A method for detecting cracks in a boiler heating surface based on millimeter wave radar imaging

Abstract

The disclosure provides a boiler heating surface crack detection method based on millimeter wave radar imaging. The millimeter wave radar is used for boiler heating surface detection, breaking through the limitations of traditional visual means in smoke shielding and weak light environment. The millimeter wave radar detection equipment is integrated into a small unmanned aerial vehicle / robot platform to form a compact mobile intelligent terminal. The terminal can be used to detect narrow or high-risk areas such as boiler furnace, horizontal flue and tail flue which are difficult for personnel to enter. Compared with the traditional method of disassembling equipment or erecting scaffolding for inspection, the mobile terminal solution greatly reduces the maintenance preparation and implementation difficulty. Through multi-angle multi-frame radar image super-resolution fusion, the crack in the image is significantly enhanced, and the subtle abnormalities can be clearly identified in a complex background. The convolutional neural network is introduced to automatically extract defect features, which is more objective and accurate than artificial rule discrimination.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power plant equipment maintenance, in particular to a boiler heating surface crack detection method based on millimeter wave radar imaging. BACKGROUND

[0002] In the long-term operation of thermal power boilers, the heating surface tube wall (such as water wall tube, superheater tube, etc.) is prone to cracks due to high temperature, high pressure and thermal cycling, causing leakage and becoming one of the main reasons for unplanned unit shutdown. This problem is more prominent under deep peak shaving conditions: frequent and large load changes of the boiler lead to repeated thermal expansion and contraction of the components, intensifying the formation and expansion of cracks.

[0003] To prevent the heating surface tube from leaking, the traditional method is to conduct manual inspection by maintenance personnel during unit shutdown maintenance, including visual inspection, wall thickness measurement by thickness gauge, expansion detection by plug gauge, and auxiliary non-destructive testing such as ultrasonic and eddy current. However, manual detection has obvious limitations: it requires experienced personnel and is labor-intensive, is limited by environmental factors such as narrow space, high temperature residual heat and insufficient lighting, and has a wide detection range that is difficult to cover comprehensively and timely, and defects are often missed due to human error. In addition, camera-based machine vision detection methods are often affected by smoke, dust and lighting in the boiler, making it difficult to detect small cracks or defects covered by dust, and the reliability is insufficient.

[0004] Millimeter wave radar has good material penetration and environmental adaptability: it can work in smoky, high-temperature and dark conditions, is not affected by light, and can still obtain clear measurement data even in harsh environments with a lot of smoke and dust. Studies have shown that millimeter wave high-frequency signals are very sensitive to small discontinuities on metal surfaces, and even tiny cracks covered by a coating or oxide layer can cause changes in scattering characteristics and be detected by radar. By integrating millimeter wave radar sensors into mobile vehicles such as drones, robots or robotic arms, a perception and motion control integrated system can be built to overcome the spatial limitations of manual inspection, achieve automatic inspection of large areas of the heating surface, and significantly improve the coverage and safety of defect detection.

[0005] To address the above challenges, the present application introduces millimeter wave radar combined with intelligent mobile vehicles to achieve remote and automatic detection of boiler heating surface cracks. SUMMARY

[0006] The first aspect of the present disclosure provides a boiler heating surface crack detection method based on millimeter wave radar imaging, comprising the following steps:

[0007] An autonomous mobile vehicle carrying a millimeter wave radar array is controlled to move along a predetermined trajectory, continuously emitting frequency-modulated continuous wave signals to the boiler heating surface and receiving its reflected echo signals during the movement.

[0008] solving the radial distance of the heating surface of the boiler according to the phase difference and the frequency difference of the reflected echo signal;

[0009] fusing the motion trajectory of the autonomous mobile carrier and the radial distance, and performing synthetic aperture radar imaging processing on the fused data to generate a radar image;

[0010] performing multi-modal enhancement processing on the radar image to obtain an enhanced image, inputting the enhanced image into a pre-trained crack identification model, and outputting a crack identification result.

[0011] In combination with the first aspect, the solving of the radial distance of the heating surface of the boiler comprises:

[0012] performing fast Fourier transform on the beat signal of the reflected echo signal;

[0013] calculating the radial distance of the heating surface of the boiler according to the beat frequency corresponding to the spectral peak.

[0014] In combination with the first aspect, before the solving of the radial distance of the heating surface of the boiler according to the phase difference and the frequency difference of the reflected echo signal, the method further comprises:

[0015] estimating the real-time pose of the autonomous mobile carrier by fusing inertial measurement unit data and radar positioning data through a Kalman filter;

[0016] performing motion compensation on the reflected echo signal based on the real-time pose.

[0017] In combination with the first aspect, the synthetic aperture radar imaging processing comprises:

[0018] performing range pulse compression on the reflected echo signal to generate a range compressed signal;

[0019] performing range migration correction on the range compressed signal;

[0020] performing azimuth phase focusing processing on the corrected signal to generate the radar image.

[0021] In combination with the first aspect, the multi-modal enhancement processing comprises:

[0022] fusing an optical image collected by a visible light camera or an infrared camera with the radar image;

[0023] performing geometric distortion correction on the fused image by using a pre-set three-dimensional model of the boiler structure;

[0024] improving the image resolution through multi-frame super-resolution reconstruction.

[0025] In combination with the first aspect, the crack identification model is trained in the following manner:

[0026] constructing a training dataset containing millimeter wave radar images and corresponding crack annotations;

[0027] adopting a transfer learning strategy to utilize pre-trained model weights of visible light pipeline defect images;

[0028] fine-tuning the pre-trained model with the training dataset to optimize crack recognition accuracy.

[0029] In combination with the first aspect, the method further comprises assigning a preset trajectory of the autonomous mobile vehicle according to the crack recognition result, specifically comprising:

[0030] when a crack is recognized, calculating a running angle of the crack in the image;

[0031] adjusting the motion direction of the autonomous mobile vehicle according to the running angle, so that the radar beam azimuth forms a preset included angle with the extension direction of the crack.

[0032] In combination with the first aspect, the method further comprises:

[0033] controlling the autonomous mobile vehicle to hover above the crack area;

[0034] acquiring at least two radar images of different perspectives through multi-angle scanning;

[0035] comparing the crack feature consistency of multi-perspective images to verify the recognition result.

[0036] A second aspect of the present disclosure provides an electronic device, comprising:

[0037] one or more processors;

[0038] a storage unit for storing one or more programs, which when executed by the one or more processors, can cause the one or more processors to implement the millimeter wave radar imaging-based boiler heating surface crack detection method.

[0039] A third aspect of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, which when executed by a processor, can implement the millimeter wave radar imaging-based boiler heating surface crack detection method.

[0040] Beneficial effects: The present disclosure provides a boiler heating surface crack detection method based on millimeter wave radar imaging. By using millimeter wave radar for boiler heating surface detection, the limitations of traditional visual methods in smoky and weak light environments are broken through. The millimeter wave radar detection equipment is integrated into a small unmanned aerial vehicle / robot platform to form a compact mobile intelligent terminal. The terminal can be used to detect narrow or high-risk areas such as boiler furnaces, horizontal flues, and tail flues where personnel cannot enter. Compared with the traditional method of disassembling equipment or erecting a scaffold for inspection, the mobile terminal solution greatly reduces the difficulty of maintenance preparation and implementation. Through multi-angle multi-frame radar image super-resolution fusion, the crack in the image is significantly enhanced, and subtle abnormalities can be clearly identified in a complex background. The convolutional neural network is introduced to automatically extract defect features, which is more objective and accurate than manual rule-based discrimination. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A flowchart of a boiler heating surface crack detection method based on millimeter wave radar imaging according to an embodiment of the present disclosure.

[0042] Figure 2 An electronic device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0043] Hereinafter, exemplary embodiments will be described in detail with reference to the accompanying drawings. In the following description, unless otherwise expressly specified, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments are not representative of all embodiments consistent with the present disclosure.

[0044] The terminology used in the present disclosure is merely for the purpose of describing particular embodiments and is not intended to limit the present disclosure. The singular forms "a," "an," and "the" used in the present disclosure and the appended claims are 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.

[0045] It should be understood that although the terms first, second, third, etc. can be used in this disclosure to describe various information, these terms are not intended to limit the information. These terms are only used to distinguish one type of information from another type of information. For example, without departing from the scope of the present disclosure, first information can also be referred to as second information, and similarly, second information can also be referred to as first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to determining".

[0046] As Figure 1As shown, it is a flowchart of a boiler heating surface crack detection method based on millimeter wave radar imaging, including:

[0047] S101: Control the autonomous mobile vehicle carrying the millimeter wave radar array to move along the preset trajectory, continuously emit frequency-modulated continuous wave signals to the boiler heating surface during the movement, and receive the reflected echo signals;

[0048] Specifically, control the autonomous mobile vehicle (including unmanned aerial vehicle, wall climbing robot or mechanical arm) carrying the millimeter wave radar array to move along the preset trajectory, penetrate the boiler heating surface soot layer by emitting linear frequency-modulated continuous wave signals, and receive the reflected echo signals.

[0049] Suppose the frequency of the radar-emitted linear frequency-modulated signal changes with time from the initial frequency to in the modulation period . Then the instantaneous frequency function can be expressed as .

[0050] Where is the frequency modulation slope (Hz / s), is the sweep bandwidth, is the time length of a single Chirp. The instantaneous phase of the transmitted signal can be obtained by integrating the frequency with respect to time:

[0051] .

[0052] Where is the initial phase. Then the transmitted signal can be expressed as:

[0053] .

[0054] It can also be expressed in complex exponential form as .

[0055] The radar signal propagates to the front surface of the object and reflects. If the target distance is , then the echo signal produces a round-trip time delay ( is the speed of light) relative to the transmitted signal. Ignoring the reflection attenuation, the received signal can be regarded as the delayed transmitted signal:

[0056]

[0057] Where represents the delay of the signal due to the round-trip propagation delay .

[0058] S102: According to the phase difference and frequency difference of the reflected echo signal, solve the radial distance of the boiler heating surface;

[0059] The calculated radial distance of the boiler heating surface includes:

[0060] Performing a fast Fourier transform on the beat signal of the reflected echo signal;

[0061] According to the beat frequency corresponding to the spectral peak, the radial distance of the boiler heating surface is calculated.

[0062] Specifically, the received signal is mixed (multiplied) with the current transmitted signal and low-pass filtered to obtain a beat signal (intermediate frequency signal)

[0063]

[0064] After low-pass filtering (LPF), the high-frequency term is filtered out, and the remaining beat signal mainly contains frequency The beat frequency is directly proportional to the target distance:

[0065]

[0066] Thus, the target distance (radial distance) can be calculated.

[0067] The radar receiver finds the beat frequency peak by performing FFT spectrum analysis on the mixed intermediate frequency signal, and obtains the target distance information.

[0068] For a moving target, a plurality of Chirp signal sequences are introduced, and the Doppler shift can be extracted using the phase difference of adjacent Chirp echoes, so as to estimate the relative speed (Doppler shift , Wavelength).

[0069] S103: Fuse the motion trajectory of the autonomous mobile vehicle and the radial distance, and perform synthetic aperture radar imaging processing on the fused data to generate a radar image;

[0070] The synthetic aperture radar imaging processing includes:

[0071] Performing range compression on the reflected echo signal to generate a range compressed signal;

[0072] Performing range migration correction on the range compressed signal;

[0073] Performing azimuth phase focusing processing on the corrected signal to generate the radar image.

[0074] Specifically, the goal of synthetic aperture radar (SAR) imaging is to obtain multi-view echoes by utilizing vehicle motion, coherently superimposing the energy of the same target point at different observation positions, thereby focusing to obtain a high-resolution image. To achieve this, the raw echo data needs to be decoupled and focused in the range and azimuth directions. The main steps include:

[0075] For each echo, in fast time Matched filtering is then performed (see step 4 of the implementation process) to compress the long pulse broadened echoes and improve range resolution. This is equivalent to performing a range-direction FFT and phase correction on the echoes at each azimuth position, aligning the pulses of different range units.

[0076] In oblique geometry, the range migration of a target (the curve of the target projected onto different range cells with azimuth and time) changes with its azimuth and position, requiring compensation. A range migration compensation phase factor can be designed to correct the range migration of all scattering points to a trajectory consistent with a certain reference range. For example, by multiplying the range-Doppler domain by a pre-calculated phase function, the range migration curves of all targets can be made to have a consistent shape.

[0077] In azimuth (slow time) Multiple observations of the same target are subjected to phase correction and superposition. Due to the viewing angle difference introduced by vehicle motion, echo data exhibits phase differences and time delays in the azimuth direction, requiring correction. Specifically, azimuth matched filtering / phase correction is introduced to eliminate the relative phase difference between data from different sub-apertures, focusing them onto the same target point. This step is typically achieved by transforming the data to the Doppler frequency domain using azimuth FFT, multiplying by a compensating phase, and then performing an inverse transform. In classic SAR algorithms, this step is called azimuth compression, achieving high azimuth resolution.

[0078] The processed data is projected onto image coordinates and summed to form the image intensity. Back-projection is used, targeting each pixel in the image grid (assuming it corresponds to a scattering point). ), calculate its theoretical distance from each radar location. Extract the corresponding delay from the echo data. The signal at that location is then phase-corrected (multiplied by) (equal factors), then for all Sum.

[0079] This process is equivalent to "projecting" the echo signal back to the scattering point and accumulating it. The amplitude of the accumulated pixel is the reflection intensity imaging value of that point.

[0080] S104: Perform multimodal enhancement processing on the radar image to obtain an enhanced image, input the enhanced image into a pre-trained crack recognition model, and output the crack recognition result.

[0081] The radar image may have limited resolution and severe speckle noise. To improve the visibility of anomalies (such as cracks) in the image, the image needs to be enhanced and combined with other sensor information for multi-modal fusion.

[0082] Super-resolution reconstruction aims to break through the original resolution limit of the radar image and obtain higher clarity by fusing multiple images. The implementation method is as follows.

[0083] Autonomous vehicles image the same area at different times, angles, or positions, and fusing these images can improve the details. Due to the difference in perspective, image registration is needed: aligning multiple millimeter wave radar images in space and calculating the translation, rotation, and scale transformation between them. Registration can use feature point matching (such as SURF / SIFT features) or template matching based on correlation. If the radar image resolution is low and the features are few, the known changes in the pose of the autonomous vehicle can also be used to calculate the projective transformation relationship between the images.

[0084] Pixel-level fusion of multiple registered images can use simple averaging to improve signal-to-noise ratio or Laplacian pyramid fusion to preserve high-frequency details. The information of multiple frames can complement the details: for example, the noise in one frame may be smoothed out in another frame, resulting in clearer texture after fusion.

[0085] Using a super-resolution algorithm (such as a super-resolution network based on sparse representation or deep learning), input multiple low-resolution images, and output a high-resolution image. This requires that multiple images have been registered to a common high-resolution grid, and then the details are inferred by the algorithm. Due to the speckle noise of millimeter wave radar images, denoising may be needed before super-resolution.

[0086] Through the above methods, enhanced images with higher pixel resolution and lower noise than single-frame original SAR images can be obtained.

[0087] After completing image reconstruction and enhancement, the final step is to automatically detect and identify cracks in the image by intelligent algorithms.

[0088] The crack recognition model in this scheme uses a convolutional neural network (CNN) for image anomaly detection and classification. To improve the recognition effect of CNN on millimeter wave image cracks, targeted solutions are made in input feature design and training strategy.

[0089] Input feature design: considering the characteristics of millimeter wave radar images, the original radar intensity image and some derived features are considered as input channels of CNN to enrich the information available to the model:

[0090] Raw radar image: This is the base input of CNN, providing the intensity distribution of the scene. The enhanced and fused image contains the main visual clues of cracks (bright-dark lines, edges, etc.).

[0091] Edge / gradient map: Cracks appear as abrupt changes in local grayscale on the image. Extracting the gradient magnitude map or binary edge map after Canny edge detection can highlight the outline of the cracks. Inputting it as the second channel, CNN can learn the linear edge features more easily.

[0092] Multi-modal channel: Aligned visible light image or other modal data can be input into CNN as an additional channel. For example, cracks are often visible in visible light images (if the lighting and dirt allow), which complements the radar image.

[0093] Distance depth map: Radar imaging also provides distance depth information (such as 3D point cloud projection map), which can be input into CNN to help distinguish noise and real targets.

[0094] CNN architecture design: Crack detection can be considered as an anomaly detection or semantic segmentation task. Possible network designs include:

[0095] Global classification network: Input the image into CNN in blocks or sliding windows for binary classification (with or without cracks). This network structure is similar to the regular image classification CNN, such as ResNet, VGG, etc., only the end needs to be binary judged. However, this method cannot accurately locate the crack position and can only determine the existence, usually requiring post-processing to locate the detection area.

[0096] Target detection network: Use methods similar to Faster R-CNN or YOLO to directly regress the bounding box of the crack in the image. However, cracks are not closed targets, and using a rectangular box to represent them is not accurate.

[0097] Semantic segmentation network: Suitable for detecting cracks and other thin regions, it can use U-Net, SegNet, etc. Encoder-decoder structure to classify each pixel (crack pixel or background). This pixel-level labeling can accurately depict the shape of the crack, but the labeling process is time-consuming. For thinner cracks, the problem can also be converted into edge detection, using a network like HED (Holistically-nested Edge Detection) to detect the crack edges in the image.

[0098] Considering that cracks are usually thin and long, a multi-scale feature fusion CNN structure can be used to capture the continuous morphology of cracks under the receptive field of different convolution layers. An effective approach is to increase the hollow convolution or multi-scale pooling in the network to simultaneously obtain local texture and global structure information. Another consideration is to introduce an attention mechanism to focus on high-contrast and thin line areas to improve the sensitivity to cracks.

[0099] Due to the limited amount of actual crack sample data, a series of enhancement and transfer means are used in the training of the CNN model to improve the performance of the model. The crack recognition model is trained in the following ways:

[0100] A training data set containing millimeter wave radar images and corresponding crack labels is constructed;

[0101] A transfer learning strategy is used to use the pre-trained model weights of the visible light pipeline defect image;

[0102] The pre-trained model is fine-tuned with the training data set to optimize the crack recognition accuracy.

[0103] Specifically, various boiler pipe wall surface images are collected, including normal surface and millimeter wave radar images containing cracks, rust, wear and other abnormalities. Since it is not easy to obtain millimeter wave images, we also collect a general visible light pipeline defect image library for auxiliary training (see step 6.3.3 for details). Each radar image is manually labeled with crack location or class label. If segmentation is performed, the crack pixel mask needs to be labeled.

[0104] The data set is cleaned and data augmented. The augmentation methods include but are not limited to: random rotation (any angle, which helps the network learn the robustness of crack inclination in all directions), translation, scaling, mirror flipping, adding Gaussian noise, adjusting contrast, etc. These operations increase the diversity of training samples and alleviate the problem of insufficient training set. For radar images, different noise levels or blur levels can also be simulated to make the model robust to changes in imaging quality.

[0105] Considering the limited amount of radar crack image data, a transfer learning strategy is used. First, use a general image pre-trained model (such as ResNet pre-trained on ImageNet) as a basis, or specifically, use a model pre-trained on visible light pipeline defect data to initialize the CNN. For example, first train the CNN with a large number of pipeline crack and rust photos taken by a general camera, so that it learns basic texture and shape features (including the fact that cracks are usually thin and long, with high-contrast edges). Then, transfer the weights of this model to the millimeter wave radar task and continue to fine-tune it with millimeter wave images. This can expand the distribution of training data and speed up convergence.

[0106] The CNN is supervised trained using the enhanced millimeter wave crack data. Cross-entropy loss (classification) or Dice / BCE combined loss (segmentation) is adopted according to the task selection. The performance of the validation set is monitored during training to avoid overfitting. If the sample is extremely unbalanced (normal area is much more than crack pixel), the weight of the abnormal class can be increased in the loss function or a difficult sample mining strategy (such as focal loss) can be used.

[0107] The accuracy, recall rate and other indicators of the model in detecting cracks are evaluated on an independent test set (or using cross-validation). Special attention is paid to the missed detection rate (cracks misjudged as normal must be extremely low). If the performance is not up to standard, analyze the error cases and consider adding corresponding samples or adjusting the model (for example, if a certain type of rust is misjudged as a crack, add samples of this type or adjust the network's ability to distinguish textures).

[0108] Preferably, before solving the radial distance of the boiler heating surface according to the phase difference and the frequency difference of the reflected echo signal, it further comprises:

[0109] Fusing inertial measurement unit data and radar positioning data through a Kalman filter to estimate a real-time pose of the autonomous mobile vehicle;

[0110] Motion compensation is performed on the reflected echo signal based on the real-time pose.

[0111] Specifically, the autonomous vehicle (such as a drone) needs accurate attitude and motion parameter estimation during the detection process to stabilize the control movement and ensure the geometric accuracy of radar imaging. To this end, the system uses multi-source sensor data fusion (such as IMU inertial measurement unit, barometer, odometer, etc.) and estimates the state of the vehicle through Kalman filtering, including position, velocity, acceleration and attitude angle, etc. Kalman filtering provides an optimal recursive estimation method for the state of the system in noisy measurements.

[0112] First, a state space model of the vehicle needs to be established, including the state equation of the system (kinematic model) and the observation equation (sensor measurement model):

[0113] Let be the state vector at discrete time , containing vehicle position, velocity and attitude, etc. For example, can be defined as , representing position coordinates , velocity and attitude Euler angles (roll , pitch , yaw ). If acceleration bias or sensor bias is considered, it can also be included in the state.

[0114] The evolution of the state over time is described for a linear system with process noise:

[0115]

[0116] where is the state estimate at time is the prior estimate of the current state predicted from the model; is the state transition matrix (determined by the system kinematics, e.g. under constant velocity model will add the position by the velocity at the last time step multiplied by the sampling period); is the control input (e.g. motor thrust command etc.) acting through the control matrix ; is the process noise, assumed to have zero mean and covariance .

[0117] The relationship between the sensor observation and the state is expressed, also in linear form with measurement noise:

[0118]

[0119] where is the sensor observation vector at time (e.g. acceleration and angular velocity measured by IMU, position features measured by vision / radar etc.), is the observation matrix (mapping the state to the observation space, e.g. when directly measuring the position may be a subset of the identity matrix), is the measurement noise, assumed to have zero mean and covariance .

[0120] The Kalman filter contains two stages, prediction and update, at each time step:

[0121] From the posterior state estimate and covariance at the last time step, the current time step is predicted using the state equation:

[0122] ,

[0123] ,

[0124] The current measurement is obtained, which is fused with the prediction to correct it:

[0125] ,

[0126] where is the gain matrix, reflecting the weighting proportion between the prediction uncertainty and the measurement uncertainty.

[0127] ,

[0128] i.e. the prediction plus a correction weighted by the residual (the difference between the measurement and the prediction).

[0129] - ,

[0130] The covariance matrix of the posterior estimate is obtained (uncertainty usually decreases due to the increased information).

[0131] The above recursion proceeds from the initial state, gradually correcting the vehicle's attitude and position estimates. By fusing IMU inertial data (high frequency but with drift) and radar / visual data (low frequency but absolute positioning), the Kalman filter can obtain an accurate and smooth trajectory and attitude. In this scheme, the filtering result is used in two aspects: ① Provide motion parameters (position, attitude, etc.) to the SAR imaging module for motion compensation and coordinate conversion; ② Provide the moving state to the autonomous control module to realize precise hovering, fixed-point scanning, etc.

[0132] Further, the multi-modal enhancement processing further comprises:

[0133] Fusing the optical image collected by the visible light camera or infrared camera with the radar image;

[0134] Geometric distortion correction is performed on the fused image using a pre-set three-dimensional model of the boiler structure;

[0135] The image resolution is improved through multi-frame super-resolution reconstruction.

[0136] Specifically, the autonomous vehicle is equipped with a visible light camera or an infrared thermal imager, which can obtain an optical image of the pipe wall in the boiler. The optical image is aligned (registered) with the millimeter wave radar image, for example, by aligning the two image coordinates by finding common features such as pipe edges or flanges. Then, the optical image can be fused by layer superposition or feature splicing: the texture details of the optical image are superimposed on the radar image to assist in identifying cracks. Another embodiment is to use the data of the two modalities as two input channels of the CNN in the algorithm, and let the model learn the fused features by itself.

[0137] Having a CAD model or 3D point cloud model of the structure under inspection beforehand allows for image correction and reconstruction. For example, knowing the spacing and radius of the boiler water-cooled wall tubes, geometric distortions in radar images can be corrected: mapping radar imaging coordinates to actual physical coordinates flattens the tube wall surface. This mapping can be obtained through ray tracing simulation or calibration. Once the image is aligned with the real structure, the known structure can be used to eliminate certain artifacts and highlight real anomalies. For example, if the model predicts a smooth and continuous path under normal conditions, but the radar image shows abrupt changes, an anomaly can be identified. Prior models can also be used for image reconstruction: by mapping millimeter-wave radar echoes onto the model surface (similar to texture mapping), a radar reflection map of the structural surface can be obtained, making it easier for the human eye to inspect.

[0138] Through multimodal fusion, the resulting image is more comprehensive and clearer than a single millimeter-wave radar image, visually enhancing contrast and detail, making linear anomalies such as fine cracks more prominent. This provides higher-quality input for subsequent automatic recognition algorithms.

[0139] The method further includes specifying a preset trajectory for the autonomous mobile vehicle based on the crack identification result, specifically including:

[0140] When a crack is detected, the direction and angle of the crack in the image are calculated.

[0141] The autonomous mobile vehicle's movement direction is adjusted according to the aforementioned directional angle, so that the radar beam azimuth and the crack extension direction form a preset angle.

[0142] When a suspected crack is detected during cruise, the vehicle needs to adjust its movement to obtain the optimal imaging view and more information, ensuring that the crack is accurately imaged and verified. Specific strategies include:

[0143] Optimization of Angular Imaging for Crack Directionality: Cracks typically extend along a specific direction (e.g., vertical or horizontal). To image the crack more clearly, the azimuth resolution direction of the synthetic aperture radar (SAR) should be aligned with or at an angle to the crack's direction. Empirically, SAR imaging provides the highest resolution in directions parallel to the moving track. Therefore, if a crack is detected to be roughly vertical, the autonomous vehicle's movement path should be controlled to scan the area as vertically as possible (with the radar's movement direction parallel to the crack) to obtain higher resolution imaging along the crack's direction. If the crack is highly inclined, the track angle should be adjusted to achieve the optimal imaging angle of the SAR relative to the crack. This orientation-adaptive scanning can achieve track angle optimization by analyzing the crack's azimuth angle in the image in real time. Adjustments, such as (Move the radar perpendicular to the crack direction, and observe from the side) or (Parallel to the crack direction, depending on specific imaging requirements).

[0144] The method further comprises:

[0145] controlling the autonomous mobile vehicle to hover above the crack region;

[0146] acquiring at least two radar images of different perspectives by multi-angle scanning;

[0147] comparing the consistency of crack features in multi-perspective images to verify the identification result.

[0148] Specifically, when the CNN detects a suspected crack feature, the autonomous vehicle enters a hovering fine scanning mode. It hovers at a fixed point near the crack region and collects radar views at multiple angles by changing its posture or circling in a small range. This is similar to "wrap-around shooting" in photography: keep the sensor distance from the target unchanged, change the perspective angle and take multiple images. Multi-angle data can be used for the aforementioned super-resolution superposition (to improve image details from different perspectives), and also to verify the stability of crack features (real cracks will still appear as continuous lines from different angles, while noise or false positives may not be consistent). For example, the autonomous vehicle can be made to move left and right around the crack center axis for two times of imaging, or move a few cm forward and backward to image at slightly different incident angles. Then the obtained multiple images are compared and fused to enhance the signal characteristics of real cracks.

[0149] The electronic device 200 can be a desktop computer, a notebook, a palm computer, a cloud server, and the like. The electronic device 200 can include, but is not limited to, a processor 201 and a memory 202. Those skilled in the art can understand that Figure 2 The electronic device 200 is only an example and does not constitute a limitation on the electronic device 200, and can include more or fewer components than shown, or combine certain components, or different components, for example, the electronic device can also include an input / output device, a network access device, a bus, and the like.

[0150] The processor 201 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0151] The memory 202 can be an internal storage unit of the electronic device 200, for example, a hard disk or a memory of the electronic device 200. The memory 202 can also be an external storage device of the electronic device 200, for example, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the electronic device 200. Further, the memory 202 can include both an internal storage unit and an external storage device of the electronic device 200. The memory 202 is used to store the computer program 203 and other programs and data required by the electronic device. The memory 202 can also be used to temporarily store data that has been output or will be output.

[0152] The above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than limit them; although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure.

Claims

1. A method for detecting cracks in a heating surface of a boiler based on millimeter wave radar imaging, characterized by, The method comprises the following steps: controlling an autonomous mobile vehicle carrying a millimeter wave radar array to move along a preset trajectory, continuously emitting a frequency-modulated continuous wave signal to the boiler heating surface during movement, and receiving the reflected echo signal thereof; calculating the radial distance of the boiler heating surface according to the phase difference and frequency difference of the reflected echo signal; fusing the movement trajectory of the autonomous mobile vehicle and the radial distance, and performing synthetic aperture radar imaging processing on the fused data to generate a radar image, the synthetic aperture radar imaging processing comprising: performing distance pulse compression on the reflected echo signal to generate a distance compressed signal, performing range migration correction on the distance compressed signal, performing azimuth phase focusing processing on the corrected signal to generate the radar image; performing multi-modal enhancement processing on the radar image to obtain an enhanced image, inputting the enhanced image into a pre-trained crack identification model, and outputting a crack identification result, the multi-modal enhancement processing further comprises: fusing an optical image collected by a visible light camera or an infrared camera with the radar image, performing geometric distortion correction on the fused image using a preset three-dimensional model of the boiler structure, and improving the image resolution through multi-frame super-resolution reconstruction, the crack identification model is trained in the following manner: a training data set containing millimeter wave radar images and corresponding crack labels is constructed, a transfer learning strategy is adopted, the pre-trained model weight of the visible light pipeline defect image is used, the pre-trained model is fine-tuned using the training data set, and the crack identification accuracy is optimized.

2. The method of claim 1, wherein, The calculation of the radial distance of the boiler heating surface comprises: performing fast Fourier transform on the beat signal of the reflected echo signal; calculating the radial distance of the boiler heating surface according to the beat frequency corresponding to the spectral peak.

3. The method of claim 2, wherein, Before calculating the radial distance of the boiler heating surface according to the phase difference and frequency difference of the reflected echo signal, the method further comprises: estimating the real-time pose of the autonomous mobile vehicle by fusing inertial measurement unit data and radar positioning data through a Kalman filter; motion compensating the reflected echo signal based on the real-time pose.

4. The method of claim 1, wherein, The method further comprises specifying the preset trajectory of the autonomous mobile vehicle according to the crack identification result, specifically comprising: when a crack is identified, calculating the strike angle of the crack in the image; adjusting the movement direction of the autonomous mobile vehicle according to the strike angle, so that the radar beam direction forms a preset included angle with the extension direction of the crack.

5. The method of claim 4, wherein, The method further comprises: controlling the autonomous mobile vehicle to hover above the crack area; collecting at least two radar images of different viewing angles through multi-angle scanning; verifying the identification result by comparing the consistency of crack features in multi-view images.

6. An electronic device, comprising: comprise: one or more processors; a storage unit for storing one or more programs, which when executed by the one or more processors, can cause the one or more processors to implement the boiler heating surface crack detection method based on millimeter wave radar imaging according to any one of claims 1 to 5.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, can implement the boiler heating surface crack detection method based on millimeter wave radar imaging according to any one of claims 1 to 5.

Images