the thinned area to obtain a two-dimensional image includes calculating by the following formula:

[0026] ,

[0027] in, is the coordinate of any pixel point on the tube wall of the two-dimensional image The intensity value at is the radar position, is the distance between the pixel and the radar, It is the echo signal received by the radar after pulse compression.

[0028] In combination with the first aspect, the method further includes performing feature enhancement on the blurred out-of-focus area of ​​the two-dimensional image, specifically including introducing a gain factor according to the pixel signal intensity of the two-dimensional image. , and the new two-dimensional image is obtained by the following formula at coordinates The intensity value at :

[0029] .

[0030] In combination with the first aspect, the step of simultaneously registering the two-dimensional image with the three-dimensional point cloud in a coordinate system and calculating the thickness of the thinned area along the normal direction of the point cloud includes:

[0031] Perform coordinate transformation on the point cloud: ,in represents the LiDAR point coordinates, is the coordinate after conversion to radar coordinate, is the rotation matrix, is the translation vector;

[0032] For the outer surface point at the same position on the pipe wall and inner surface points or echoes , calculate the thickness using the following formula: ,in The outer surface of the tube wall The unit normal vector at .

[0033] Beneficial effects: The present disclosure provides a boiler tube wall thinning detection method based on multimodal sensor fusion. Through the synchronous acquisition and deep fusion of multi-source data of millimeter-wave radar, infrared thermal imager, lidar and visible light camera, combined with dynamic Bayesian optimization scanning strategy and four-channel convolutional neural network classification algorithm, it realizes high-precision identification and full-coverage monitoring of boiler tube wall thinning defects, solves the problem of missed detection of traditional single sensors under complex working conditions, eliminates the need for shutdown or manual intervention, improves detection efficiency, and provides safe, reliable and intelligent decision-making non-destructive testing technology support for high-risk scenarios such as thermal power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The present invention is a flowchart of a method for detecting boiler tube wall thinning based on multimodal sensor fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] Exemplary embodiments are described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present disclosure.

[0036] The terms used in the embodiments of the present disclosure are for the purpose of describing specific embodiments only and are not intended to limit the embodiments of the present disclosure. The singular forms "a," "the," and "the" used in the embodiments of the present disclosure and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0037] It should be understood that although the terms first, second, third, etc. may be used to describe various information in the embodiments of the present disclosure, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of the embodiments of the present disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0038] like Figure 1 FIG. 1 is a flow chart of a method for detecting boiler tube wall thinning based on multimodal sensor fusion according to an embodiment of the present disclosure, including:

[0039] S101: Uses millimeter-wave radar, infrared thermal imager, lidar, and visible light camera to simultaneously collect millimeter-wave echo signals, infrared thermal images, 3D point clouds, and visible light images of the boiler tube wall;

[0040] S102: performing time-frequency analysis on the millimeter wave echo signal, extracting the echo time difference and phase difference between the outer surface and the inner surface of the boiler tube wall, calculating the thickness of the boiler tube wall in combination with the dielectric properties of the boiler tube wall, and marking areas with abnormal thickness;

[0041] S103: performing noise reduction and threshold segmentation on the infrared thermal image, extracting the temperature abnormality area, matching it with the thickness abnormality area, and marking the successfully matched overlapping area as a thinning area;

[0042] S104: Controlling the millimeter-wave radar to scan the thinned area to obtain a two-dimensional image, and simultaneously performing coordinate registration on the two-dimensional image and the three-dimensional point cloud, and calculating the thickness of the thinned area along the normal direction of the point cloud;

[0043] S105: Inputting the thickness value of the thinned area into a convolutional neural network, outputting the classification result and confidence of the thinned area, combining the visible light image to generate a three-dimensional thickness thermal map of the boiler tube wall and marking the position and thickness of the thinned area.

[0044] Specifically, millimeter-wave radar uses its high-frequency signal to perform non-contact measurement of metal pipe walls. When millimeter-wave radiation strikes a metal pipe, most of the energy is reflected from the outer surface. However, due to the thinness of the pipe wall or material attenuation, some energy can penetrate to the inner surface and be reflected a second time. This can result in double echoes in the received signal. Pipe wall thickness is estimated by analyzing the time delay and amplitude ratio of these double echoes.

[0045] Since conductive metals such as steel have limited penetration capabilities for millimeter waves, the present disclosure selects a higher frequency band (such as the Ka band or higher) and uses oblique incidence and multiple measurements to increase the probability of capturing inner wall echoes.

[0046] Linear frequency modulated continuous wave (FMCW) millimeter wave radar is used to measure the distance between the pipe wall and the wall.

[0047] The transmitted signal can be represented as a linear frequency modulated sine wave: ,in is the initial frequency, For time, is the FM slope (bandwidth In the frequency modulation period If the radar echo is reflected by the pipe wall, the delay (Round trip distance , the speed of light ), the received signal is .

[0048] The transmitted signal is mixed (multiplied) with the delayed echo and low-pass filtered to obtain a beat signal with a fixed beat frequency. .

[0049] The target distance can be calculated using the beat frequency: ,

[0050] The beat signal is processed by FFT to obtain the distance spectrum peak, which can accurately extract the distance information of the pipe wall.

[0051] Furthermore, for double-wall structures such as pipes, the millimeter-wave radar may receive two echoes from the outer surface and the inner surface.

[0052] Assume the beat frequency corresponding to the two echoes is and , and the distance between the outer wall and the inner wall is obtained , then the tube wall thickness is approximately .

[0053] Consider the propagation speed of electromagnetic waves in the medium inside the wall ( is the relative dielectric constant), dielectric correction is required, and the thickness calculation formula is: ,in is the arrival time difference of the two echoes.

[0054] If the phase method is used to measure thickness, the phase difference between the two echoes can be compared. , using relationships Inversion thickness: ( is the millimeter wave wavelength).

[0055] In addition, since the second echo will produce greater attenuation when penetrating the material, an amplitude correction factor is introduced to compensate for the inner wall echo amplitude attenuation, thereby improving the accuracy of thickness estimation.

[0056] Preferably, in order to reduce the noise interference of radar ranging, the echo signal is smoothed in space and time domains. Spatial filtering uses weighted convolution smoothing, such as using Windowed weighted average filter: .

[0057] in After filtering, the coordinate The pixel value at The original image is at coordinates The pixel value at is the filter kernel weight (Gaussian kernel or mean kernel can be used, ), is the radius of the filter window, It is the horizontal offset, which is used to control the movement of pixels on the X axis (width direction). It is the vertical offset, which is used to control the movement of pixels on the Y axis (height direction).

[0058] Smoothing of millimeter wave imaging or distance matrix is ​​achieved to reduce isolated noise points. In the time domain, recursive filtering of multi-frame measurement results can be performed, for example, using a first-order recursive low-pass filter:

[0059] ,

[0060] in is the filtered thickness output value at the current moment (nth time), For the The estimated thickness of the measurement, is the filtered thickness value at the previous moment (n-1th time).

[0061] is the smoothing coefficient. Through exponential averaging over time, the measurement jitter is reduced and the stability of thickness estimation is improved.

[0062] Furthermore, infrared thermal imaging is used to capture infrared thermal images of the pipe wall surface, and abnormal temperature areas are detected through the images.

[0063] Specifically, the infrared image is subjected to noise reduction and smoothing, such as by using Gaussian filtering.

[0064] Convolve with the original image to remove high-frequency noise while retaining temperature gradient information.

[0065] The Otsu method, with its adaptive threshold, is used to binarize infrared images and segment abnormal areas of relatively high or low temperatures. The Otsu method maximizes the inter-class variance between the foreground (thinned area) and the background (normal area) by traversing grayscale thresholds to determine the optimal threshold, enabling automatic detection of thinned pipe wall areas on thermal maps. Its criterion function is:

[0066] ,

[0067] in, is the between-class variance, is the proportion of pixels divided by the normal area threshold, is the proportion of pixels in the thinning area threshold division, is the average grayscale of the normal area, is the average grayscale of the thinning area, is the global average grayscale, and The largest threshold, that is, the best segmentation threshold Through threshold segmentation, the mask of the temperature abnormality area can be extracted to provide a priori for subsequent thickness calculation and multimodal fusion.

[0068] The change of pipe wall thickness will affect its steady-state and transient temperature distribution. According to Fourier's law of heat conduction and thermal resistance model, the thickness difference can be inferred from the surface temperature. Under steady-state conditions, if the temperature of the medium in the pipe is , the outer surface temperature is , and the heat flux density Approximately uniform, the one-dimensional heat conduction equation is simplified to linear: ,in is the thermal conductivity of the material, is the wall thickness. The relationship between local thickness and temperature difference is obtained as follows:

[0069] ,

[0070] When the internal heat source is stable, the thermal impedance of the thinner area decreases, making the outer surface temperature closer to the inner surface temperature (i.e. Therefore, the areas with higher temperatures on the infrared image often correspond to thinning of the pipe wall.

[0071] Using the above model, the measured temperature field can be inverted into thickness distribution: when the material properties and internal temperature are known, the thickness of each point is calculated. If there is external convection heat transfer in the actual environment, the model can be expanded to a composite thermal resistance: ,in is the ambient temperature, is the convective heat transfer coefficient. By measuring the surface temperature in the steady state and estimating , it can still be deduced For transient analysis, we use the partial differential equation for heat conduction to determine thickness based on the time-dependent temperature evolution pattern (thinner walls heat up / cool down faster). By combining steady-state and transient information, we create a temperature-thickness map and identify areas of temperature anomalies.

[0072] Furthermore, the millimeter wave radar is controlled to scan the thinning area to obtain a two-dimensional image.

[0073] The radar antenna moves along the pipe wall to obtain multi-view echoes and uses the back-projection algorithm to reconstruct a high-resolution two-dimensional image.

[0074] Radar transmits signals, any pixel on the pipe wall The distance from the radar is , echo delay After the range pulse compression process, the signal at each position is projected onto the imaging grid according to the corresponding delay and coherently accumulated to form the image intensity. The two-dimensional image reconstruction formula is: .

[0075] in, is the coordinate of any pixel point on the tube wall of the two-dimensional image The intensity value at is the radar position, is the distance between the pixel and the radar, It is the echo signal received by the radar after pulse compression.

[0076] When the above integral is realized discretely, the echo at each antenna position is calculated according to the distance The two-way distance is sampled and added, which is equivalent to matched filter focusing. The energy of the same target is superimposed in the imaging, improving the signal-to-noise ratio and resolution.

[0077] Thinning of the pipe wall can alter the electromagnetic reflection characteristics at that location, often manifesting as a localized decrease in intensity or a blurred, out-of-focus "depression" in a 2D image. This is because the thinned area may have weaker echo intensity (due to a reduced reflection cross-section or increased transmission due to thickness reduction), or because surface deformation causes the echo phase to not conform to the assumed flat wall model, resulting in poor imaging focus.

[0078] Optionally, to highlight these abnormal areas, the imaging algorithm is enhanced:

[0079] During the imaging process, the reference signal of the matched filter is adjusted to take into account the possible phase shift in the thinning area, so that abnormal echoes can also be focused correctly.

[0080] Apply adaptive gain to the reconstructed SAR image, that is, introduce a gain factor based on the pixel signal intensity , increase the gain for low-intensity areas: (For example, ), to enhance its contrast.

[0081] The expected echo of the normal thickness area is used as a background model and subtracted from the reconstructed image to highlight the abnormal residual area.

[0082] By adjusting the above-mentioned imaging algorithm, the concave / blurred area caused by thinning will be more obvious in the image, facilitating subsequent identification and positioning.

[0083] Furthermore, the two-dimensional image and the three-dimensional point cloud are aligned in a coordinate system, and the thickness value of the thinned area is calculated along the normal direction of the point cloud.

[0084] The LiDAR provides 3D point cloud data of the pipe wall's outer surface, while the millimeter-wave radar's 2D image provides information about the distance or thickness of the target. To integrate the two, they need to be unified into the same coordinate system. The rigid body transformation from the laser coordinate system to the radar coordinate system is obtained through calibration. , perform coordinate transformation on the point cloud:

[0085] ,

[0086] in represents the LiDAR point coordinates, is the coordinate after conversion to radar coordinate, is the rotation matrix, is the translation vector. After this step, the data from different sensors will be geometrically aligned. Further, the iterative closest point (ICP) algorithm is used to refine the registration: minimize the mean square error of the distance between the corresponding point sets , improve fusion accuracy.

[0087] In the registration coordinate system, the laser point cloud and millimeter wave ranging results are fused to calculate the pipe wall thickness. Specifically, for the outer surface point at the same position on the pipe wall (from laser point cloud) and inner surface points or echoes (derived from millimeter-wave radar detection, such as the inner wall position obtained by double echo), the thickness can be expressed as the distance between two points:

[0088] ,

[0089] In order to reduce the angle error, the thickness is usually calculated along the normal direction: The outer surface of the tube wall The thickness is approximately the projection of the difference vector in the normal direction:

[0090] ,

[0091] This avoids measurement errors caused by sensor viewing angle deviations and ensures that thickness measurements are perpendicular to the surface. In practice, if millimeter waves directly acquire thickness values ​​(for example, by inverting the thickness at a specific point through an algorithm), these values ​​can also be mapped to the corresponding point cloud location. The laser point cloud provides precise geometry, while the millimeter wave provides in-wall information. By fusing the two, a 3D model with thickness annotations can be generated: each external surface point of the pipe wall is associated with a thickness value, enabling a visual representation of the thinning distribution.

[0092] Furthermore, the thickness value of the thinned area is input into a convolutional neural network, and the classification result and confidence of the thinned area are output. Combined with the visible light image, a three-dimensional thickness thermal map of the boiler tube wall is generated and the position and thickness of the thinned area are marked.

[0093] Specifically, the raw data collected by the multimodal sensor (including millimeter-wave 2D images, infrared thermal images, 3D point clouds, and visible light images) is first preprocessed and aligned. The 2D image generated by the millimeter-wave radar is grayscale normalized and then spatially aligned with the temperature distribution of the infrared thermal image, the thickness matrix generated by laser point cloud projection, and the edge texture features of the visible light image, forming four-channel input data.

[0094] Exemplarily, the convolutional neural network employs a multi-branch design to extract features based on the characteristics of data from different modalities. The millimeter-wave SAR image branch captures electromagnetic reflection anomalies using a 3×3 convolution kernel, the infrared branch analyzes temperature gradient features, the laser thickness branch learns geometric deformation patterns, and the visible light branch focuses on surface texture changes. The feature maps of each branch undergo channel splicing and cross-modal fusion in the intermediate layer, and the correlation between different sensor data is learned through convolution operations. The network's end outputs the classification probability (normal / thinned) of the thinned area through global average pooling and a fully connected layer, and calculates the confidence level based on the variance of repeated inferences (e.g., a variance below 0.05 is considered a high-confidence result).

[0095] During the training phase, a weighted cross-entropy loss function was used to address the scarcity of thinning samples in real-world scenarios. Positive samples were weighted threefold to balance data distribution. During optimization, the Adam optimizer, combined with a dynamic learning rate adjustment strategy (initial learning rate 0.001, decaying by 50% every 10 epochs), effectively improved model convergence speed and generalization capabilities. After training, the network was able to distinguish true thinning from interference signals (such as surface rust or dirt) against complex backgrounds, achieving classification accuracy exceeding 98%.

[0096] The confidence calculation is implemented using the Monte Carlo Dropout method. During the inference phase, the same input data is forward propagated 10 times, each time randomly dropping 50% of the neurons. The variance of the "thinning" probability is then calculated. If the variance is below a preset threshold (e.g., 0.05) and the average probability exceeds 0.85, the region is identified as a high-confidence thinning region. Otherwise, a review mechanism is triggered. This design significantly reduces the false alarm rate and ensures the credibility of the detection results.

[0097] Ultimately, the system fuses the classification results with visible light images and laser point cloud data to generate a three-dimensional thickness heat map. Specifically, the center coordinates of the thinned area are mapped to the corresponding position in the visible light image. Combined with the three-dimensional geometric information of the laser point cloud, the precise location of the thinned area and the remaining thickness value (such as the percentage of thickness loss represented by a color gradient) are annotated in the 3D model. The heat map supports real-time display and interactive operation, allowing users to intuitively view the distribution, severity, and confidence assessment results of the thinned area. At the same time, it triggers graded alarms (such as yellow warning, orange high risk, and red emergency shutdown), providing direct decision-making basis for safe boiler operation and maintenance.

[0098] Through the above technical solution, the present invention realizes full-process automated detection from multimodal data acquisition, intelligent classification to visual output. While improving detection accuracy, it greatly reduces the need for manual intervention and downtime. It is suitable for boiler tube wall health monitoring in harsh industrial environments such as high temperature and high dust.

[0099] The above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than to limit them. Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present disclosure, and should all be included in the scope of protection of the present disclosure.

Claims

1. A method for detecting boiler tube wall thinning based on multimodal sensor fusion, characterized in that: The following steps are involved: Millimeter-wave radar, infrared thermal imager, lidar and visible light camera are used to synchronously collect millimeter-wave echo signals, infrared thermal images, 3D point clouds and visible light images of the boiler tube wall; Performing time-frequency analysis on the millimeter wave echo signal to extract the echo time difference and phase difference between the outer surface and the inner surface of the boiler tube wall, calculating the thickness of the boiler tube wall in combination with the dielectric properties of the boiler tube wall, and marking areas with abnormal thickness; Performing noise reduction and threshold segmentation on the infrared thermal image, extracting the temperature abnormality area, and matching it with the thickness abnormality area, marking the successfully matched overlapping area as the thinning area; Controlling the millimeter-wave radar to scan the thinned area to obtain a two-dimensional image, and simultaneously performing coordinate registration on the two-dimensional image and the three-dimensional point cloud, and calculating the thickness value of the thinned area along the normal direction of the point cloud; The thickness value of the thinned area is input into a convolutional neural network, and the classification result and confidence of the thinned area are output. The three-dimensional thickness thermal map of the boiler tube wall is generated in combination with the visible light image, and the position and thickness of the thinned area are marked.

2. The method according to claim 1, characterized in that The operating frequency band of the millimeter wave radar is larger than the Ka band, and an oblique incidence method is adopted.

3. The method according to claim 1, characterized in that The echo time difference and phase difference between the outer surface and the inner surface of the boiler tube wall are extracted, and the thickness of the boiler tube wall is calculated by combining the dielectric properties of the boiler tube wall using the following formula: ， in, is the speed of light, is the echo time difference, is the dielectric constant of the boiler tube wall.

4. The method according to claim 1, wherein Performing threshold segmentation on the infrared thermal image includes: By traversing the grayscale threshold, the inter-class variance between the thinned area and the normal area is maximized, and the threshold at this time is used as the segmentation threshold. The calculation formula is: ， in, is the between-class variance, is the proportion of pixels divided by the normal area threshold, is the proportion of pixels in the thinning area threshold division, is the average grayscale of the normal area, is the average grayscale of the thinning area, is the global average grayscale, and The maximum threshold, the threshold at this time as the segmentation threshold.

5. The method according to claim 4, characterized in that The performing of noise reduction and threshold segmentation on the infrared thermal image and extracting the temperature abnormality area further includes inverting the temperature field of the infrared thermal image into a thickness distribution and establishing a mapping from temperature to thickness.

6. The method according to claim 1, characterized in that The matching with the thickness abnormal area and marking the successfully matched overlapping area as a thinning area includes: mapping the coordinates of the thickness abnormal area to the coordinate system of the infrared thermal image, calculating the overlapping area ratio of the two, and determining that the match is successful if the ratio exceeds 70%.

7. The method according to claim 1, characterized in that The controlling the millimeter-wave radar to scan the thinning area to obtain a two-dimensional image includes calculating by the following formula: ， in, is the coordinate of any pixel point on the tube wall of the two-dimensional image The intensity value at is the radar position, is the distance between the pixel and the radar, It is the echo signal received by the radar after pulse compression.

8. The method according to claim 7, characterized in that The method further includes enhancing the features of the blurred and out-of-focus areas of the two-dimensional image, specifically including introducing a gain factor according to the pixel signal intensity of the two-dimensional image. , and the new two-dimensional image is obtained by the following formula at coordinates The intensity value at : 。 9. The method according to claim 1, characterized in that The step of simultaneously registering the two-dimensional image with the three-dimensional point cloud in a coordinate system and calculating the thickness of the thinned area along the normal direction of the point cloud comprises: Perform coordinate transformation on the point cloud: ,in represents the LiDAR point coordinates, is the coordinate after conversion to radar coordinate, is the rotation matrix, is the translation vector; For the outer surface point at the same position on the pipe wall and inner surface points or echoes , calculate the thickness using the following formula: ,in The outer surface of the pipe wall The unit normal vector at .

Images