Method for detecting blockage of spray nozzles of a spray layer of a desulfurization tower based on machine vision

Abstract

The application discloses a method for detecting nozzle blockage of a spray layer of a desulfurization tower based on machine vision, which comprises the following steps: S1, image acquisition and preprocessing; S2, difference feature extraction and suspected area positioning; S3, blockage state verification based on fluid simulation; and S4, blockage degree quantification and detection result output. Firstly, the pixel-level difference between a real-time image and a reference image is extracted by a deep neural network to quickly lock the geometric features and spatial position of a suspected blockage area. On this basis, the visual detection result is converted into a quantifiable fluid mechanics model, and key flow field parameters such as pressure drop, flow rate and pressure distribution reflecting the actual working performance of the nozzle are obtained through computational fluid dynamics simulation. This mechanism overcomes the defect that pure visual detection is prone to false detection caused by light and fog interference, and makes up for the deficiency that pure physical models cannot be quickly positioned, realizes double confirmation of apparent abnormalities and physical performance, and greatly improves the accuracy of blockage determination.

Description

Technical Field

[0001] This application belongs to the field of desulfurization tower spray layer detection technology, specifically involving a machine vision-based method for detecting nozzle blockage in desulfurization tower spray layers. Background Technology

[0002] Wet flue gas desulfurization (FGD) technology is the mainstream process for flue gas treatment in coal-fired power plants. The operating status of the nozzles in the spray layer of the desulfurization tower directly affects the desulfurization efficiency and system energy consumption. The nozzles in the spray layer are exposed to a high-concentration slurry, high-temperature and high-humidity, and acidic corrosive environment for a long time, which makes them prone to scaling, crystallization, or blockage by foreign objects. This leads to a decrease in nozzle atomization performance, uneven spray coverage, and consequently, reduced desulfurization efficiency, increased local corrosion, or even system shutdown.

[0003] Currently, the detection of nozzle blockage in the spray layer of desulfurization towers mainly relies on the following methods:

[0004] Indirect monitoring based on pressure or flow rate. This method collects pressure and flow rate signals from the main or branch sprinkler pipes and combines them with historical data to determine if the sprinkler system is abnormal. However, this method only reflects the overall condition of the entire pipeline or sprinkler layer and cannot pinpoint the specific blocked nozzle, let alone distinguish the degree of blockage of an individual nozzle. Furthermore, it is easily affected by factors such as fluctuations in pipeline pressure loss and changes in slurry density, resulting in a high false alarm rate.

[0005] Visual inspection based on image processing. Some existing technologies attempt to use industrial cameras to capture images of nozzles and identify blockages through image segmentation or template matching. However, the actual spray layer has uneven lighting and severe mist scattering, making it easy to produce false positives or false negatives by relying solely on image features. More importantly, visual inspection can only obtain surface information of the nozzle end face and cannot know the actual impact of blockage on the internal flow field, making it difficult to accurately quantify the degree of blockage. This results in inspection results lacking physical basis and insufficient credibility for maintenance decisions.

[0006] In summary, existing nozzle blockage detection methods suffer from problems such as poor real-time performance, inaccurate positioning, and lack of physical quantitative evidence. There is an urgent need for an intelligent detection method that can achieve non-contact, precise positioning, and physical verification capabilities. Summary of the Invention

[0007] This application provides a machine vision-based method for detecting nozzle blockage in the spray layer of a desulfurization tower, aiming to solve the problems of existing technologies being susceptible to interference, having a high false alarm rate, being difficult to accurately quantify the degree of blockage, resulting in a lack of physical basis for detection results and insufficient credibility of maintenance decisions.

[0008] A machine vision-based method for detecting nozzle blockage in the spray layer of a desulfurization tower, the method comprising:

[0009] S1. Image acquisition and preprocessing: Acquire real-time images of the nozzles of the spray layer and pre-stored reference images, and preprocess the real-time images and reference images respectively to obtain preprocessed real-time images and preprocessed reference images with spatial alignment, consistent grayscale and noise suppression.

[0010] S2. Difference feature extraction and suspected region localization: The preprocessed real-time image and the preprocessed reference image are input into a pre-trained difference detection network to obtain a difference score map that characterizes the degree of pixel-level difference; the difference score map is processed to locate and extract the geometric features and location information of at least one suspected blockage area on the nozzle end face;

[0011] S3. Verification of clogging status based on fluid simulation: Based on the geometric features and location information of the suspected clogging area, construct a nozzle physical model containing clogging features; perform computational fluid dynamics simulation on the nozzle physical model to obtain simulation feature quantities that reflect the working performance of the nozzle;

[0012] S4. Quantification of blockage degree and output of detection results: Based on the simulation feature quantity, calculate the blockage influence coefficient used to comprehensively characterize the blockage degree, compare the blockage influence coefficient with a preset threshold, and output the blockage detection result according to the comparison result.

[0013] Optionally, S1 includes:

[0014] Obtain the pre-stored reference image Real-time images are collected at preset intervals during the operation of the desulfurization tower. ;

[0015] For the reference image With the real-time image The preprocessed reference image is obtained by sequentially performing size normalization, grayscale conversion, and filtering denoising. With preprocessed real-time images ;

[0016] During image acquisition, the illuminance of the coaxial light source is adjusted in a closed loop to maintain the overall grayscale average of the image within a preset range.

[0017] Optionally, S2 includes:

[0018] The preprocessed reference image With preprocessed real-time images After concatenation along the channel dimensions, the data is input into the difference detection network, and a difference score map is obtained through forward propagation. ;

[0019] For the difference score map Binarization and connected component analysis are performed to extract the geometric features and location information of each connected component. The geometric features and location information include the minimum bounding rectangle, the area of ​​the outline, the perimeter of the outline, the centroid coordinates, and the region shape description.

[0020] Based on a preset pixel-physical scale mapping coefficient, the geometric features and location information are converted into the actual area, width, and height of the blockage area in terms of physical dimensions.

[0021] Optionally, the difference detection network is trained using a virtual sample training strategy, and the training data construction process includes:

[0022] Select images from publicly available image datasets as background templates;

[0023] Generate a binary mask image corresponding to the background template, and randomly generate irregular polygonal regions in the mask image as simulated abnormal regions;

[0024] The background template is copied as the image to be inspected, and the image to be inspected is abnormally synthesized according to the mask image. The synthesis method includes filling the mask area with random RGB colors, or cropping the pixel block corresponding to the mask area from another image and embedding it with Poisson fusion.

[0025] Optionally, S3 includes:

[0026] Based on the geometric features and location information of the suspected blockage area, a blockage body is constructed on the original geometric model of the nozzle to form the nozzle physical model. ;

[0027] Set simulation boundary conditions and medium properties for the nozzle physical model. Perform flow field simulation calculations;

[0028] After the simulation converges, simulation features are extracted, including the pressure drop across the nozzle. Average outlet velocity and the coefficient of variation of pressure distribution in the sprinkler coverage area .

[0029] Optionally, the construction rule of the blockage body is:

[0030] When the aspect ratio of the suspected congestion area Between 0.8 and 1.2 and with roundness When the value is greater than 0.7, the blockage body is constructed as a spherical or cylindrical protrusion;

[0031] When the aspect ratio When the value is greater than 2.0, the blockage body is constructed as a cuboid or elliptical cylinder structure;

[0032] For other shapes, fit them as irregular polyhedra.

[0033] Optionally, S4 includes:

[0034] Obtain pre-stored reference flow field parameters of the nozzle under normal operating conditions, including the pressure drop across the nozzle under normal operating conditions. Average outlet velocity and the coefficient of variation of pressure distribution in the sprinkler coverage area ;

[0035] Based on the relative deviation between the simulated characteristic quantities and the reference flow field parameters, a weighted fusion method is used to calculate the blockage influence coefficient. .

[0036] Optionally, the blockage influence coefficient The calculation formula is:

[0037] in, , , Let be the weight coefficient, and satisfy... ;

[0038] The method further includes: dynamically adjusting the weighting coefficient based on the morphological characteristics of the suspected congestion area. , , The value of .

[0039] Optionally, in step S4, the detection results include the location of the blockage, the geometry of the blockage, the blockage influence coefficient, a comparison of key flow field parameters, and maintenance recommendations.

[0040] The method further includes: storing the detection results in a structured data format in a historical database and pushing them to the visualization interface of the monitoring system;

[0041] The real-time images, difference score maps, and simulated flow field clouds from this test were used. Figure 1 And archive it.

[0042] Compared with the prior art, this application has at least the following beneficial effects:

[0043] This application first extracts pixel-level differences between real-time and reference images using a deep neural network to quickly pinpoint the geometric features and spatial location of suspected blockage areas. Based on this, the visual detection results are transformed into a quantifiable fluid dynamics model. Through computational fluid dynamics simulation, key flow field parameters such as pressure drop, flow velocity, and pressure distribution, which reflect the actual working performance of the nozzle, are obtained. This mechanism overcomes the shortcomings of pure visual detection, which is susceptible to interference from lighting and fog, leading to false detections. At the same time, it makes up for the inability of pure physical models to quickly locate the blockage, achieving dual verification of apparent anomalies and physical performance, and significantly improving the accuracy of blockage determination.

[0044] This application constructs a blockage impact coefficient and performs multi-parameter weighted fusion of visually extracted geometric features (area, morphology) and simulated flow field parameters (pressure drop change, velocity decay, and spray uniformity deterioration) to ultimately output a quantified blockage level and specific maintenance recommendations. This quantification method gives the detection results clear physical meaning, providing objective data support for maintenance personnel to determine the timing of cleaning and assess the severity of blockages, avoiding the uncertainty caused by subjective experience.

[0045] In actual desulfurization tower operation, it is difficult to obtain a large number of nozzle blockage samples, and the morphology varies greatly, making traditional supervised training methods unsuitable. This application, during the training phase of the difference detection network, uses data synthesis techniques such as randomly generated irregular masks and Poisson fusion to simulate blockage images with varying shapes, sizes, and locations. Combined with data augmentation techniques such as brightness adjustment and noise addition, the network can learn robust difference features without requiring real blockage samples. This strategy not only reduces data acquisition costs but also improves the network's generalization ability to unknown blockage morphologies.

[0046] This application uses pre-calibrated camera parameters and working distance to accurately convert pixel coordinates in the image into physical coordinates on the nozzle end face, achieving precise location of blockages and facilitating quick identification by maintenance personnel. Furthermore, during the calculation of the blockage impact coefficient, the weighting coefficients of indicators such as pressure drop, flow velocity, and pressure distribution are dynamically adjusted according to the blockage morphology (circular, strip-shaped, or irregular), making the evaluation model more closely reflect the actual impact of different blockage types on spray performance and further enhancing the reliability of the evaluation results. Attached Figure Description

[0047] Figure 1 This is a flowchart of a machine vision-based method for detecting nozzle blockage in a desulfurization tower spray layer, provided as an embodiment of this application. Detailed Implementation

[0048] 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.

[0049] The machine vision-based method for detecting nozzle blockage in the spray layer of a desulfurization tower provided in this application is as follows: Figure 1 As shown, it includes the following steps:

[0050] Step S1: Image acquisition and preprocessing, providing standardized image data with spatial alignment, consistent grayscale, and noise suppression for subsequent differential feature extraction;

[0051] First, construct the image acquisition unit, including:

[0052] A high-resolution industrial area scan camera (5472×3648 pixels);

[0053] A fixed-focus optical lens with a 12mm focal length;

[0054] Coaxial ring light source system;

[0055] The coaxial ring light source is installed at the front of the lens, providing uniform illumination that is consistent with the camera's optical axis, effectively suppressing uneven lighting caused by reflections from the metal surface and scattering of mist within the spray layer.

[0056] The acquisition unit is fixed to the end of the adjustable pose robotic arm. During system initialization, the working parameters are precisely calibrated using a laser rangefinder: the distance from the camera optical center to the nozzle end face is maintained at 150±2mm, and the parallelism between the imaging plane and the nozzle end face is controlled within 0.5° to ensure the accuracy and consistency of image acquisition;

[0057] During the system initialization phase, the target nozzles, where the spray layer is clean and unblocked during the desulfurization tower shutdown and maintenance, are sequentially moved by the robotic arm and camera to the preset acquisition points of each nozzle, and the images are captured and saved as reference images, denoted as . Where i is the nozzle number, the image format is BMP, and the bit depth is 8 bits. After the reference image is acquired, the relative pose parameters of the camera and the robotic arm, and the spatial coordinates of each nozzle are written into the system calibration file as the reference for subsequent real-time image acquisition;

[0058] During normal operation of the desulfurization tower, the system reads the calibration file according to the set inspection cycle (e.g., every 2 hours), controls the robotic arm to repeat the above posture trajectory, and sequentially performs real-time image acquisition on each nozzle to obtain real-time images. During the acquisition process, the ambient light intensity is monitored simultaneously. When the light fluctuation is detected to exceed a preset threshold (such as a change of 5%), the light source controller is triggered to perform closed-loop adjustment of the illuminance of the coaxial light source, so that the overall grayscale average of the image is maintained within the preset range of 128±15.

[0059] After image acquisition is complete, the following preprocessing steps are performed on each set of reference images and real-time images:

[0060] Size normalization: The original image is uniformly scaled to 640×480 pixels using a bilinear interpolation algorithm. This size matches the input dimension of the difference detection network in subsequent step S2, while preserving the local detail features of the nozzle end face and orifice area;

[0061] Grayscale processing: The RGB three-channel image is converted into a grayscale image using a weighted average method. The conversion formula is G=0.299R+0.587G+0.114B, in order to eliminate the interference of color information on difference detection and make the network more focused on structural changes.

[0062] Filtering and Denoising: The grayscale image is smoothed by applying median filtering and Gaussian filtering in sequence. The median filter kernel size is 3×3 pixels, which is used to remove salt-and-pepper noise caused by smoke or fog; the Gaussian filter kernel size is 5×5 pixels, with a standard deviation of 1.2, which is used to suppress Gaussian noise introduced by the camera sensor while preserving high-frequency details such as nozzle edges.

[0063] After the above preprocessing, a preprocessed reference image is obtained. With preprocessed real-time images The two images maintain consistency in spatial scale, grayscale dynamic range, and noise level, forming the input image pair for the difference detection network in step S2. Furthermore, the preprocessed image resolution, camera intrinsic parameters, and working distance parameters are synchronously transmitted to the fluid simulation module in step S3 to establish a pixel-to-physical scale mapping relationship between the nozzle in the simulation physical model and the nozzle in the actual image (e.g., 1 pixel corresponds to 0.05 mm physical size), ensuring that the geometric features of the suspected blockage area detected visually can be accurately converted into the boundary conditions of the simulation model.

[0064] Step S2: Difference feature extraction and suspected region localization. A deep neural network is used to extract the pixel-level differences between the preprocessed real-time image output in step S1 and the preprocessed reference image, and the suspected blockage area is located based on the difference score map, providing geometric features and spatial location information for subsequent fluid simulation verification.

[0065] First, a difference detection network is constructed. This network adopts an encoder-decoder symmetric architecture. The encoder part contains 10 convolutional blocks, each consisting of a convolutional layer, an instance normalization layer, and a Gaussian error linear unit activation function connected in series. The decoder part consists of 4 deconvolutional blocks, 5 convolutional layers, and 5 transposed convolutional layers. During the decoding stage, a skip connection is used to concatenate the feature maps of the corresponding layers of the encoder with the feature maps of the current layer of the decoder along the channel dimension, achieving multi-scale feature fusion. The network input layer size is 640×480×2, corresponding to the preprocessed reference image and the preprocessed real-time image after concatenation along the channel dimension. After the network output layer is activated by the Sigmoid function, the output is a difference score map with a size of 640×480×1, where the value of each pixel represents the degree of difference between the two images at that location, with a value range of [0,1]. The closer to 1, the greater the difference.

[0066] To overcome the problem of scarce and diverse blockage samples in real-world industrial scenarios, the difference detection network employs a virtual sample training strategy. The training data construction process is as follows: An image is randomly selected from a public image dataset (such as VOC2017) as a background template, and its size is uniformly scaled to 640×480 pixels. A binary mask image corresponding to this template size is generated. One to four irregular polygonal regions are randomly generated in the mask image, with 3 to 8 vertices and an area of ​​50 to 2000 pixels. The pixel values ​​of the corresponding regions in the mask are set to 1, and the rest to 0. The template image is copied as the image to be inspected, and anomaly synthesis is performed on the image to be inspected based on the mask image. Two synthesis methods are used: First, the mask region is filled with random RGB colors (each channel value is uniformly distributed between 0 and 255). Second, another image is randomly selected from the training dataset, and the pixel blocks corresponding to the mask region in that image are cropped and embedded into the image to be inspected using Poisson fusion. In addition, during the training phase, data augmentation is applied simultaneously to the template image, the image to be inspected, and the mask image. The augmentation methods include random brightness adjustment (adjustment coefficient 0.7 to 1.3), random rotation (-15° to 15°), random translation (-10 to 10 pixels), random addition of Gaussian noise (standard deviation 0 to 0.05), and random lighting simulation, in order to improve the network's robustness to lighting fluctuations, camera micro-vibrations, and fog interference under actual working conditions.

[0067] The network training employs a binary cross-entropy loss function, combined with label smoothing regularization, with a smoothing coefficient set to 0.05. The loss function expression is as follows:

[0068]

[0069] Where H and W are the image height and width, respectively. The labels are the true difference labels after label smoothing. This represents the difference score predicted by the network. The training process uses the AdamW optimizer with an initial learning rate of 0.001, a batch size of 8, and 10,000 iterations. After network training is complete, the weight parameters of the encoder and decoder are fixed for subsequent inference stages.

[0070] During the inference phase, the preprocessed reference image output in step S1 is... With preprocessed real-time images The data is concatenated along the channel dimension to form a tensor of size 640×480×2, which is then input into a trained dissimilarity detection network. After forward propagation, the dissimilarity score map is obtained. Its dimensions are 640×480×1;

[0071] right Binarization is performed. An adaptive threshold segmentation method is used to... The segmentation threshold is 1.2 times the global mean. , will be greater than or equal to The pixels with the specified value are set to 1, and the rest are set to 0, to obtain a binary mask image. .right Perform connected component analysis to extract all connected regions with a pixel value of 1, and denote the k-th connected region as . ;

[0072] For each connected component Record the following geometric features and location information:

[0073] Minimum bounding rectangle: Calculate the bounding rectangle Find the bounding rectangle with the smallest area and record the coordinates of its center point. ,width With height ;

[0074] Outline area: statistics Total number of pixels ;

[0075] Perimeter of the outline: Calculation Number of boundary pixels ;

[0076] Centroid coordinates: Calculation The first moment is used to obtain the centroid coordinates. ;

[0077] Region shape description: based on the aspect ratio of the minimum bounding rectangle. and roundness Preliminary identification of the morphological characteristics of the blocked area (such as approximately circular, strip-shaped, or irregular);

[0078] All the above geometric features are recorded in pixels. Using the pixel-to-physical scale mapping coefficients ss (unit: mm / pixel) pre-calibrated in step S1, the pixel size is converted to the physical size to obtain the actual area of ​​the blockage region. Actual width and actual height Simultaneously, based on the camera-in-nozzle spatial coordinates calibrated in step S1, the pixel coordinates are... Converted to physical coordinates on the nozzle end face This is used to locate the specific blockage position of the nozzle in step S3;

[0079] The aforementioned geometric features and location information are organized into structured data, which serves as the input parameters for constructing the physical model of local nozzle blockage in step S3. For the minimum circumscribed rectangle aspect ratio... Approximately 1 (i.e., approximately circular) and area Exceeding a preset threshold (e.g., 2mm) 2 Connected components are preferentially marked as high-confidence suspected blockage regions; for connected components with aspect ratios deviating from 1 or with small areas, lower initial weights are assigned during simulation verification in step S3 to avoid wasting computational resources due to false detections.

[0080] Step S3: Verification of clogging status based on fluid simulation. The suspected clogging area located in step S2 is transformed into a quantifiable fluid dynamics model. Flow field parameters reflecting the actual working performance of the nozzle are obtained through computational fluid dynamics simulation, providing a physical basis for the final determination of the degree of clogging.

[0081] First, a physical model of the localized nozzle blockage is constructed. Based on the geometric features and location information of the k-th connected region recorded in step S2, a 3D simulation model corresponding to the suspected blockage area is established. The modeling process is as follows: Using the original design drawings of the target nozzle in the desulfurization tower spray layer as a reference, a complete geometric model of the nozzle is constructed in the 3D modeling software. This model includes the nozzle inlet section, constriction section, orifice section, and outlet extension section. The orifice diameter is set to 3.0 mm according to the actual nozzle specifications, and the total nozzle length is 25 mm. The centroid physical coordinates of the suspected blockage area output in step S2 are then used to... Map to the corresponding position on the nozzle exit face, and based on the width of the minimum bounding rectangle. With height And contour shape description parameters (aspect ratio) Circularity Construct the three-dimensional geometry of the blockage. The construction rules for the blockage are: when... Between 0.8 and 1.2 and When the value is greater than 0.7, it is judged as an approximately circular blockage, and constructed as a spherical or cylindrical protrusion; when When the value is greater than 2.0, it is determined to be a strip-shaped blockage, and a cuboid or elliptical cylindrical structure is constructed; for other shapes, an irregular polyhedron is fitted. The adhesion height of the blockage at the nozzle outlet end face is set as the ratio of the statistical mean of the profile perimeter to the nozzle diameter multiplied by 0.2 mm to reflect the protrusion degree of the actual blockage. The constructed blockage is then subjected to Boolean difference or intersection operations with the original nozzle geometric model to form a nozzle physical model containing local blockage characteristics, denoted as . ;

[0082] Secondly, simulation boundary conditions and medium properties are set. The nozzle physical model is then... The flow field was analyzed using a steady-state solver in computational fluid dynamics simulation software (such as COMSOL Multiphysics or ANSYS Fluent). The fluid medium properties were set according to the actual operating conditions of the desulfurization tower: the medium was desulfurization slurry or flue gas, with a density of... (Regarding flue gas) or (For slurry), dynamic viscosity (Smoke) or (Slurry). The inlet boundary condition is set to pressure inlet, and the inlet pressure value is set according to the actual gas or liquid supply system parameters. (Gauge pressure); The outlet boundary condition is set as a pressure outlet, and the outlet pressure is ambient atmospheric pressure. (Gauge pressure). The wall condition is set as a no-slip wall, and the standard wall function is used to handle the near-wall flow. The standard turbulence model is selected. The model, where the turbulent kinetic energy k and the turbulent dissipation rate The initial value is determined based on the inlet hydraulic diameter and turbulence intensity;

[0083] Flow field simulation calculations were then performed. An unstructured tetrahedral mesh was used to divide the computational domain, with mesh refinement applied near the nozzles and clogging bodies. The minimum mesh size was set to 0.05 mm to ensure accurate capture of flow field details around the clogging bodies. Mesh quality was controlled using skewness and aspect ratio, with skewness less than 0.85 and aspect ratio less than 5. The solver employed a pressure-velocity coupled algorithm (SIMPLE or PISO), and second-order upwind discretization was used to improve computational accuracy. The convergence criterion was set to a minimum residual curve density of 10. -5 Below, and the export quality flow monitoring value tends to stabilize (relative change less than 0.1%).

[0084] After the simulation converges, the following flow field parameters are extracted as simulation characteristic quantities of the nozzle under the clogging state:

[0085] Pressure drop across the nozzle ,in The average total pressure at the nozzle inlet cross-section. The average static pressure at the nozzle outlet cross section;

[0086] Average outflow velocity ,in The nozzle outlet cross-sectional area is... For axial velocity components;

[0087] The pressure distribution over the spray coverage area was analyzed, and the pressure contour map at a cross-section 10 mm downstream of the nozzle was extracted to calculate the coefficient of variation of the pressure distribution. ,in This represents the standard deviation of the pressure at that cross section. This represents the average pressure.

[0088] Local flow velocity in the blocked area Extract the maximum velocity value near the surface of the blockage;

[0089] The simulated flow field parameters described above, along with the geometric features of the suspected blockage region recorded in step S2, are stored together to form the complete feature vector corresponding to the connected domain. If a nozzle in step S2 has multiple connected regions (i.e., multiple suspected blockage points), the above modeling and simulation process is performed on each connected region, and the relative positional relationships of each blockage point are marked to support the subsequent evaluation of the coupling effect of multiple blockages. All simulation results will be passed to step S4 as input for quantifying the degree of blockage and making the final judgment.

[0090] Step S4: Quantification of blockage degree and output of detection results. Based on the simulation flow field parameters obtained in step S3, a quantitative evaluation model for the blockage degree is established. The blockage influence coefficient is calculated by multi-parameter fusion and compared with the preset threshold. Finally, the detection results with clear physical meaning are output, providing a quantitative basis for the maintenance decision of the desulfurization tower spray system.

[0091] First, a baseline library of flow field parameters under normal operating conditions is established. When the desulfurization tower spray layer is in a brand-new, unblocked state, the baseline flow field parameters of each nozzle under normal operating conditions are obtained through the same data acquisition and simulation process as steps S1 to S3. Specifically, for each nozzle i, the fluid simulation of step S3 is performed under unblocked conditions to extract the pressure drop across the nozzle under normal operating conditions. Average outlet velocity Coefficient of variation of pressure distribution in the sprinkler coverage area and local flow velocity in the blocked area (Under normal operating conditions with no blockage, this value is set to 0). The above benchmark parameters are stored in the database as a reference baseline for subsequent comparisons of blockage levels. Considering that each nozzle in the spray layer may have individual differences due to factors such as installation location and pipe length, the benchmark parameter library stores them separately according to nozzle number, rather than using a uniform global value;

[0092] Secondly, calculate the congestion impact coefficient. The feature vector of the kth suspected congestion area output in step S3. Based on the baseline parameters of the corresponding nozzle i, a weighted fusion method is used to calculate the clogging influence coefficient. The calculation formula is as follows:

[0093]

[0094] in, , , For the weighting coefficients, satisfying The values ​​of each coefficient range from [0.2, 0.6]. The specific values ​​of the weighting coefficients are dynamically adjusted according to the blockage type: when the blockage shape is close to circular and the area is small, the weighting coefficient is increased. (Flow velocity change weight) to reflect local flow obstruction characteristics; when the blockage is strip-shaped or irregular, increase the weight. (Weight of pressure distribution variation coefficient) to reflect the deterioration of spray coverage uniformity; when the blockage area is large, increase (Pressure drop weight) is used to reflect the increase in overall flow resistance. The dynamic adjustment rules for the weighting coefficients are determined in advance through sensitivity analysis and are fixed in the system. In the above formula, each relative change is normalized with the reference parameter as the denominator to eliminate the dimensional effects caused by individual differences between different nozzles;

[0095] Furthermore, for connected components identified as high-confidence suspected congestion regions in step S2, during the calculation... When introducing a confidence correction factor The value range is [0.8, 1.2], with the high confidence region taking [value]. Low confidence region The corrected congestion impact coefficient is This correction mechanism aims to reduce the impact of false positives caused by image noise or lighting fluctuations on the final judgment result;

[0096] Next, set and compare the clogging detection thresholds. Based on the actual operating requirements of the desulfurization tower spray system, pre-set the clogging detection thresholds. The threshold was determined as follows: 30 nozzle samples with known clogging states were selected (including 10 groups of mild clogging, 10 groups of moderate clogging, and 10 groups of severe clogging), and their clogging influence coefficients were calculated. The optimal threshold, which minimizes the sum of the false positive rate and the false negative rate, was determined using receiver operating characteristic (ROC) curve analysis. After calibration... The value is set to 0.35. This represents the congestion impact coefficient after correction for a certain connected component. At that time, it was determined that the nozzle was clogged, and The higher the value, the more severe the congestion; when If the suspected area is determined to be a false alarm or the degree of blockage is negligible, no alarm will be triggered.

[0097] For cases where multiple connected regions exist within the same nozzle (i.e., multiple points of blockage), the blockage influence coefficient for each connected region is calculated separately. The maximum value is taken as the comprehensive clogging influence coefficient of the nozzle. Meanwhile, if the spatial distance between multiple blockage points is less than 5mm, it is considered a coupled blockage area and needs to be cleaned as a whole during subsequent maintenance;

[0098] Finally, the detection results are output. When a nozzle blockage is detected, the system generates a detection report containing the following information:

[0099] Clogging location: The physical coordinates of the clogging area on the nozzle end face. Combined with the nozzle spatial coordinates stored in step S1, the absolute position of the desulfurization tower in the overall coordinate system is converted, which makes it easier for maintenance personnel to locate the position quickly.

[0100] Blockage geometry: Outputs the minimum bounding rectangle width of the blockage region. With height Outline area Aspect Ratio and roundness And describe the morphological category in text form ("nearly circular blockage", "strip blockage" or "irregular blockage");

[0101] Congestion impact coefficient: Output the corrected congestion impact coefficient and comprehensive congestion impact coefficient At the same time, it gives the degree of blockage according to the preset classification standard. For no blockage or slight blockage For moderate congestion, (For severe congestion);

[0102] Key flow field parameters: Pressure drop across the output nozzle Average outlet velocity and pressure distribution variation coefficient The comparison with the benchmark value is presented in chart form for review by professional technicians.

[0103] Maintenance recommendations: Based on the level of blockage, maintenance recommendations are automatically generated (such as "It is recommended to clean during the next shutdown maintenance" or "It is recommended to schedule cleaning immediately").

[0104] The above test results are stored in a structured data format in a historical database and pushed to the visualization interface of the desulfurization tower monitoring system in real time. At the same time, intermediate results such as real-time images, difference score maps, and simulated flow field cloud maps of this test are also archived to form a complete test traceability chain, which facilitates continuous optimization and calibration of test thresholds and weighting coefficients in the future.

[0105] 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 machine vision-based method for detecting nozzle blockage in the spray layer of a desulfurization tower, characterized in that, The method includes: S1. Image acquisition and preprocessing: Acquire real-time images of the nozzles of the spray layer and pre-stored reference images, and preprocess the real-time images and reference images respectively to obtain preprocessed real-time images and preprocessed reference images with spatial alignment, consistent grayscale and noise suppression. S2. Difference feature extraction and suspected region localization: The preprocessed real-time image and the preprocessed reference image are input into a pre-trained difference detection network to obtain a difference score map that characterizes the degree of pixel-level difference; the difference score map is processed to locate and extract the geometric features and location information of at least one suspected blockage area on the nozzle end face; S3. Verification of clogging status based on fluid simulation: Based on the geometric features and location information of the suspected clogging area, construct a nozzle physical model containing clogging features; perform computational fluid dynamics simulation on the nozzle physical model to obtain simulation feature quantities that reflect the working performance of the nozzle; S4. Quantification of blockage degree and output of detection results: Based on the simulation feature quantity, calculate the blockage influence coefficient used to comprehensively characterize the blockage degree, compare the blockage influence coefficient with a preset threshold, and output the blockage detection result according to the comparison result.

2. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 1, characterized in that, S1 includes: Obtain the pre-stored reference image Real-time images are collected at preset intervals during the operation of the desulfurization tower. ; For the reference image With the real-time image The preprocessed reference image is obtained by sequentially performing size normalization, grayscale conversion, and filtering denoising. With preprocessed real-time images ; During image acquisition, the illuminance of the coaxial light source is adjusted in a closed loop to maintain the overall grayscale average of the image within a preset range.

3. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 1, characterized in that, S2 includes: The preprocessed reference image With preprocessed real-time images After concatenation along the channel dimensions, the data is input into the difference detection network, and a difference score map is obtained through forward propagation. ; For the difference score map Binarization and connected component analysis are performed to extract the geometric features and location information of each connected component. The geometric features and location information include the minimum bounding rectangle, the area of ​​the outline, the perimeter of the outline, the centroid coordinates, and the region shape description. Based on a preset pixel-physical scale mapping coefficient, the geometric features and location information are converted into the actual area, width, and height of the blockage area in terms of physical dimensions.

4. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 1, characterized in that, The difference detection network is trained using a virtual sample training strategy. The training data construction process includes: Select images from publicly available image datasets as background templates; Generate a binary mask image corresponding to the background template, and randomly generate irregular polygonal regions in the mask image as simulated abnormal regions; The background template is copied as the image to be inspected, and the image to be inspected is abnormally synthesized according to the mask image. The synthesis method includes filling the mask area with random RGB colors, or cropping the pixel block corresponding to the mask area from another image and embedding it with Poisson fusion.

5. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 1, characterized in that, S3 includes: Based on the geometric features and location information of the suspected blockage area, a blockage body is constructed on the original geometric model of the nozzle to form the nozzle physical model. ; Set simulation boundary conditions and medium properties for the nozzle physical model. Perform flow field simulation calculations; After the simulation converges, simulation features are extracted, including the pressure drop across the nozzle. Average outlet velocity and the coefficient of variation of pressure distribution at a specified cross section downstream of the nozzle .

6. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 5, characterized in that, The construction rules for the blockage are as follows: When the aspect ratio of the suspected blockage area Between 0.8 and 1.2 and with roundness When the value is greater than 0.7, the blockage body is constructed as a spherical or cylindrical protrusion; When the aspect ratio When the value is greater than 2.0, the blockage body is constructed as a cuboid or elliptical cylinder structure; For other shapes, fit them as irregular polyhedra.

7. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 1, characterized in that, S4 includes: Obtain pre-stored reference flow field parameters of the nozzle under normal operating conditions, including the pressure drop across the nozzle under normal operating conditions. Average outlet velocity and the coefficient of variation of pressure distribution in the sprinkler coverage area ; Based on the relative deviation between the simulated characteristic quantities and the reference flow field parameters, a weighted fusion method is used to calculate the blockage influence coefficient. .

8. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 7, characterized in that, The blockage impact coefficient The calculation formula is: in, , , Let be the weight coefficient, and satisfy... ; The method further includes: dynamically adjusting the weighting coefficient based on the morphological characteristics of the suspected congestion area. , , The value of .

9. The method for detecting nozzle blockage in the spray layer of a desulfurization tower based on machine vision according to claim 1, characterized in that, In step S4, the detection results include the location of the blockage, the geometry of the blockage, the blockage influence coefficient, a comparison of key flow field parameters, and maintenance recommendations. The method further includes: storing the detection results in a structured data format in a historical database and pushing them to the visualization interface of the monitoring system; The real-time images, difference score maps, and simulated flow field cloud maps from this test will be archived together.

Images