Early identification method for corrosion area of desulfurization flue based on infrared thermal imaging

Abstract

The application discloses a method for early identification of corrosion area of desulfurization flue in infrared thermal imaging, which comprises the following steps: S1, multi-modal thermal excitation and original data acquisition; S2, pretreatment and multi-dimensional feature extraction; S3, quantitative identification of corrosion area based on a physical model: the extracted electromagnetic thermal response features and the variable frequency phase-locked features are input into a pre-constructed heat-corrosion physical model, a corrosion probability map is generated through feature-level fusion to identify the corrosion area, and the depth estimation value of the corrosion area is obtained based on the quantitative relationship between the variable frequency phase-locked features and the corrosion depth; and S4, three-dimensional space mapping and defect reconstruction. The application introduces a quantitative identification mechanism based on a heat-corrosion physical model, establishes a mapping relationship between feature parameters and corrosion geometric parameters through a pre-constructed finite element simulation database, and adopts a hierarchical fusion strategy to cooperatively analyze the electromagnetic thermal response features and the phase-locked features.

Description

Technical Field

[0001] This application belongs to the field of non-destructive testing technology for equipment, specifically relating to an early identification method for corrosion zones in desulfurization flues based on infrared thermal imaging. Background Technology

[0002] As a key component of environmental protection facilities in industries such as coal-fired power plants and steel smelting, desulfurization flue gas ducts operate under complex and harsh conditions of high humidity, high sulfide levels, and fluctuating temperatures, making internal wall corrosion a prominent issue. Corrosion not only leads to thinning of the flue gas wall and decreased structural strength, but in severe cases, it can even cause safety accidents such as perforation and leakage, resulting in unplanned shutdowns and significant economic losses and safety hazards. Therefore, early, accurate, and quantitative identification and assessment of corrosion areas in desulfurization flue gas ducts are crucial for ensuring safe equipment operation and implementing predictive maintenance.

[0003] Currently, the conventional technical methods for detecting corrosion in desulfurization flues mainly include the following:

[0004] One method is point-based thickness measurement using ultrasonic waves. This method measures the wall thickness point by point using a handheld ultrasonic probe, which can obtain relatively accurate local thickness data. However, its detection efficiency is extremely low, making it difficult to achieve rapid surveys of large areas and the entire range. Furthermore, it carries the risk of missing early pitting corrosion and uniform thinning, and cannot provide continuous distribution information of the corroded area.

[0005] Second, surface defect detection based on eddy current testing. Eddy current testing is sensitive to surface and near-surface cracks, but its detection depth is limited, it is greatly affected by the lift-off effect, and the detection results are easily interfered with by factors such as dirt and coatings on the flue wall surface. In scenarios such as desulfurization flues with complex surface conditions and large curvature changes, the reliability and coverage of eddy current testing are difficult to guarantee.

[0006] Thirdly, there is manual visual inspection based on visible light or infrared thermal imaging. Traditional infrared thermal imaging inspections mostly rely on manual handheld thermal imagers to take pictures and qualitatively identify abnormal areas by observing temperature distribution images. This method is greatly affected by the operator's experience and can only obtain temperature information of the surface layer of the wall, failing to distinguish between surface dirt, uneven heating, and early internal corrosion, making it difficult to achieve quantitative identification and depth assessment of corrosion. In addition, manual inspection methods make it difficult to accurately correlate the detection results with the three-dimensional spatial location of the flue, which is not conducive to long-term tracking of corrosion status and maintenance decisions.

[0007] In summary, existing detection technologies suffer from the following main technical problems: First, it is difficult to balance detection efficiency and detection range, lacking rapid, full-coverage scanning methods for large-area flue walls; second, detection results are mostly qualitative or semi-quantitative, lacking the ability to accurately quantify key indicators such as corrosion depth; third, it cannot effectively distinguish between surface artifacts and early internal corrosion, resulting in high false positive and false negative rates; and fourth, the detection data is disconnected from the spatial location of the flue, making it difficult to form an intuitive and traceable three-dimensional corrosion distribution map. Therefore, a new method is urgently needed to achieve early, rapid, quantitative, and visual identification of corrosion areas in desulfurization flues. Summary of the Invention

[0008] This application provides an early identification method for corrosion zones in desulfurization flues based on infrared thermal imaging, aiming to solve the problems of low detection efficiency, difficulty in quantification, easy misjudgment, and lack of spatial correlation in existing desulfurization flue corrosion detection technologies.

[0009] An early identification method for corrosion zones in desulfurization flues based on infrared thermal imaging, the method comprising:

[0010] S1. Multimodal thermal excitation and raw data acquisition: A transient temperature field is constructed on the wall of the desulfurization flue through a composite thermal excitation method, and a first set of thermal image sequences and a second set of thermal image sequences are acquired; wherein, the composite thermal excitation method includes electromagnetic induction heating and modulated laser heating, the first set of thermal image sequences is the temperature response data of the heating and initial cooling process acquired under electromagnetic induction heating, and the second set of thermal image sequences is the temperature response data of multiple preset frequencies acquired under modulated laser heating;

[0011] S2. Preprocessing and multi-dimensional feature extraction: The first set of thermal image sequences and the second set of thermal image sequences are preprocessed to eliminate noise interference, and electromagnetic thermal response features are extracted from the preprocessed first set of thermal image sequences, and variable frequency phase-locked features are extracted from the preprocessed second set of thermal image sequences. The electromagnetic thermal response features include a temperature gradient map and a thermal decay time constant map at the end of heating, and the variable frequency phase-locked features include a phase map under the optimal excitation frequency.

[0012] S3. Quantitative identification of corrosion areas based on physical models: The extracted electromagnetic thermal response features and the variable frequency phase-locked loop features are input into the pre-constructed heat transfer-corrosion physical model. Corrosion probability maps are generated through feature-level fusion to identify corrosion areas. The depth estimate of the corrosion area is obtained by inversion based on the quantitative relationship between the variable frequency phase-locked loop features and the corrosion depth.

[0013] S4. Three-dimensional spatial mapping and defect reconstruction: The two-dimensional corrosion feature map containing the corrosion area and depth estimate is mapped onto the three-dimensional geometric model of the desulfurization flue to generate a composite model that integrates quantitative corrosion information.

[0014] Optionally, in S1, the electromagnetic induction heating is achieved by passing an alternating current through a planar helical coil to generate eddy currents inside the metal wall of the flue to achieve heating; the modulated laser heating is achieved by modulating the output power of a laser to generate a laser beam of a preset waveform and projecting it onto the wall of the flue to achieve heating.

[0015] Optionally, in S1, during data acquisition, the control unit performs the following tasks according to a preset timing sequence: starting electromagnetic induction heating and simultaneously acquiring wall temperature changes at the maximum frame rate to generate the first set of thermal image sequences; pausing electromagnetic induction heating and waiting for the wall temperature to recover, then starting laser heating and sequentially applying multiple preset frequency modulated lasers, synchronously acquiring the wall temperature response, and generating the second set of thermal image sequences.

[0016] Optionally, in S2, the preprocessing includes: background subtraction to eliminate the effects of fixed background thermal radiation and thermal imager dark current offset; spatial domain filtering for noise reduction to suppress salt-and-pepper noise and preserve defect edges; and DC trend elimination to eliminate the heat accumulation effect of the heating source and ambient temperature drift.

[0017] Optionally, in step S2, extracting the variable frequency phase-locked feature includes: for each preset frequency in the second set of thermal image sequences, extracting the phase information of each pixel through phase-locked analysis to construct a differential phase map; analyzing the curve of the phase difference of each pixel changing with frequency, identifying blind frequencies, and selecting the frequency corresponding to the maximum root mean square value of the phase difference of all pixels as the optimal excitation frequency, and extracting the phase map at that frequency.

[0018] Optionally, in step S2, extracting the electromagnetic thermal response features includes: at the instant the electromagnetic induction heating ends, extracting the temperature field and calculating its spatial gradient amplitude to obtain the temperature gradient map at the moment the heating ends; and during the cooling stage after the electromagnetic induction heating stops, fitting the temperature decay curve of each pixel to obtain the thermal decay time constant map.

[0019] Optionally, in S3, the feature-level fusion includes: normalizing the temperature gradient map at the end of heating and the thermal decay time constant map in the electromagnetic thermal response features, and constructing the corrosion probability map by weighted summation, wherein the corrosion probability map is used to determine candidate points of the corrosion region.

[0020] Optionally, step S3 further includes: using the phase diagram at the optimal excitation frequency to perform secondary verification on the candidate points of the corrosion region, and determining the corrosion type as surface opening corrosion or internal early corrosion based on the combined response of the electromagnetic thermal response characteristics and the variable frequency phase-locked loop characteristics.

[0021] Optionally, in step S3, the inversion of the corrosion depth includes: extracting the correspondence between phase difference and defect depth from a pre-constructed feature-defect mapping database, and establishing phase-depth calibration curves for different corrosion types; for the identified corrosion area, extracting its phase difference, and substituting it into the corresponding calibration curve according to its corrosion type to perform numerical solution to obtain the depth estimate.

[0022] Optionally, in step S4, the three-dimensional spatial mapping and defect reconstruction includes: constructing a real-world three-dimensional geometric model of the desulfurization flue based on visible light images and positioning and attitude data collected by the UAV; establishing a mapping relationship between infrared images and visible light images, and between visible light images and three-dimensional spatial coordinates through heterogeneous image registration, accurately mapping the two-dimensional corrosion feature map to the corresponding surface position of the real-world three-dimensional geometric model; and assigning the corrosion depth value as a texture attribute to the corresponding three-dimensional spatial point to form the composite model.

[0023] Compared with the prior art, this application has at least the following beneficial effects:

[0024] This application employs a composite thermal excitation method, organically combining electromagnetic induction heating and modulated laser heating. Within the same detection process, it acquires electromagnetic thermal response data sensitive to wall thickness reduction and lock-in thermal response data sensitive to subsurface defects, constructing a multimodal, complementary raw data acquisition system. Compared to a single thermal excitation method, this approach can simultaneously capture the thermal characteristics of surface thinning and early internal corrosion, significantly improving the detection capability for different corrosion morphologies and laying a data foundation for subsequent high-reliability identification.

[0025] This application extracts multi-dimensional feature parameters such as phase map, temperature gradient map, and thermal decay time constant map from the original thermal image sequence through variable frequency phase-locked loop feature extraction and electromagnetic thermal response feature extraction. Among them, the phase feature is sensitive to deep defects and can effectively suppress the interference of surface emissivity changes, the thermal decay time constant is directly related to the wall thickness, and the temperature gradient can accurately characterize the corrosion boundary. The synergistic use of multi-dimensional features overcomes the shortcomings of single features being susceptible to noise interference and having unclear physical meaning under complex working conditions, and provides a feature input with high signal-to-noise ratio and strong robustness for quantitative identification.

[0026] This application introduces a quantitative identification mechanism based on a heat transfer-corrosion physical model. It establishes a mapping relationship between characteristic parameters and corrosion geometric parameters through a pre-constructed finite element simulation database, and employs a hierarchical fusion strategy to collaboratively analyze electromagnetic thermal response characteristics and phase-locked loop characteristics. This method can not only generate corrosion probability maps for accurate segmentation of corrosion regions, but also effectively distinguish between surface-opening corrosion and early-stage internal corrosion based on characteristic combination responses. Attached Figure Description

[0027] Figure 1 A flowchart of an early identification method for corrosion zones in desulfurization flues based on infrared thermal imaging, provided as an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0029] The method for early identification of corrosion zones in desulfurization flues based on infrared thermal imaging provided in this application is as follows: Figure 1 As shown, it includes the following steps:

[0030] S1: Multimodal thermal excitation and raw data acquisition. Through composite thermal excitation, a controllable and distinguishable transient temperature field is constructed on the wall of the desulfurization flue to be detected, providing a high signal-to-noise ratio raw thermal image sequence for subsequent corrosion feature extraction.

[0031] The inspection process utilizes an industrial-grade hexacopter UAV as its mobile platform. This platform is equipped with a high-precision attitude stabilization system and an RTK positioning module, ensuring stable flight in the complex environment outside the flue and maintaining the stability of the distance and angle between the inspection equipment and the flue wall. The inspection pod carried by the UAV integrates a composite thermal excitation module and a high-resolution infrared thermal imager.

[0032] The composite thermal excitation module is composed of an electromagnetic induction heating unit and a modulated laser heating unit working together. The two units use independent control channels and can work alternately or in combination according to a preset program.

[0033] Electromagnetic induction heating unit: This unit contains a planar helical coil wound with Litz wire, and a magnetic core is provided around the coil to enhance the magnetic field directivity. The coil is connected to a high-frequency inverter power supply mounted on the ground or a drone via a flexible shielded cable. During electromagnetic induction heating, an alternating current within a specific frequency range is passed into the coil, generating eddy currents inside the flue metal wall. Due to the skin effect, the eddy currents are mainly concentrated on the surface layer of the wall. The heating depth can be controlled by adjusting the excitation frequency, typically within the range of 1-3 mm. This heating method has advantages such as uniform heating, no contact with the wall surface, and less susceptibility to surface contaminants, making it suitable for rapidly screening large areas of wall thickness reduction. In practical applications, an optimal set of frequencies will be pre-calibrated based on the magnetic permeability and conductivity of the flue material for automatic configuration before testing.

[0034] Modulated Laser Heating Unit: This unit consists of a continuous fiber laser (wavelength 1064nm, maximum output power 50W) and a two-dimensional galvanometer scanning system. The laser output power is modulated by an arbitrary waveform generator, capable of generating sine waves, square waves, or user-defined periodic waveforms. The modulation frequency range is set from 0.01Hz to 10Hz to adapt to the thermal wave diffusion characteristics of defects at different depths. The galvanometer scanning system can project the laser beam onto the flue wall in point, line, or area patterns. By changing the scanning path and residence time, precise heating of localized micro-areas can be achieved. Compared to electromagnetic induction, laser heating has the advantages of high spatial resolution and programmable heat source morphology, making it particularly suitable for phase-locked analysis of suspected corrosion areas to obtain phase information and estimate corrosion depth.

[0035] The infrared thermal imager employs a cooled focal plane detector with a thermal sensitivity better than 20mK and a full-frame acquisition rate of no less than 100Hz. The optical axis of the thermal imager, the central axis of the electromagnetic coil, and the light output center of the laser scanning system are pre-calibrated to ensure that the three coincide in space, thereby guaranteeing pixel-level alignment accuracy of temperature response data for the same area under different excitation methods.

[0036] During data acquisition, the control unit performs the following tasks according to a preset timing sequence:

[0037] The electromagnetic induction heating unit is activated to continuously heat the flue wall at a selected frequency for a duration of 5 to 30 seconds. Simultaneously, the thermal imager acquires wall temperature changes at the maximum frame rate, recording the complete heating and initial cooling process and generating the first set of thermal image sequences. ;

[0038] Electromagnetic heating is paused, and once the wall temperature returns to ambient temperature, the laser heating unit is activated. Multiple sinusoidal modulated lasers at preset frequencies (e.g., 0.05Hz, 0.1Hz, 0.5Hz, 1Hz, 5Hz) are applied sequentially, with each frequency lasting for at least three complete modulation cycles. The thermal imager simultaneously acquires the wall temperature response, generating a second set of thermal image sequences. ;

[0039] All the above thermal image sequences, along with the corresponding excitation parameters (frequency, power, time) and UAV pose data (GPS coordinates, altitude, gimbal tilt angle), are stored in the airborne data storage unit as the raw data input for subsequent processing.

[0040] S2: Preprocessing and multi-dimensional feature extraction. Signal enhancement and feature decoupling are performed on the two sets of original thermal image sequences acquired in S1 to eliminate interference from non-defect factors (such as environmental noise, inherent noise of the thermal imager, non-uniform heating, and background thermal radiation) on the temperature signal. Multi-dimensional feature parameters sensitive to corrosion defects are extracted to provide a reliable data foundation for subsequent quantitative identification. This includes the following sub-steps:

[0041] S2.1: Image preprocessing, original thermal image sequence and Various noise components are inevitably introduced during the data acquisition process, including thermal imager readout noise, fluctuations in ambient thermal radiation, local variations in the emissivity of the flue gas duct wall surface, and spatial inhomogeneities of the heating source itself. To suppress these interferences, this step performs the following preprocessing operations sequentially:

[0042] Background subtraction: Before each heating cycle begins, 10 consecutive frames of thermal images in an unheated state are acquired, and their pixel-level average value is calculated as the background thermal image B(x,y). This background thermal image is then subtracted pixel by pixel from the thermal images at all subsequent time points to eliminate the effects of fixed background thermal radiation and thermal imager dark current offset.

[0043] Spatial domain denoising: Median filtering is used to perform spatial domain denoising on each frame of the thermal image. The filter window size is set to 3×3 or 5×5 pixels depending on the image resolution and defect size. Compared with mean filtering, median filtering can more effectively suppress salt-and-pepper noise while better preserving the sharpness of defect edges and avoiding edge blurring caused by smoothing.

[0044] DC trend elimination: Temperature time series curve for each pixel (x,y) A second-order polynomial was used to fit the overall trend, which reflects the heat accumulation effect of the heating source itself and the slow drift of the ambient temperature. Subtracting the fitted curve from the original temperature time series yields the temperature response signal T(x,y,t) with DC drift eliminated. The choice of the second-order polynomial is based on prior knowledge of the output characteristics of the thermal excitation source in actual testing, which can better balance fitting accuracy and computational efficiency.

[0045] After the above preprocessing, a pure temperature response signal T(x,y,t) is obtained. This signal contains only the temperature fluctuation component caused by the difference in internal thermal properties of the material (i.e. corrosion defects), providing a high signal-to-noise ratio data input for subsequent feature extraction.

[0046] S2.2: Variable frequency phase-locked loop feature extraction, targeting laser excitation data. Since the laser heat source has the characteristics of controllable waveform and adjustable frequency, phase lock analysis technology is used to extract the phase and amplitude information of each pixel to distinguish defect areas of different depths and different thermal properties.

[0047] The specific steps are as follows:

[0048] Multi-frequency phase-locked demodulation: For each preset excitation frequency Correlation operations are performed on the temperature response time series T(x,y,t) of each pixel (x,y) with reference sine and cosine signals of the same frequency. The sampling duration is assumed to be N complete cycles, and the sampling frequency is... Then phase and amplitude The following is calculated using the discrete Fourier transform:

[0049]

[0050]

[0051] Where M is the number of sampling points. This yields the phase diagram at each frequency. Amplitude diagram ;

[0052] Phase-frequency differential map construction: Select a corrosion-free area on the flue wall, confirmed visually or through prior inspection, as a reference area, and calculate the average phase value of all pixels within this area at each frequency. For each frequency Construct a differential phase diagram:

[0053]

[0054] It reflects the phase lag or lead of the defective region relative to the defect-free region. The larger the absolute value, the more significant the difference between the thermal properties of the region and the matrix.

[0055] Blind frequency identification and optimal frequency selection: analyzing each pixel The curve shows the change with frequency ff. When the excitation frequency is too high, the thermal diffusion length is less than the defect burial depth, and the thermal wave cannot effectively penetrate to the defect layer, leading to... The frequency approaching zero is called the "blind frequency." This is determined by iterating through the frequencies... Value, select all pixels The frequency corresponding to the maximum root mean square value The optimal excitation frequency is selected, and the phase diagram at that frequency is extracted. As the primary feature for identifying corrosion zones, phase features are sensitive to deep defects but insensitive to changes in surface emissivity, effectively suppressing artifacts caused by uneven heating.

[0056] S2.3: Electromagnetic thermal response feature extraction, targeting electromagnetic excitation data. Since electromagnetic induction heating primarily acts on the surface and near-surface layers, its temperature response characteristics are closely related to wall thickness reduction (i.e., material loss due to corrosion). This step extracts the following two types of characteristics:

[0057] Temperature gradient diagram at the end of heating Extract the temperature field at the instant electromagnetic heating ends (i.e., the first frame after the coil current is turned off). Calculate the spatial gradient magnitude of the temperature field:

[0058]

[0059] Because the local heat capacity decreases in the thinned wall region, the temperature rises faster under the same heat input, therefore... A distinct gradient peak appears at the boundary of the corrosion defect in the image. This feature can be used to quickly locate the edge of the corrosion area.

[0060] Thermal decay time constant diagram During the cooling phase after electromagnetic heating stops, for each pixel... Its temperature decay curve approximately follows an exponential decay law:

[0061]

[0062] in For ambient temperature, Let be the thermal decay time constant. A nonlinear least squares fitting algorithm is used to fit the cooling section temperature sequence of each pixel to obtain... The thermal decay time constant is positively correlated with the material's thermal conductivity and wall thickness: As the wall thickness decreases in the corrosion-thinned region, the heat capacity decreases, and heat dissipation is faster, manifesting as… The value decreases. This feature exhibits high sensitivity to uniformly thinned corrosion and is less affected by uneven surface heating;

[0063] Through the above steps, a multi-dimensional feature dataset is obtained. , as the input for S3 quantitative identification;

[0064] S3: Quantitative identification of corrosion regions based on physical models, using the multi-dimensional feature dataset extracted in step two. The data is input into a pre-built heat transfer-corrosion physical model. Through feature-level fusion and model inversion, accurate identification, type determination, and depth quantification of the corrosion region are achieved. The core of this step lies in mapping the measured temperature response characteristics with the physical simulation database, thereby giving the detection results a clear physical meaning and quantitative indicators, as detailed below:

[0065] (1) Heat transfer-corrosion model construction and eigenmap database establishment: To establish a quantitative relationship between eigenvalues ​​and corrosion geometric parameters, a series of heat transfer-corrosion numerical models reflecting the actual working conditions of the desulfurization flue were constructed in advance using finite element simulation software (such as ANSYS or COMSOL Multiphysics). The model construction process is as follows:

[0066] Geometric parametric modeling: A three-dimensional geometric model is established using the flue wall substrate material (Q235 carbon steel or weathering steel) as the object. Corrosion defects of different shapes and sizes are pre-defined on or inside the model surface, including:

[0067] Defect shapes are defined as three typical types: V-shaped (simulating sharp-bottomed pits formed by erosion corrosion), U-shaped (simulating round-bottomed pits formed by pitting corrosion or local thinning), and rectangular (simulating uniform thinning).

[0068] Defect depth d: The value range is set from 0.2 mm to 5.0 mm, covering the entire process from early pitting to severe thinning;

[0069] Defect width and width-to-depth ratio: The width value is set from 1 mm to 20 mm, and the width-to-depth ratio (width / depth) value is set from 0.5 to 10 to cover corrosion morphologies at different expansion stages;

[0070] Material properties and boundary conditions were assigned: The thermal properties of the matrix material, such as thermal conductivity, specific heat capacity, and density, were obtained through actual sampling and testing. The defect area was set to be filled with air (with a thermal conductivity much lower than that of metal) to simulate the thermal resistance effect of the medium within the corrosion pit. Heat flux boundary conditions were applied to the upper surface of the model, corresponding to two excitation methods: electromagnetic induction heating (uniform surface heat flux) and laser phase-locked loop heating (periodic modulated heat flux). The lower surface and sides were set as convective heat transfer boundaries with air, and the heat transfer coefficients were calibrated based on the actual operating environment of the flue (temperature, wind speed).

[0071] Solution and Eigenvalue Extraction: The transient heat conduction solver is used to calculate the temperature field evolution under different excitation conditions. For each combination of defect geometric parameters, the same eigenvalue as in step two is extracted: the temperature gradient at the end of heating. Thermal decay time constant during the cooling stage and the phase difference obtained from phase-locked analysis The aforementioned feature values ​​are then associated with their corresponding defect geometric parameters (shape, depth, width) to form a feature-defect mapping database. ;

[0072] The establishment of this database provides a physical basis for the subsequent inversion of measured data. For example, simulation results show that, under the same depth and aspect ratio, the U-shaped defect, due to its larger bottom curvature and stronger heat flow convergence effect, has a longer thermal decay time constant. The difference is 15%-30% higher than that of V-type defects; while the phase difference The sensitivity to defect shape decreases exponentially with increasing depth, and the sensitivity to defect shape is lower than that to depth. This characteristic provides a basis for the step-by-step quantization strategy.

[0073] (2) Multi-feature fusion recognition algorithm: In order to overcome the problem of false detection and false negative detection of single features under complex working conditions, a hierarchical fusion strategy is adopted to combine electromagnetic excitation features ( , ) and laser phase-locked loop characteristics ( Collaborative analysis is used to accurately identify and initially determine the type of corrosion areas;

[0074] Feature-level fusion and erosion probability map generation: and Normalization is performed on each feature to unify its value range to the [0,1] interval. The normalized features are denoted as follows: and Constructing a corrosion probability map :

[0075]

[0076] Among them, the weighting coefficient and The sensitivity to corrosion defects is set based on two types of characteristics. In this method, the sensitivity is set as follows: , The physical significance is that the temperature rises faster in areas where the wall thickness is reduced. (Increased) and dissipates heat faster ( (Decrease), and the two exhibit a superimposed enhancing effect in the probability diagram. When When the value exceeds the preset threshold (set according to the actual detection signal-to-noise ratio, typically 0.65), the pixel is determined to be a candidate point of the erosion region.

[0077] Phase verification and corrosion type identification: using phase diagrams Secondary verification was performed on the candidate corrosion regions. Phase characteristics are sensitive to deep defects but not to changes in surface emissivity, thus they can serve as an important basis for determining corrosion types.

[0078] If the candidate region simultaneously satisfies and Significant phase shift (relative to the reference region) occurs If the result is negative, it is determined to be surface opening corrosion (wall thickness reduction type defect).

[0079] If candidate region The change was not significant (<0.3), but Significant phase shift occurs ( If the corrosion rate is 0, it is determined to be early internal or subsurface corrosion (a buried defect that has not yet developed to the surface).

[0080] If the candidate region is only higher and If there is no significant change, it may be an artifact caused by surface dirt or uneven local heating, which can be eliminated through this verification step;

[0081] The above-mentioned fusion recognition algorithm effectively reduces the false alarm rate of single feature detection, and at the same time realizes the preliminary identification of corrosion type (surface corrosion / early internal corrosion), providing richer information for subsequent in-depth quantification and maintenance decision-making.

[0082] Furthermore, to adapt to fluctuations in excitation conditions and testing environments under different on-site working conditions, an adaptive adjustment mechanism for weighting coefficients is introduced. Specifically, before on-site testing, a non-corroded area on the flue wall is selected for pre-testing, and temperature response data under electromagnetic excitation is collected to calculate the average temperature gradient amplitude of this area. With average thermal decay time constant The measured values ​​are compared with the theoretical values ​​of the simulation model under the same material and working conditions to evaluate the actual signal-to-noise ratio level of the two types of features under the current detection conditions.

[0083] If actual measurement A value significantly lower than the theoretical value (e.g., below 70% of the theoretical value) indicates limited electromagnetic excitation effectiveness. Possible causes include insufficient power output, excessive distance between the coil and the wall, or the presence of a non-metallic coating on the wall reducing eddy current efficiency. In this case, the system automatically reduces... The weight was adjusted to the range of 0.2 to 0.3, and increased accordingly. The weights make the corrosion probability map more dependent on the thermal decay time constant feature, thereby maintaining the robustness of the fusion algorithm;

[0084] Conversely, if measured in actual... A value significantly higher than the theoretical value (e.g., 130% higher than the theoretical value) indicates that the cooling process is affected by ambient thermal radiation or airflow disturbance, resulting in a decrease in the signal-to-noise ratio of the thermal attenuation characteristic. The system then automatically improves this. The weights are set to the range of 0.5 to 0.6, making the probability map more dependent on temperature gradient features, so as to reduce the impact of cooling section noise on the recognition results;

[0085] The above adaptive adjustment process is automatically executed once before each detection task begins, and the adjusted weight coefficients are applied to the complete data processing flow of this detection task. The adjustment thresholds (such as 70%, 130%) and the upper and lower limits of the weights (0.2~0.6) can be modified through the system configuration interface according to the on-site calibration results to adapt to the differentiated needs of different detection scenarios;

[0086] By combining the above-mentioned strategy of fixed initial weights with adaptive dynamic adjustment, this invention, while ensuring the consistency of the algorithm, endows the detection system with the ability to adapt to different working conditions, effectively improving the construction quality of the corrosion probability map and the reliability of the identification results.

[0087] For identified corrosion areas, this invention uses the physical relationship between lock-in phase characteristics and depth to quantitatively invert corrosion depth by querying a simulation database.

[0088] Phase-Depth Calibration Curve Construction: From Feature Map Database In the process, the phase difference under all simulation conditions is extracted. The relationship between phase difference and defect depth d. Since the phase difference decreases approximately exponentially with depth, the following function is used for fitting:

[0089]

[0090] Where a, b, and c are fitting coefficients related to material thermophysical properties, excitation frequency, and defect shape. Independent calibration curves are established for different types of corrosion (V-shaped, U-shaped, and rectangular). In practical applications, since U-shaped corrosion is the most common, the U-shaped calibration curve is preferentially used as the default curve.

[0091] For each pixel (x, y) in the identified eroded region, extract its phase difference. Substituting this phase difference value into the corresponding calibration curve, the depth estimate is obtained through numerical solution. :

[0092]

[0093] If the phase difference of a certain pixel exceeds the fitting range of the calibration curve (i.e., less than the minimum simulated phase or greater than the maximum simulated phase), then the extrapolation method is used or it is marked as an area that needs to be manually verified.

[0094] Depth Correction and Confidence Assessment: Considering the potential noise and model bias in actual detection, this invention introduces a confidence factor C(x,y) to correct the depth estimation results. The confidence factor is determined by a combination of the following factors:

[0095] The higher the probability, the higher the confidence level.

[0096] Whether the phase verification step is passed, and the confidence level of the area that passes the verification is increased;

[0097] The consistency of depth estimation between adjacent pixels reduces the confidence of isolated outliers.

[0098] The final output is the estimated corrosion depth. The confidence level information is stored together in the corrosion feature map. In this context, it serves as the input for the S4 three-dimensional spatial mapping;

[0099] Through the above steps, the transformation from raw feature data to physically meaningful corrosion area identification, type discrimination, and depth quantification was achieved, providing a quantitative basis for subsequent 3D visualization and maintenance decision-making;

[0100] S4: Three-dimensional spatial mapping and defect reconstruction, converting the two-dimensional corrosion feature map generated in S3... Mapping the data to a three-dimensional spatial model of the desulfurization flue allows for the spatial localization, geometric quantification, and visualization of corrosion defects. By integrating the detection results with the actual flue structure, intuitive and accurate corrosion distribution information is provided to maintenance personnel, supporting subsequent maintenance decisions and condition assessments. This includes the following sub-steps:

[0101] S4.1: 3D geometric model construction. Based on visible light images and POS (positioning and attitude determination) data collected by the UAV during the same flight, an oblique photogrammetry technique is used to construct a real-world 3D geometric model of the desulfurization flue. The specific process is as follows:

[0102] Image Data Acquisition: In the flight mission of step one, in addition to the infrared thermal imager acquiring a sequence of thermal images, the UAV also carries a high-resolution visible light camera (with a resolution of no less than 20 million pixels) to acquire visible light images of the outer wall of the flue gas duct using a preset flight path and overlap rate (heading overlap rate of no less than 75% and lateral overlap rate of no less than 65%). The camera and the infrared thermal imager are rigidly connected and jointly calibrated to ensure that their optical axes are parallel and their relative positions are fixed, laying the foundation for subsequent registration of images from different sources.

[0103] POS data synchronization: The UAV flight control system records real-time POS data such as GPS coordinates, altitude, heading angle, pitch angle, and roll angle corresponding to each frame of visible light image, with timestamps and image frame numbers stored synchronously. The accuracy of POS data directly affects the geometric accuracy of 3D reconstruction. This invention uses real-time positioning (RTK) technology to improve positioning accuracy to the centimeter level.

[0104] Aerial triangulation and dense matching: The acquired visible light image sequence and POS data are imported into 3D reconstruction software (such as ContextCapture or Agisoft Metashape) for aerial triangulation. First, feature points (such as SIFT or AKAZE features) are extracted from each image, and corresponding points between images are established through feature matching. Combining the initial exterior orientation elements provided by the POS data, bundle adjustment is used to optimize the camera's interior orientation parameters and the exterior orientation parameters of each image. Based on this, a multi-view dense matching algorithm is used to generate 3D point cloud data.

[0105] 3D Model Generation: Based on dense point cloud data, a triangular mesh model is generated using surface reconstruction algorithms (such as Poisson surface reconstruction). The texture of the original visible light image is then mapped onto the mesh surface, forming a high-resolution real-world 3D geometric model with realistic texture. The model fully records the geometric features of the flue, including its external outline, cylinder slope, and platform location, serving as a spatial carrier for subsequent corrosion feature mapping.

[0106] S4.2: Texture mapping and data fusion, which integrates the two-dimensional erosion feature map generated in S3. Precise mapping to 3D geometric model Determining the corresponding surface location is the core technical challenge of this step. Because infrared thermal imagers and visible light cameras differ in field of view, resolution, and imaging band, and because there may be slight shifts in the acquisition time, it is necessary to establish a precise pixel-level mapping relationship.

[0107] Heterogeneous image registration: First, the infrared thermal imager and the visible light camera were jointly calibrated. A checkerboard calibration board was used to perform intrinsic parameter calibration (focal length, principal point coordinates, distortion coefficients) and stereo calibration (rotation matrix, translation vector) of the two sensors on the ground before detection. The calibration results recorded the transformation relationship between the pixel coordinates of the infrared image and the pixel coordinates of the visible light image. In actual testing, since the distance between the drone and the flue wall remains relatively constant during flight, this calibration relationship can be considered fixed in a single testing mission.

[0108] Feature point matching and coordinate association: For each frame of infrared thermal image, through calibration relationships... Locate its corresponding pixel region in the visible light image. Further, utilize the 3D model generated by S4. A mapping relationship between pixel coordinates in a visible light image and three-dimensional spatial coordinates (X, Y, Z) is established using collinearity equations. Combining these two steps, a method is achieved to extract features from a two-dimensional erosion image. The complete mapping from each pixel (x,y) to a point (X,Y,Z) in three-dimensional space;

[0109] Texture blending and data smoothing: adjusting erosion depth values Assigned as texture attributes to corresponding 3D spatial points. For cases where the same spatial location appears repeatedly in multiple frames, multiple depth estimates are fused using a weighted average method, with the weights determined based on the confidence factor C(x,y) of the pixel in the corresponding image. For texture-deficient regions caused by occlusion or registration errors, interpolation of adjacent triangular faces is used to fill in the missing areas, ensuring the continuity and integrity of the erosion features in the 3D model.

[0110] After the above processing, the three-dimensional geometric model Each surface unit (triangular facet or mesh vertex) is assigned a corrosion depth attribute, forming a composite model that incorporates quantitative corrosion information. .

[0111] S4.3: Visualization and labeling of corrosion areas. To facilitate on-site maintenance personnel in quickly understanding the corrosion status, this invention integrates the three-dimensional model... It presents information visually and provides automated information annotation capabilities;

[0112] Color stop rendering: Rendering is performed using a pseudo-color mapping method based on the numerical range of the erosion depth d. The color stop mapping rules are set as follows:

[0113] Light yellow ( ): This indicates early, minor corrosion, which is within an acceptable range;

[0114] orange color( ): This indicates moderate corrosion and should be included in the monitoring plan;

[0115] Dark red ( ): This indicates severe corrosion; repairs are recommended soon.

[0116] purplish-black ( ): This indicates critical corrosion, requiring immediate attention.

[0117] The color mark mapping rule can be customized and modified in the system configuration interface according to the actual engineering criteria;

[0118] Corrosion Region Identification and Segmentation: Based on Corrosion Probability Map The threshold segmentation results are used to automatically extract continuous eroded regions and connected components from the 3D model. For each independent connected component, the system automatically calculates the following geometric parameters:

[0119] Corrosion area: obtained by summing the areas of all triangular facets within the region, in square meters (m²). 2 );

[0120] Average erosion depth: The arithmetic mean of the depth estimates of all pixels in this area, in millimeters (mm).

[0121] Maximum corrosion depth: The maximum value of the estimated depth within this area;

[0122] Spatial coordinates: Three-dimensional coordinates of the geometric center of the corroded region ( Establish a coordinate system with the center of the bottom of the flue as the origin;

[0123] Annotation and Report Output: In the 3D model interface, the area and average depth information of each corroded area are annotated as floating labels. Users can view detailed parameters of this area through interactive operations (mouse hover or click). The system supports generating inspection reports, including:

[0124] A three-dimensional model view of the entire flue, showing the corrosion distribution using color-coded rendering;

[0125] A summary table of corroded areas lists the area, average depth, maximum depth, and spatial coordinates of each area.

[0126] Focus on magnified views of key areas and their corresponding original thermal images;

[0127] A deep quantification confidence distribution map helps users identify areas that require manual review.

[0128] Through the above steps, the process of transforming raw inspection data into a 3D visualized inspection report is achieved. The complete data flow provides an intuitive and accurate basis for decision-making regarding the assessment of the corrosion status of desulfurization flues and the formulation of maintenance strategies.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging, characterized in that, The method includes: S1. Multimodal thermal excitation and raw data acquisition: A transient temperature field is constructed on the wall of the desulfurization flue through a composite thermal excitation method, and a first set of thermal image sequences and a second set of thermal image sequences are acquired; wherein, the composite thermal excitation method includes electromagnetic induction heating and modulated laser heating, the first set of thermal image sequences is the temperature response data of the heating and initial cooling process acquired under electromagnetic induction heating, and the second set of thermal image sequences is the temperature response data of multiple preset frequencies acquired under modulated laser heating; S2. Preprocessing and multi-dimensional feature extraction: The first set of thermal image sequences and the second set of thermal image sequences are preprocessed to eliminate noise interference, and electromagnetic thermal response features are extracted from the preprocessed first set of thermal image sequences, and variable frequency phase-locked features are extracted from the preprocessed second set of thermal image sequences. The electromagnetic thermal response features include a temperature gradient map and a thermal decay time constant map at the end of heating, and the variable frequency phase-locked features include a phase map under the optimal excitation frequency. S3. Quantitative identification of corrosion areas based on physical models: The extracted electromagnetic thermal response features and the variable frequency phase-locked loop features are input into the pre-constructed heat transfer-corrosion physical model. Corrosion probability maps are generated through feature-level fusion to identify corrosion areas. The depth estimate of the corrosion area is obtained by inversion based on the quantitative relationship between the variable frequency phase-locked loop features and the corrosion depth. S4. Three-dimensional spatial mapping and defect reconstruction: The two-dimensional corrosion feature map containing the corrosion area and depth estimate is mapped onto the three-dimensional geometric model of the desulfurization flue to generate a composite model that integrates quantitative corrosion information.

2. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In S1, the electromagnetic induction heating is achieved by passing an alternating current through a planar spiral coil to generate eddy currents inside the metal wall of the flue to achieve heating; the modulated laser heating is achieved by modulating the output power of a laser to generate a laser beam with a preset waveform and projecting it onto the wall of the flue to achieve heating.

3. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In step S1, during data acquisition, the control unit performs the following tasks according to a preset timing sequence: starting electromagnetic induction heating and simultaneously acquiring wall temperature changes at the maximum frame rate to generate the first set of thermal image sequences; pausing electromagnetic induction heating and waiting for the wall temperature to recover, then starting laser heating and sequentially applying multiple preset frequency modulated lasers, synchronously acquiring the wall temperature response, and generating the second set of thermal image sequences.

4. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In S2, the preprocessing includes: background subtraction to eliminate the effects of fixed background thermal radiation and thermal imager dark current offset; spatial domain filtering to suppress salt-and-pepper noise and preserve defect edges; and DC trend elimination to eliminate the heat accumulation effect of the heating source and ambient temperature drift.

5. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In step S2, extracting the variable frequency phase-locked features includes: for each preset frequency in the second set of thermal image sequences, extracting the phase information of each pixel through phase-locked analysis and constructing a differential phase map; analyzing the curve of the phase difference of each pixel changing with frequency, identifying blind frequencies, and selecting the frequency corresponding to the maximum root mean square value of the phase difference of all pixels as the optimal excitation frequency, and extracting the phase map at that frequency.

6. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In step S2, extracting the electromagnetic thermal response features includes: at the instant the electromagnetic induction heating ends, extracting the temperature field and calculating its spatial gradient amplitude to obtain the temperature gradient map at the moment the heating ends; and during the cooling stage after the electromagnetic induction heating stops, fitting the temperature decay curve of each pixel to obtain the thermal decay time constant map.

7. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In S3, the feature-level fusion includes: normalizing the temperature gradient map at the end of heating and the thermal decay time constant map in the electromagnetic thermal response features, and constructing the corrosion probability map by weighted summation, wherein the corrosion probability map is used to determine candidate points of the corrosion region.

8. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 7, characterized in that, S3 further includes: using the phase diagram at the optimal excitation frequency to perform secondary verification on the candidate points of the corrosion region, and determining the corrosion type as surface opening corrosion or internal early corrosion based on the combined response of the electromagnetic thermal response characteristics and the variable frequency phase-locked loop characteristics.

9. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In step S3, the inversion of the corrosion depth includes: extracting the correspondence between phase difference and defect depth from a pre-constructed feature-defect mapping database, and establishing phase-depth calibration curves for different corrosion types; for the identified corrosion area, extracting its phase difference, and substituting it into the corresponding calibration curve according to its corrosion type to perform numerical solution to obtain the depth estimate.

10. The method for early identification of corrosion zones in desulfurization flue gas ducts based on infrared thermal imaging according to claim 1, characterized in that, In step S4, the three-dimensional spatial mapping and defect reconstruction includes: constructing a real-world three-dimensional geometric model of the desulfurization flue based on visible light images and positioning and attitude data collected by the UAV; establishing a mapping relationship between infrared images and visible light images, as well as between visible light images and three-dimensional spatial coordinates, through heterogeneous image registration, and accurately mapping the two-dimensional corrosion feature map to the corresponding surface position of the real-world three-dimensional geometric model; and assigning the corrosion depth value as a texture attribute to the corresponding three-dimensional spatial point to form the composite model.

Images