A method for positioning and detecting local blockage of an air preheater by fusing infrared images

Abstract

The application discloses a kind of local blockage positioning detection methods of fusion infrared image air preheater, belong to industrial detection and image recognition technical field.Method includes: acquisition infrared video stream and radiation correction get standardization temperature signal;Dynamic reference temperature field is constructed using space-time graph convolution network;Actual and reference temperature are compared and differential to obtain residual characteristics;Residual characteristics are executed clustering and outlier analysis, and spatial mutation boundary signal and local outlier heat source signal are extracted respectively;Two kinds of signal evidence fusion are obtained after feedback correction, and the confidence of blockage is obtained;Grading determination outputs positioning coordinate and severity.The application realizes the accurate positioning of air preheater local blockage under non-steady state condition by infrared image recognition, false alarm rate is low, and interference rejection capability is strong.

Description

Technical Field

[0001] This invention belongs to the field of industrial inspection and image recognition technology, specifically relating to a method for locating and detecting local blockages in an air preheater by fusing infrared images. Background Technology

[0002] As a core heat exchange device in large-scale thermal power boiler systems, the air preheater directly determines the boiler's thermal efficiency and the unit's operational stability by absorbing waste heat from the boiler's flue gas to heat the air needed for combustion. Under actual operating conditions, a large amount of fly ash, incompletely burned carbon particles, and ammonium bisulfate crystals generated from fuel combustion carried in the flue gas easily deposit and adhere to the cold-end heating surface of the air preheater. This accumulation of deposits leads to localized blockages, obstructing fluid channels. The flue gas flow in the blocked area is sharply reduced, hindering the heat exchange process, causing increased ventilation resistance, a drop in heat exchange efficiency, and potentially inducing induced draft fan stall or surge, threatening the unit's output safety.

[0003] Existing methods for detecting blockages in air preheaters suffer from technical drawbacks such as large detection blind spots and low positioning accuracy. Traditional methods based on monitoring the pressure difference between inlet and outlet flue gas can only reflect changes in the overall flow resistance of the air preheater, representing macroscopic overall state monitoring, and cannot detect localized minor blockages or obtain the precise physical coordinates of the blockage. Contact thermocouple temperature measurement methods are limited by sensor deployment density, resulting in low spatial resolution and numerous temperature measurement point blind spots. Conventional infrared thermal imaging detection schemes can acquire continuous surface temperature distribution, but existing data processing logic relies on fixed empirical static threshold segmentation, directly classifying areas exceeding specific temperature or gradient values ​​as blockage areas. Frequent load fluctuations in thermal power generating units and changes in ambient supply air temperature can cause a global and drastic drift in the air preheater's reference temperature field. Under non-steady-state conditions, static fixed thresholds are highly susceptible to inducing large-scale false alarms or missed alarms. Relying solely on image spatial gradient features for judgment cannot distinguish between normal physical temperature abrupt changes caused by thermal bridging effects of internal metal support structures and pathological heat transfer anomalies caused by actual blockages, resulting in weak anti-interference capabilities.

[0004] Some improved solutions introduce deep learning models for dust accumulation recognition, but most adopt an end-to-end single-network architecture, directly classifying or segmenting the input image, resulting in poor interpretability. Furthermore, the model baseline is prone to drift over time, and its sensitivity to slowly evolving early blockages is insufficient. Another solution uses dynamic temperature difference fields for monitoring, but it still relies on calculating the difference between measured temperature and historical averages, failing to establish a spatiotemporal predictive baseline. This limits its ability to cope with drastic fluctuations in operating conditions, and it lacks a mechanism for reverse correction of the baseline based on detection results, leading to a continuous accumulation of baseline deviations after long-term operation. Summary of the Invention

[0005] The purpose of this invention is to provide a method for locating and detecting local blockages in air preheaters by integrating infrared images. This method solves the technical problems of existing infrared detection technologies that rely on static empirical thresholds, resulting in high false alarm rates, weak anti-interference capabilities, and poor positioning accuracy in detecting minor blockages in air preheaters under non-steady-state conditions such as boiler load changes and environmental changes.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for locating and detecting local blockages in an air preheater by fusing infrared images includes the following steps:

[0008] The initial infrared video stream signal of the heated surface of the air preheater is collected, and the initial infrared video stream signal is preprocessed with emissivity dynamic correction to obtain a standardized infrared thermal spatiotemporal signal.

[0009] Based on standardized infrared thermal spatiotemporal signals, a spatiotemporal graph convolutional network is used to perform spatiotemporal graph convolutional mapping processing to extract the evolution features of the time dimension and the correlation features of the spatial dimension, thereby obtaining the dynamic reference signal of the heated surface.

[0010] The actual temperature grid signal is extracted from the standardized infrared thermal spatiotemporal signal. The actual temperature grid signal is compared with the dynamic reference signal of the heated surface and subjected to first-order differential operation in multiple spatial directions to obtain the temperature residual characteristic signal.

[0011] Adaptive density clustering is performed on the temperature residual feature signal. The set of connected high-density anomalous gradient grids is searched by calculating the core distance to obtain the spatial abrupt boundary signal. At the same time, local outlier feature extraction is performed on the temperature residual feature signal. The degree of outlier of the grid relative to the neighborhood is evaluated by comparing the actual distance with the nearest neighbor reachable distance to obtain the local outlier heat source signal.

[0012] The spatial mutation boundary signal is fed back to the spatiotemporal graph convolutional network for cross-correction until the interaction termination condition is met. After the interaction termination condition is met, the spatial mutation boundary signal and the local outlier heat source signal are subjected to evidence chain fusion processing. The orthogonal combination rule is used to eliminate conflicts and contradictions and calculate the joint basic probability allocation value to obtain the blockage situation confidence signal.

[0013] The blockage situation confidence signal is graded and judged according to its amplitude range. When the amplitude exceeds the judgment threshold, the local blockage location coordinate signal and severity signal of the air preheater are output.

[0014] As a preferred technical solution of the present invention, the initial infrared video stream signal is subjected to emissivity dynamic correction preprocessing, specifically including: extracting the reflectivity constant of the heated surface material and the atmospheric transmittance attenuation constant, and constructing the Planck correction function; using the Planck correction function to perform nonlinear inversion on the apparent radiance recorded by the initial infrared video stream signal, and restoring the standardized infrared thermal spatiotemporal signal characterizing the physical absolute temperature.

[0015] As a preferred technical solution of the present invention, a spatiotemporal graph convolutional network is used for spatiotemporal graph convolutional mapping processing, specifically including: dividing the heating surface of the air preheater into discrete temperature grids; constructing a topological undirected graph structure corresponding to the heating surface with the discrete temperature grids as nodes and the physically adjacent heat conduction paths of the grids as edges; and performing forward propagation on the topological undirected graph structure using Chebyshev polynomial approximate graph convolution kernels with self-loop adjacency matrices.

[0016] As a preferred technical solution of the present invention, the actual temperature grid signal is compared with the dynamic reference signal of the heated surface, specifically including: performing node registration according to the same physical coordinate system, and calculating the absolute difference between the actual temperature grid signal value and the predicted value of the dynamic reference signal of the heated surface at each coordinate node.

[0017] As a preferred technical solution of the present invention, the spatial abrupt boundary signal is fed back to the spatiotemporal graph convolutional network for cross-correction, specifically including: extracting the set of grid nodes contained in the spatial abrupt boundary signal, and calculating the average density confidence corresponding to the set of grid nodes; feeding the average density confidence as a penalty weight back to the loss function of the spatiotemporal graph convolutional network in real time for joint correction of model parameters; the interaction termination condition is that the value of the first derivative of the interactive correction loss function output in two consecutive calculation cycles is lower than a preset convergence threshold.

[0018] As a preferred technical solution of the present invention, the interactive correction loss function on which the joint correction of model parameters depends is constructed according to the following rules: The interactive correction loss function includes a basic mean square error loss function and an additional feedback penalty term; The feedback penalty term is obtained by calculating the squared difference between the predicted baseline temperature and the historical unblocked real physical temperature within the grid node set, and multiplying it by the global penalty hyperparameter and the average density confidence value.

[0019] As a preferred embodiment of the present invention, the global penalty hyperparameter is a dynamically adjustable parameter. Obtaining the global penalty hyperparameter involves the following logic: extracting the actual temperature grid signal of the grid node set within a set historical time window to generate an actual temperature time series characterizing the true temperature evolution; extracting the dynamic reference signal of the heated surface of the grid node set within the set historical time window to generate a reference temperature time series characterizing the theoretical temperature evolution; calculating the temporal temperature gradient decay rate of the actual temperature time series relative to the reference temperature time series, and converting the temporal temperature gradient decay rate into the global penalty hyperparameter using a nonlinear mapping function. The calculation model employed includes:

[0020] ;

[0021] In the formula, The global penalty hyperparameter; The base penalty coefficient is set; It is the hyperbolic tangent activation function; This is the mapping steepness adjustment coefficient; The total number of grid nodes contained in the set of grid nodes; The total number of discrete time steps included in the set historical time window; The set of grid nodes; For the first node in the set of grid nodes The grid node at the ... The actual temperature grid signal values ​​corresponding to each discrete time step; For the first node in the set of grid nodes The grid node at the ... The dynamic reference signal values ​​of the heated surface corresponding to each discrete time step; The partial derivative of the actual temperature grid signal value in the time dimension; The partial derivative of the dynamic reference signal value of the heated surface in the time dimension; This is the preset thermal inertia tolerance threshold.

[0022] As a preferred technical solution of the present invention, the core distance is calculated by searching for a set of high-density abnormal gradient grids. When calculating the core distance, the following logic is executed: The difference in the physical space span of the horizontal and vertical coordinates of the grid nodes in the heated surface plane coordinate system is comprehensively considered, as well as the difference in the numerical space span of the temperature residual characteristic value adjusted by introducing the temperature difference weighting coefficient.

[0023] As a preferred technical solution of the present invention, the evaluation of the outlier degree of the grid relative to its neighborhood specifically includes: extracting the local reachability density value of the evaluation node; calculating the ratio of the sum of the reachability density values ​​of all neighbors in the nearest neighborhood of the evaluation node to the corresponding reachability density value of the evaluation node, and generating a local anomaly factor score characterizing independent thermal anomalies.

[0024] As a preferred technical solution of the present invention, calculating the joint basic probability allocation value includes: extracting a first basic probability value for the allocation of congestion events based on the results of adaptive density clustering; extracting a second basic probability value for the allocation of congestion events based on the local anomaly factor score; calculating the orthogonal product of the first basic probability value and the second basic probability value, and dividing by the normalized denominator after conflict coefficient correction.

[0025] As a preferred embodiment of the present invention, the following logic is executed when calculating the conflict coefficient: First probability values ​​of non-blocking events determined by the density clustering algorithm are extracted and cross-producted with second probability values ​​of blocking events determined by the local anomaly factor algorithm; Third probability values ​​of blocking events determined by the density clustering algorithm are extracted and cross-producted with fourth probability values ​​of non-blocking events determined by the local anomaly factor algorithm; The results of the two cross-product operations are added together to quantify the inconsistency of spatially overlapping regions.

[0026] As a preferred technical solution of the present invention, the classification judgment is performed based on the amplitude range of the congestion situation confidence signal, specifically including: establishing a mapping lookup table from continuous confidence space to discrete severity interval, performing matrix inverse transformation on the coordinates of grid nodes exceeding the judgment threshold, and outputting two-dimensional geometric positioning parameters in Cartesian coordinate system.

[0027] This invention also discloses a local blockage detection system for air preheaters based on infrared images, comprising: a video stream processing module configured to acquire initial infrared video stream signals of the heated surface of the air preheater, and perform emissivity dynamic correction preprocessing on the initial infrared video stream signals to obtain a standardized infrared thermal spatiotemporal signal; a dynamic mapping module configured to perform spatiotemporal graph convolutional mapping processing based on the standardized infrared thermal spatiotemporal signal using a spatiotemporal graph convolutional network, extracting temporal dimension evolution features and spatial dimension correlation features to obtain a dynamic reference signal for the heated surface; and a feature comparison module configured to extract actual temperature grid signals from the standardized infrared thermal spatiotemporal signal, compare the actual temperature grid signals with the dynamic reference signal for the heated surface, and perform spatial multi-directional first-order differential operations to obtain a temperature residual feature signal. The dual-line extraction module is configured to perform adaptive density clustering on the temperature residual feature signal, searching for a set of connected high-density anomalous gradient grids by calculating core distance to obtain the spatial abrupt boundary signal; simultaneously, it performs local outlier feature extraction on the temperature residual feature signal, evaluating the degree of outlier relative to the neighborhood by comparing the actual distance with the nearest reachable distance to obtain the local outlier heat source signal; the evidence fusion module is configured to feed the spatial abrupt boundary signal back to the spatiotemporal graph convolutional network for cross-correction until the interaction termination condition is met. After the interaction termination condition is met, the spatial abrupt boundary signal and the local outlier heat source signal are subjected to evidence chain fusion processing, using the orthogonal combination rule to eliminate conflicts and contradictions and calculate the joint basic probability allocation value to obtain the blockage status confidence signal; the status output module is configured to make a graded judgment based on the amplitude range of the blockage status confidence signal, and output the local blockage location coordinate signal and severity signal of the air preheater when the amplitude exceeds the judgment threshold.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This invention constructs a four-layer collaborative detection architecture consisting of dynamic benchmark prediction, dual-branch anomaly extraction, closed-loop feedback correction, and evidence fusion, which differs from existing detection schemes that rely on static thresholds or single models. Through emissivity dynamic correction preprocessing, a standardized infrared thermal spatiotemporal signal characterizing the physical absolute temperature is reconstructed from the initial infrared video stream signal, eliminating temperature measurement system errors caused by flue gas attenuation and material reflection, providing a reliable data foundation for subsequent analysis. Based on this, a spatiotemporal graph convolutional network is used for spatiotemporal graph convolutional mapping processing to extract temporal evolution features and spatial correlation features, generating a dynamic benchmark signal for the heated surface that dynamically changes with operating conditions. This effectively eliminates global thermal field drift interference caused by unit load fluctuations and ambient temperature changes, solving the problem of high false alarm rates under unsteady conditions. By comparing the actual temperature grid signal with the dynamic benchmark signal of the heated surface and performing first-order differential operations in multiple spatial directions, a temperature residual feature signal is obtained, achieving accurate separation of temperature anomalies from normal operating conditions in the blocked area.

[0030] Adaptive density clustering is performed on the temperature residual feature signal to extract the spatial abrupt boundary signal. Simultaneously, local outlier feature extraction is performed to obtain local outlier heat source signals. Two independent computational branches cover two typical morphologies: continuous sheet-like ash accumulation and point-like crystalline blockage, respectively. This eliminates the blind spots of single algorithms when dealing with false features or subtle pathological features of structural thermal bridges, improving the detection capability of small blockages. A closed-loop mechanism of feeding the spatial abrupt boundary signal back to the spatiotemporal graph convolutional network for cross-correction avoids the assimilation of long-period, slow blockage evolution by the prediction network, keeping the blockage component in the temperature residual feature signal pure and improving the positioning accuracy and robustness of the judgment. Evidence chain fusion processing is performed on the spatial abrupt boundary signal and the local outlier heat source signal. Orthogonal combination rules are used to eliminate conflicts and contradictions and calculate the joint basic probability allocation value to obtain the blockage status confidence signal. This effectively suppresses the false alarm probability of metal-supported thermal bridge areas and enhances the confidence response of point-like blockages. The blockage situation confidence signal is graded and judged according to its amplitude range, and the local blockage location coordinate signal and severity signal of the air preheater are output, providing a refined basis for operation and maintenance decision-making.

[0031] This invention further introduces a dynamically adjustable global penalty hyperparameter. By calculating the time-series temperature gradient decay rate and using a nonlinear mapping function to adjust the penalty intensity in real time, it avoids excessive or insufficient feedback correction under extreme unsteady conditions such as deep peak shaving or rapid load changes. This ensures that the predicted trajectory of the dynamic reference signal is always anchored in a non-blocking state, enhancing the system's adaptive anti-interference capability and operational stability in complex industrial scenarios. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the method of the present invention.

[0034] Figure 2 This is a flowchart of the closed-loop feedback correction sub-process of the present invention.

[0035] Figure 3 This is a flowchart of the adaptive density clustering processing branch of the present invention.

[0036] Figure 4 This is a flowchart of the local outlier feature extraction processing branch of the present invention. Detailed Implementation

[0037] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0038] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0039] The overall process of this invention is divided into six core steps: data preprocessing, dynamic benchmark prediction, residual feature generation, dual-branch anomaly extraction, closed-loop feedback correction, and evidence fusion judgment.

[0040] The initial infrared video stream signal is converted into a real temperature spatiotemporal sequence after radiometric correction and then input into a spatiotemporal graph convolutional network to generate a theoretical temperature benchmark that dynamically changes with operating conditions. The difference between the measured temperature and the benchmark temperature is then subtracted and spatially differentiated to obtain residual features. These residual features are used to extract different types of anomalous features through two independent branches: density clustering and outlier factor. The clustering results are fed back as penalty weights to the graph convolutional network to correct the benchmark model, forming a closed-loop iteration until convergence. The two anomalous features after convergence are fused using evidence theory to obtain the final blockage confidence score, and the location coordinates and severity level are output. The overall architecture forms a closed-loop link of benchmark prediction, anomaly detection, and benchmark correction, which combines detection accuracy with adaptability to operating conditions.

[0041] Furthermore, the acquired initial infrared video stream signal first enters the emissivity dynamic correction preprocessing stage. The initial infrared video stream signal consists of consecutive frames of infrared thermal images, each frame containing pixel coordinates and corresponding radiance values, and is continuously acquired at fixed sampling intervals in the time dimension.

[0042] The high dust concentration in the flue gas inside the air preheater causes infrared radiation to be absorbed and scattered by the flue gas particles during transmission, resulting in an apparent radiation brightness reaching the camera that is lower than the actual radiation brightness of the heated surface. Simultaneously, the varying degrees of oxidation on the heated surface metal leads to spatial differences in material reflectivity, further affecting the accuracy of radiation temperature measurement. To eliminate these interferences, the video stream processing module performs dynamic emissivity correction preprocessing. The preprocessing first extracts the reflectivity constant of the heated surface material and the atmospheric transmittance attenuation constant. The reflectivity constant of the heated surface material is determined based on the type of substrate of the heat storage element; carbon steel substrates and enamel substrates correspond to different emissivity values. For example, the emissivity of enamel coating can be taken as 0.92, while the emissivity of carbon steel substrates can be taken as 0.85. The atmospheric transmittance attenuation constant is calculated based on the dust concentration in the flue gas and the transmission path length. The dust concentration can be obtained from the unit's operating parameters, and the attenuation coefficient can be determined through on-site calibration.

[0043] Based on the above parameters, a Planck correction function is constructed to perform nonlinear inversion on the apparent radiance recorded in the initial infrared video stream signal. The Planck correction function is based on the blackbody radiation law, converting the apparent radiance values ​​into corresponding physical absolute temperature values. The inversion process is performed pixel-by-pixel and frame-by-frame, ultimately outputting a standardized infrared thermal spatiotemporal signal stripped of smoke attenuation and material reflection interference. The standardized infrared thermal spatiotemporal signal is stored in the form of a three-dimensional tensor, with the three dimensions corresponding to the horizontal physical coordinates, vertical physical coordinates, and time series of the heated surface, respectively. Each tensor element represents the actual physical temperature value at the corresponding location and time, in °C. The standardized infrared thermal spatiotemporal signal is transmitted to both the dynamic mapping module for baseline prediction and the feature comparison module for residual calculation.

[0044] The dynamic mapping module receives standardized infrared thermal spatiotemporal signals, performs spatiotemporal graph convolutional mapping processing using a spatiotemporal graph convolutional network, and outputs a dynamic reference signal for the heated surface. This dynamic reference signal characterizes the theoretically expected temperature field under current operating conditions without blockages, serving as a reference baseline for subsequent anomaly detection. The dynamic mapping module first divides the two-dimensional physical cross-section of the air preheater's heated surface into discrete temperature grids. The grid edge length is determined based on the spatial resolution of the infrared camera, ensuring that each grid corresponds to at least one infrared pixel, balancing computational efficiency and spatial accuracy.

[0045] Using discrete temperature grids as nodes and physically adjacent heat conduction paths as edges, a topological undirected graph structure is constructed for the corresponding heated surfaces. Physical adjacency includes lateral and vertical adjacency. Each internal grid node connects to four adjacent nodes, and boundary grid nodes connect to the corresponding number of adjacent nodes. The topological graph structure accurately reflects the heat conduction relationships of the heated surfaces, with heat transfer between adjacent grids corresponding to the edge connections in the graph structure.

[0046] A forward propagation is performed on a topological undirected graph structure using Chebyshev polynomial approximation graph convolution kernels with self-loop adjacency matrices. The graph convolutional network simultaneously processes multiple frames of temperature data within a continuous historical time window, extracting temperature evolution patterns in the temporal dimension and temperature correlation features in the spatial dimension. The temporal dimension captures the global temperature drift trend caused by load changes and ambient temperature fluctuations, while the spatial dimension captures the temperature coupling relationship between adjacent grids caused by heat conduction paths.

[0047] Chebyshev polynomial approximation reduces the computational complexity of spectral graph convolution and improves forward propagation speed. Self-loop adjacency matrices ensure that node features participate in aggregation calculations, avoiding information loss. The number of network layers and the polynomial order are set according to the size of the heated surface and the complexity of the temperature field; for example, it can be set to three convolutional layers with a third-order Chebyshev polynomial to ensure sufficient extraction of spatiotemporal correlation features.

[0048] Furthermore, through multi-layer graph convolution forward extrapolation, the dynamic mapping module outputs a theoretically expected temperature field adapted to the current load conditions, i.e., the dynamic reference signal for the heated surface. The dimensions and coordinate system of the dynamic reference signal for the heated surface are completely consistent with the standardized infrared thermal spatiotemporal signal, with each node corresponding to a reference temperature prediction value. When the heated surface experiences global temperature drift due to random load increases or decreases and changes in ambient temperature, the dynamic reference signal changes synchronously, always reflecting the normal temperature distribution under the current operating conditions.

[0049] The feature comparison module receives standardized infrared thermal spatiotemporal signals and dynamic reference signals from the heated surface, performs comparison processing and first-order differential operations in multiple spatial directions, and outputs temperature residual feature signals. The feature comparison module first extracts the actual temperature grid signal from the standardized infrared thermal spatiotemporal signal. The actual temperature grid signal represents the measured temperature distribution at the current moment and uses the same physical coordinate system and grid division rules as the dynamic reference signal from the heated surface.

[0050] Node registration is performed using the same physical coordinate system. The absolute difference between the actual temperature grid signal value and the predicted value of the dynamic reference signal of the heated surface at each coordinate node is calculated to generate an absolute residual matrix. The absolute difference calculation is performed node by node to obtain the deviation amplitude between the measured temperature and the reference temperature for each grid. The deviation amplitude is smaller in the normal region and larger in the blocked region. First-order differential operations are performed on the absolute residual matrix using multi-directional operators to obtain the spatial gradient amplitude in the two-dimensional plane. The spatial first-order differential operation extracts the steep boundaries of abnormal surface heat accumulation caused by fluid obstruction, eliminates the global gradual change trend, and highlights the edge characteristics of the blocked region.

[0051] Differentiation operations are performed simultaneously in both the horizontal and vertical directions, for example, using the Sobel operator to synthesize the final gradient magnitude matrix. After comparison and differentiation processing, the feature comparison module outputs a temperature residual feature signal. This temperature residual feature signal contains both absolute deviation information and gradient edge information, reflecting both the severity of the blockage and the boundary morphology. The temperature residual feature signal is transmitted to the dual-line extraction module as a common input for dual-branch anomaly detection.

[0052] The dual-line extraction module receives the temperature residual feature signal and sets up two parallel, independent computational branches to extract different types of blockage anomaly features. The first computational branch performs adaptive density clustering processing to extract the boundary features of continuous sheet-like blockages; the second computational branch performs local outlier feature extraction processing to extract the anomaly features of isolated point-like blockages. The two branches operate in parallel without interference, outputting two complementary types of anomaly signals.

[0053] The first computational branch performs adaptive density-based clustering. This branch targets continuously distributed sheet-like ash blockages, which manifest as contiguous temperature anomaly regions with continuous gradient changes and strong spatial clustering. Clustering is performed in a composite feature space, where each grid node's feature vector contains three dimensions: x-coordinate, y-coordinate, and temperature residual value. The core distance calculation considers both the physical spatial span difference of the grid node's x and y coordinates in the heated surface plane coordinate system and the numerical spatial span difference of the temperature residual feature value, adjusted by a temperature difference weighting coefficient. The temperature difference weighting coefficient balances the dimensional differences between physical and numerical spatial measures, ensuring that the contributions of both dimensions to the clustering results match their physical meaning. The formula for calculating the core distance is:

[0054] ;

[0055] in, This is the core distance metric between grid node v and grid node u in the composite feature space; Let v be the physical scale of the abscissa of the grid node v in the plane coordinate system of the heated surface; The physical scale for the ordinate of grid node v; The horizontal coordinate of grid node u is the physical scale. The physical scale for the ordinate of grid node u; This is a weighting coefficient used to balance the temperature difference between physical and numerical spatial measurement dimensions. This represents the temperature residual characteristic signal value corresponding to grid node v; This represents the temperature residual characteristic signal value corresponding to grid node u. After the core distance is calculated, all grid nodes on the heated surface are traversed to search for a connected set of high-density anomalous gradient grids. The density threshold is adaptively determined based on the residual distribution of the entire image, without the need for manual preset of a fixed value. Nodes that meet the density reachability condition form a cluster, and each cluster corresponds to a continuous blockage region. The outer contour boundary of each cluster is extracted, and the spatial abrupt boundary signal is output. The spatial abrupt boundary signal includes the boundary coordinates of each blockage region, the set of internal nodes, and the average density confidence value.

[0056] The second computational branch performs local outlier feature extraction. This branch targets isolated, point-like crystalline blockages, which manifest as scattered temperature anomalies with discrete spatial distribution, drastic gradient changes, and are easily missed by continuous detection algorithms. When evaluating the outlier degree of a node relative to its nearest neighbors, the local reachability density of the evaluation node is first extracted. Reachability distance is defined as the larger of the actual distance from the evaluation node to its neighboring nodes and the k-th nearest neighbor distance of the neighboring node itself, reflecting the ease of reaching the evaluation node from its neighboring nodes. Local reachability density is the reciprocal of the average reachability distance of all the evaluation node's neighbors; the lower the density, the more the node deviates from the surrounding normal area. The ratio of the sum of the reachability densities of all neighbors in the evaluation node's nearest neighborhood to the corresponding reachability density of the evaluation node is calculated to generate a local anomaly factor score characterizing independent heating anomalies. A score close to 1 indicates that the node's local density is similar to its neighbors and belongs to the normal area; a score significantly greater than 1 indicates that the node's local density is much lower than its neighbors and belongs to an anomalous outlier. The anomaly factor score is calculated for all grid nodes, and nodes with scores exceeding the judgment threshold are filtered out, outputting the local outlier heat source signal. The local outlier heat source signal includes the coordinates of all isolated anomalies and their corresponding anomaly factor scores.

[0057] The evidence fusion module receives spatially abrupt boundary signals and local outlier heat source signals. It first performs a closed-loop feedback correction operation, followed by evidence chain fusion processing. The purpose of feedback correction is to prevent slowly evolving blockages from being learned and incorporated into the baseline prediction by the spatiotemporal graph convolutional network, causing the residual signal to be canceled out. Slowly formed blockages exhibit gradual temperature changes. If the baseline model continuously learns this change, it will gradually identify abnormal temperatures as normal temperatures, eventually causing the residual signal to approach zero, resulting in missed detections. The feedback correction process first extracts the set of grid nodes contained within the spatially abrupt boundary signal and calculates the average density confidence score corresponding to the set of grid nodes. The average density confidence score reflects the probability that the corresponding area is a true blockage; a higher confidence score indicates a greater likelihood of blockage. The average density confidence score is used as a penalty weight and fed back in real-time to the loss function of the spatiotemporal graph convolutional network in the dynamic mapping module for joint correction of model parameters. The interactive correction loss function includes a basic mean squared error loss function and an additional feedback penalty term. The basic mean squared error loss function ensures the prediction accuracy of normal areas, while the feedback penalty term suppresses the model's temperature fitting of suspected blockage areas, preventing the assimilation of abnormal features. The feedback penalty term is obtained by calculating the squared difference between the predicted baseline temperature and the historical unblocked real physical temperature within the grid node set, and multiplying it by the global penalty hyperparameter and the average density confidence value. The global penalty hyperparameter is used to adjust the feedback strength; an excessively large value can lead to baseline oscillation, while an excessively small value results in insufficient suppression. The average density confidence value serves as a dynamic weight; a higher confidence value results in a stronger penalty, and a lower confidence value results in a weaker penalty, achieving adaptive adjustment of the feedback strength. The formula for calculating the interactive correction loss function is as follows:

[0058] ;

[0059] in, This is the interactive correction loss function after feedback correction based on adaptive density clustering results; The basic mean squared error loss function; The global penalty hyperparameter for adjusting the feedback strength; The set of grid nodes covered by the spatial abrupt boundary signal; The predicted baseline temperature is the output of the i-th node network within the grid node set; The historical unblocked real physical temperature corresponding to the i-th node; For a set of grid nodes The average density confidence value of all grid nodes within the range.

[0060] The interaction termination condition is set when the first derivative of the interactively corrected loss function output over two consecutive computation cycles is lower than a preset convergence threshold. The first derivative of the loss function reflects the rate of change of the loss; when the derivative approaches zero, it indicates that the loss is stabilizing and the model parameters are no longer changing significantly. Meeting this condition for two consecutive cycles avoids false convergence caused by a single fluctuation.

[0061] To avoid over- or under-correction of feedback during unsteady operating conditions such as deep peak shaving or rapid load changes, the global penalty hyperparameter further employs a dynamic adjustment mechanism. When acquiring the dynamically adjusted global penalty hyperparameter, the evidence fusion module extracts the actual temperature grid signal of the grid node set within a set historical time window, and reassembles the grid signals from different time steps into a three-dimensional spatiotemporal tensor in chronological order, forming an actual temperature time series characterizing the actual temperature evolution. Simultaneously, the dynamic reference signal of the heated surface, output in real-time by the dynamic mapping module within the set historical time window, is extracted from the grid node set, and a reference temperature time series characterizing the theoretical temperature evolution is generated using the same tensor reconstruction logic. The actual temperature time series and the reference temperature time series are strictly aligned on the time axis to eliminate time delay errors caused by system sampling jitter. Next, the temporal temperature gradient decay rate of the actual temperature time series relative to the reference temperature time series is calculated. The temporal temperature gradient decay rate characterizes the degree of hysteresis between the physical actual temperature and the theoretically expected temperature of the model within the grid node set; the degree of hysteresis is positively correlated with the blockage thickness in that region. The calculation model used to calculate the time-series temperature gradient decay rate and convert it into the global penalty hyperparameter includes:

[0062] ;

[0063] In the formula, The global penalty hyperparameters are actually substituted into the interactive correction loss function after dynamic adjustment; The base penalty coefficient is set; It is the hyperbolic tangent activation function; This is the mapping steepness adjustment coefficient; The total number of grid nodes contained in the set of grid nodes; The total number of discrete time steps included in the set historical time window; The set of grid nodes; For the first node in the set of grid nodes The grid node at the ... The actual temperature grid signal values ​​corresponding to each discrete time step; For the first node in the set of grid nodes The grid node at the ... The dynamic reference signal values ​​of the heated surface corresponding to each discrete time step; The partial derivative of the actual temperature grid signal value in the time dimension; The partial derivative of the dynamic reference signal value of the heated surface in the time dimension; This is a preset thermal inertia tolerance threshold. In the microprocessor's low-level computation, the partial derivative in the time dimension is approximated by discretization using backward difference, i.e. It is approximately equal to the difference between the current time step value and the previous time step value divided by the physical time interval between adjacent time steps.

[0064] The partial derivatives of the dynamic reference signal values ​​for the heated surface are calculated using the same backward differential logic. After calculating the partial derivatives of the actual temperature and the reference temperature at each moment, the absolute value of the difference between the two is obtained. A double summation operation is performed on all nodes within the grid node set and all time steps within the set historical time window, and the result is divided by the product of the total number of nodes and the total number of time steps to obtain the mean. The mean is then subtracted from the preset thermal inertia tolerance threshold. The thermal inertia tolerance threshold filters out the minute thermal inertia fluctuations inherent in normal metallic materials and the thermal noise from the infrared sensor. The net attenuation rate after threshold removal is multiplied by the mapping steepness adjustment coefficient. Input into the hyperbolic tangent activation function The hyperbolic tangent activation function smoothly maps the input value to... to Within the convergence interval, the penalty coefficient is limited to prevent exponential divergence under extreme conditions, ensuring gradient stability of the spatiotemporal graph convolutional network during backpropagation. Mapping value addition The base penalty coefficient for subsequent multiplication is set. In the process, the final global penalty hyperparameters are output. .

[0065] By introducing this dynamic adjustment model, during steady-state operation of the unit, the partial derivatives of both the actual temperature and the reference temperature approach zero, and the net decay rate is negative. After hyperbolic tangent mapping, this becomes a negative decimal. At this point, the global penalty hyperparameter is smaller than the basic penalty coefficient, preventing the model from overfitting to weak anomalies. When the unit operates under variable load and the blocked area is exposed to extreme thermal inertia, the absolute value of the difference in partial derivatives rises sharply, and the hyperbolic tangent mapping value rapidly approaches... The global penalty hyperparameter is amplified to nearly twice the base penalty coefficient. This amplified penalty weight dominates the interactive correction loss function, severely hindering the spatiotemporal graph convolutional network's blind learning of the current spurious benchmark and forcibly anchoring the dynamic benchmark signal to the theoretically predicted trajectory in a non-blocking state.

[0066] After the interaction termination condition is met, the weight parameters are frozen, and the dual-line extraction module outputs the steady-state spatial abrupt boundary signal and the local outlier heat source signal. At this point, the baseline model has been corrected, the blockage component in the residual features reaches its maximum, and the detection sensitivity is optimal.

[0067] After convergence, evidence chain fusion processing is performed. Spatial abrupt boundary signals and local outlier heat source signals reflect blockage characteristics from different dimensions. While they are complementary, they also exhibit local conflicts, and direct superposition can easily lead to misjudgment. The orthogonal combination rule of evidence theory is used for fusion, which can effectively resolve conflicts and improve the reliability of the judgment. First, basic probability values ​​are assigned to the two signals. The first basic probability value for the blockage event is extracted based on the adaptive density clustering results, and the second basic probability value is extracted based on the local anomaly factor score. The probability values ​​are obtained by mapping the output amplitude of the corresponding algorithm; the higher the amplitude, the greater the blockage probability. Simultaneously, a probability value and uncertainty for non-blockage events are assigned to each signal. The sum of the blockage probability and the non-blockage probability is less than or equal to 1, and the remainder is the uncertainty. The orthogonal combination result of the two signals is calculated to obtain the joint basic probability allocation value. The formula for calculating the orthogonal combination is:

[0068] ;

[0069] in, The amplitude of the congestion situation confidence signal output after fusion processing; This represents the first basic probability value assigned to the congestion event by the density clustering algorithm. This is the second basic probability value assigned to the blocking event by the local anomaly factor algorithm; The conflict coefficient is used to characterize the degree of disagreement in the algorithm's decision-making process.

[0070] The collision coefficient is used to quantify the degree of inconsistency between two signals, and the calculation logic is as follows:

[0071] The first probability value of a non-blocking event determined by the density clustering algorithm is extracted and cross-producted with the second probability value of a blocking event determined by the local anomaly factor algorithm. The third probability value of a blocking event determined by the density clustering algorithm is extracted and cross-producted with the fourth probability value of a non-blocking event determined by the local anomaly factor algorithm. The results of the two cross-product operations are added together to obtain the conflict coefficient. A larger conflict coefficient indicates a more severe divergence between the two signals, and a stronger correction effect of the normalization denominator. The fusion mechanism outputs high confidence for regions where both signals are consistently classified as blocking, medium confidence for regions where both signals are classified as conflicting, and low confidence for regions where both signals are consistently classified as normal.

[0072] Metal-supported thermal bridge regions typically exhibit boundary characteristics only in density clustering branches, showing no anomalous response in outlier branches. After fusion, the confidence level is suppressed, thus reducing the false alarm probability. Tiny point-like blockages show strong responses in outlier branches but may exhibit weak responses in density clustering branches. After fusion, the confidence level is enhanced, improving the detection rate. The evidence fusion module ultimately outputs a blockage status confidence signal. This signal is a two-dimensional matrix with the same dimension as the temperature grid, where each node corresponds to a confidence value between 0 and 1. Higher values ​​indicate a greater likelihood of blockage at the corresponding node.

[0073] The status output module receives the congestion status confidence signal, performs hierarchical judgment and coordinate transformation, and outputs the final detection result.

[0074] The status output module establishes a lookup table mapping continuous confidence space to discrete severity intervals. For example, a confidence level in the [0.6, 0.8) interval is considered initial minor blockage, corresponding to a low-level warning; a confidence level in the [0.8, 0.95) interval is considered intermediate moderate blockage, corresponding to a medium-level warning; and a confidence level in the [0.95, 1.0] interval is considered severe critical blockage, corresponding to a high-level warning. The interval division threshold is determined based on field operation experience and maintenance strategies and can be adjusted according to actual needs. For grid nodes that meet the alarm conditions, the status output module performs an inverse coordinate matrix transformation. The grid node coordinates are discrete indices in the image pixel coordinate system. Through an inverse transformation using a pre-calibrated perspective transformation matrix, the pixel coordinates are restored to two-dimensional geometric positioning parameters in the Cartesian coordinate system of the air preheater heating surface. The positioning parameters include the center coordinates, boundary contour, and area size of the blockage area, in meters (m).

[0075] The status output module ultimately outputs the location coordinates and severity signals for localized blockages in the air preheater. The location coordinates contain the geometric position information of all blocked areas, and the severity signal corresponds to the blockage level of each area. The output signals can be transmitted to the unit's distributed control system via the industrial bus for operators to view and make decisions, or they can be directly linked to the soot blowing control system to perform targeted soot blowing operations.

[0076] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to specific implementation examples.

[0077] Example 1: This example uses a three-compartment rotary air preheater of a 600MW supercritical thermal power generating unit as the application platform to fully illustrate the deployment method and operation process of the detection system. The hardware deployment architecture includes an infrared sensor array and edge computing nodes. The infrared sensor array is an uncooled long-wave infrared focal plane camera deployed on the flue gas outlet side of the cold end of the air preheater, with a working wavelength set to 8 to 14μm and a spatial resolution of 640×512 pixel matrix. To resist dust erosion inside the air preheater, the uncooled long-wave infrared focal plane camera is equipped with a nitrogen purging protective sleeve with a streamlined air guide. The camera output is connected to the edge computing server via armored optical fiber. The edge computing server internally runs a video stream processing module, a dynamic mapping module, a feature comparison module, a dual-line extraction module, an evidence fusion module, and a status output module. The camera acquisition frame rate is set to 2 frames per second, and the temporal resolution meets the monitoring requirements for slow blockage evolution. A single frame image covers a fan-shaped area of ​​the heated surface, achieving full-area scanning in conjunction with rotor rotation.

[0078] The camera's field of view and installation position are calibrated to ensure accurate mapping between image coordinates and physical coordinates. The video stream processing module receives the initial infrared video stream signal output from the camera and performs dynamic emissivity correction. The heat storage element on the heated surface is made of enamel-coated material with an emissivity set to 0.92. The flue gas transmittance is calculated in real time based on the dust content of the unit's flue gas, which is obtained from the unit's SIS system, with an attenuation coefficient of 0.015. The temperature measurement error of the standardized infrared thermal spatiotemporal signal output after correction is controlled within ±2℃, meeting the detection accuracy requirements.

[0079] The dynamic mapping module constructs a spatiotemporal graph convolutional network, dividing the heated surface into a discrete grid with a side length of 0.1m, resulting in approximately 12,000 nodes. The topology graph structure is constructed using the four-adjacency rule. The graph convolutional network has three convolutional layers, and the Chebyshev polynomial order is set to third. The input time window length is set to 30 frames, corresponding to 15 seconds of historical data. The network outputs a dynamic reference signal for the heated surface, which can update the reference temperature field in real time according to changes in unit load. During the process of the unit load ramping up from 50% to 100%, the reference signal following delay does not exceed two calculation cycles.

[0080] During the model pre-training phase, three consecutive months of normal operation data were extracted to construct a training dataset. Normal operation data was defined as data under conditions of stable unit operation, no blockage alarms, and operation for more than 4 hours after soot blowing. The dataset included temperature field samples under different loads, ambient temperatures, and coal types, covering the entire operating range. For missing frames, a cubic spline interpolation algorithm was used for smooth temporal reconstruction, and Z-score standardization was performed on the temperature data.

[0081] The model training employs supervised learning, using historical frame temperature data as input and the current frame's actual temperature as the label. The optimizer uses an adaptive moment estimation algorithm with an initial learning rate of 0.001 and a training batch size of 32. An early stopping strategy is employed to prevent overfitting. A Bayesian optimization algorithm based on Gaussian process regression is introduced to jointly optimize the number of network layers, Chebyshev polynomial order, and learning rate, resulting in the optimal combination of a three-layer network, a third-order polynomial, and an initial learning rate of 0.0012. The trained model weights are permanently stored on edge computing nodes. During online operation, only forward inference is performed, with incremental fine-tuning performed once daily during off-peak load periods using the day's normal operating data.

[0082] The feature comparison module performs node registration and difference calculation to generate an absolute residual matrix. The Sobel operator is used to perform first-order differentiation in both the horizontal and vertical directions to synthesize the gradient magnitude matrix. The temperature residual feature signal simultaneously contains both absolute deviation and gradient information, improving the signal-to-noise ratio by approximately 20% compared to the original temperature difference.

[0083] The dual-line extraction module performs two types of anomaly detection in parallel. The temperature difference weighting coefficient for the density clustering branch is set to 0.05 / ℃ to balance the dimensions of spatial distance and temperature difference. The adaptive density threshold is taken as the upper quartile of the core distance across the entire image. The nearest neighbor count k for the outlier factor branch is set to 20, and the anomaly determination threshold is set to 1.5. The two branches output spatial abrupt boundary signals and local outlier heat source signals, respectively. The evidence fusion module performs closed-loop feedback correction. (Base penalty coefficient...) The global penalty hyperparameter is set to 0.80 during initial runtime and updated in real-time according to a dynamic adjustment mechanism. The convergence threshold is set to 0.001; convergence is determined when the condition is met for two consecutive cycles. A single feedback iteration takes approximately 50ms, which can be corrected within the interval between two sampling cycles.

[0084] After convergence, evidence fusion is performed. The basic probability allocation of the two signals is mapped using the sigmoid function, normalizing the original amplitude to the range of 0 to 1. The conflict coefficient is calculated in real time, maintaining below 0.1 under normal operating conditions and not exceeding 0.4 under abnormal operating conditions. The false alarm rate of the congestion status confidence signal after fusion is reduced by more than 40% compared to the single path. The status output module sets three-level judgment thresholds: 0.6 for minor congestion, 0.8 for moderate congestion, and 0.95 for severe congestion. The inverse coordinate transformation uses a pre-calibrated homography matrix, with a positioning error not exceeding 0.05m. The output signal is transmitted to the unit's DCS system via the Modbus TCP protocol, with a refresh cycle of 1 second. This embodiment operates under three typical scenarios: full-load stable operation, variable load ramping operation, and sudden change in ambient temperature operation. The detection rate of minor congestion reaches over 95%, the false alarm rate is controlled within 3%, and the positioning accuracy meets the needs of on-site operation and maintenance.

[0085] Example 2: This example is a further optimization based on Example 1. In this example, the technical defect of the fixed value of the global penalty hyperparameter causing over-correction or under-correction of the model feedback is solved when the thermal power generating unit performs deep peak shaving or rapid load change operation.

[0086] In actual industrial fluid heat transfer environments, the ash accumulation and blockage process on the heating surface of an air preheater essentially increases the thermal resistance of the gas-solid interface, leading to a significant increase in the thermal inertia of the blocked area. When the unit load is stable, this thermal inertia manifests as a slow spatial temperature shift. However, when the unit is in a rapid load reduction or increase unsteady-state condition, the normal, unblocked heating surface area can closely follow the flue gas temperature and generate a dramatic temporal temperature gradient change. The area experiencing ash accumulation and blockage is doubly constrained by the impediment to fluid scouring and the insulating effect of the adhering layer, causing its actual physical temperature change rate to lag far behind theoretical expectations. A fixed global penalty hyperparameter cannot distinguish whether the anomaly in this area is due to a static anomaly caused by slow ash accumulation or a dynamic anomaly exposed by thermal inertia hysteresis at the moment of load change. To resolve the aforementioned physical constraint conflict, this embodiment assigns a dynamic adjustment attribute to the global penalty hyperparameter in the interactive correction loss function. After receiving the spatial abrupt boundary signal, the evidence fusion module extracts the set of mesh nodes contained within the spatial abrupt boundary signal. The historical time window is configured to span 60 consecutive discrete time steps, corresponding to a physical time span of 30 seconds. The actual temperature grid signals continuously collected within the set of grid nodes within the defined historical time window are extracted. The grid signals from different time steps are then reassembled into a three-dimensional spatiotemporal tensor in chronological order to form the actual temperature time series characterizing the evolution of the true temperature.

[0087] The dynamic reference signal of the heated surface, output in real time by the dynamic mapping module within the set historical time window, is synchronously extracted from the grid node set. The same tensor reconstruction logic is used to generate the reference temperature time series characterizing the theoretical temperature evolution. The actual temperature time series and the reference temperature time series are strictly aligned on the physical time axis to eliminate time delay errors caused by system sampling jitter. The actual temperature time series and the reference temperature time series are compared, and the temporal temperature gradient decay rate of the actual temperature time series relative to the reference temperature time series is calculated. The temporal temperature gradient decay rate characterizes the degree of hysteresis between the actual physical temperature and the theoretically expected temperature of the model within the grid node set. Based on the physical mapping relationship of the heat conduction differential equation, the degree of hysteresis is positively correlated with the blockage thickness in that region. The temporal temperature gradient decay rate is calculated and converted into the global penalty hyperparameter, which is performed using a specific nonlinear mathematical calculation model.

[0088] ;

[0089] In the formula, The global penalty hyperparameters are actually substituted into the interactive correction loss function after dynamic adjustment; The base penalty coefficient is set; It is the hyperbolic tangent activation function; This is the mapping steepness adjustment coefficient; The total number of grid nodes contained in the set of grid nodes; The total number of discrete time steps included in the set historical time window; The set of grid nodes; For the first node in the set of grid nodes The grid node at the ... The actual temperature grid signal values ​​corresponding to each discrete time step; For the first node in the set of grid nodes The grid node at the ... The dynamic reference signal values ​​of the heated surface corresponding to each discrete time step; The partial derivative of the actual temperature grid signal value in the time dimension; The partial derivative of the dynamic reference signal value of the heated surface in the time dimension; This is a preset thermal inertia tolerance threshold. In the low-level computation flow of a microprocessor, the partial derivatives in the time dimension cannot be directly obtained as continuous analytical solutions; they must be discretized and approximated using the finite difference method.

[0090] For the current discrete time step Adjacent historical time steps The partial derivatives are calculated using the backward difference formula, i.e. Approximately equal to minus The difference is divided by the physical time interval between adjacent time steps. The partial derivatives of the dynamic reference signal values ​​of the heated surface use the same backward differential logic. After calculating the partial derivatives of the actual temperature and the reference temperature at each moment, the absolute value of the difference between the two is obtained.

[0091] A double summation operation is performed on all nodes within the grid node set and all time steps within the set historical time window. The summation result is then divided by the product of the total number of nodes and the total number of time steps to obtain the mean. The preset thermal inertia tolerance threshold is then subtracted from the mean. The thermal inertia tolerance threshold filters out the minute thermal inertia fluctuations inherent in normal metallic materials and the thermal noise from the infrared sensor. The net attenuation rate after threshold removal is multiplied by the mapping steepness adjustment coefficient. Input into the hyperbolic tangent activation function The hyperbolic tangent activation function smoothly maps the input value to... to Within the convergence interval, the penalty coefficient is limited to prevent exponential divergence under extreme conditions, ensuring gradient stability of the spatiotemporal graph convolutional network during backpropagation. Mapping value addition The base penalty coefficient for subsequent multiplication is set. In the process, the final global penalty hyperparameters are output. By introducing this dynamic adjustment model, during steady-state operation of the unit, the partial derivatives of both the actual temperature and the reference temperature approach zero, and the net attenuation rate is negative. After hyperbolic tangent mapping, it becomes a negative decimal. At this point, the global penalty hyperparameter is less than the basic penalty coefficient, preventing the model from overfitting to weak anomalies. When the unit operates under variable load and the blocked area is exposed to extreme thermal inertia, the absolute value of the difference in partial derivatives rises sharply, and the hyperbolic tangent mapping value rapidly approaches... The global penalty hyperparameter is amplified to nearly twice the base penalty coefficient. This amplified penalty weight dominates the interactive correction loss function, severely hindering the spatiotemporal graph convolutional network's blind learning of the current spurious benchmark and forcibly anchoring the dynamic benchmark signal to the theoretically predicted trajectory in a non-blocking state.

[0092] In specific implementation, the operating scenario is set as a deep peak-shaving phase where a 600MW supercritical unit, under grid dispatch requirements, urgently reduces its load from 600MW to 400MW at a rate of 30MW / min. At this time, the flue gas temperature at the air preheater inlet experiences a step drop. The dual-line extraction module continuously operates, capturing the spatial abrupt boundary signal in a specific physical coordinate region at the cold end of the heated surface. The set of grid nodes within the envelope of this boundary signal is then extracted. Total number of nodes equal The memory matrix buffer pool of the edge computing node retains continuous... The raw data of the frame, setting the historical time window equal sampling frequency Hz corresponds to total time s.

[0093] Here During the rapid load reduction process of s, the non-clogging area should have experienced a rapid cooling of the flue gas temperature. For the first node in the aforementioned grid node set... The dynamic reference signal of the heated surface output by the dynamic mapping module accurately predicts this cooling trend through a discrete grid. The finite difference method is used to obtain the... Partial derivatives at the center points of each grid Its value has always been in ℃ / s to Fluctuating between ℃ / s.

[0094] The mesh node is attached with a thickness of approximately The mixture of ammonium bisulfate and fly ash forms a clump. Due to the thermal barrier effect of the clump, its actual cooling rate is severely delayed. The partial derivative of the actual temperature grid signal extracted by the feature comparison module... The value only remained at ℃ / s to Between ℃ / s.

[0095] At a single time step, the absolute difference between the partial derivatives of the two is approximately ℃ / s. Parallel accumulators in edge computing systems Each grid node experienced The absolute differences of partial derivatives at each time step are summed twice and the mean is taken to obtain the scalar value of the overall time-series temperature gradient decay rate. ℃ / s.

[0096] Proceeding to the nonlinear mapping function calculation stage. The system memory retrieves the preset algorithm basis parameters: thermal inertia tolerance threshold. Set as ℃ / s, eliminating the normal temperature measurement delay caused by the intrinsic heat capacity of the metal substrate; mapping steepness adjustment coefficient Set as Base penalty coefficient Set as .

[0097] Substitute the above scalar values ​​into the core part of the formula to calculate the activation parameters: The underlying mathematical library function is called to calculate the hyperbolic tangent. equal The final output is the global penalty hyperparameter. In a normal open-loop state or with fixed penalty parameters, the spatiotemporal graph convolutional network will assume that the measured values ​​are... Slow cooling at ℃ / s is a normal characteristic of the equipment under the current load. In the next forward mapping, the predicted baseline cooling rate is reduced to accommodate the erroneous data, causing the residual signal to be completely "eaten up" within seconds.

[0098] In this embodiment, The excess dynamic penalty weights are forcibly injected into the interactive correction loss function, and the backpropagation gradients within this grid node set are completely taken over by the penalty term. The network weights cannot be updated to accommodate the slow cooling, and the dynamic baseline signal continues to be maintained. Rapid cooling prediction at ℃ / s. As real physical time progresses, the absolute temperature difference between the actual measured temperature and the locked-in, uncontaminated predicted reference temperature increases from the initial... The temperature rapidly increased to ℃. The temperature residual characteristic signal forms an extremely steep "cliff-like" edge in this region, and the amplitude of the blockage situation confidence signal ranges from... Directly leap to .

[0099] Table 1. Performance Comparison of Static and Dynamic Global Penalty Hyperparameters:

[0100] Full load steady state operation Static fixed ( ) 150 ms 1.5% 12.4% Full load steady state operation Dynamic adjustment (model in this embodiment) 155 ms 1.2% 8.6% Rapid load reduction and deep peak shaving Static fixed ( ) 420 ms 28.7% 35.1% Rapid load reduction and deep peak shaving Dynamic adjustment (model in this embodiment) 185 ms 3.4% 4.2% Dramatic changes in ambient air supply temperature Static fixed ( ) 350 ms 21.5% 22.8% Dramatic changes in ambient air supply temperature Dynamic adjustment (model in this embodiment) 170 ms 2.8% 3.5%

[0101] Based on the large-scale test results recorded in Table 1, the global penalty hyperparameters are calculated using the dynamic adjustment model. Under stable operating conditions, this does not increase the additional computing power burden, and the model convergence time only increases. The false alarm rate remains at an extremely low level. When the unit enters typical unsteady-state conditions such as rapid load reduction for deep peak shaving or encounters drastic changes in ambient supply air temperature, the dynamic adjustment mechanism plays a crucial protective role. Static parameters, unable to adapt to the significant temperature difference caused by thermal inertia hysteresis, are highly susceptible to model collapse or large-scale misjudgments, causing the false alarm rate to soar. The dynamic adjustment model, by capturing the physical differences in partial derivatives, precisely positions anomalous nodes and blocks baseline contamination, forcibly suppressing the false alarm rate under such harsh operating conditions. Within the safe range, the model's adaptive anti-interference capability in the face of sudden operating conditions has been fundamentally improved, effectively avoiding the operational risks of the fan system caused by erroneous dust removal commands, and providing a detection closed-loop control system with extremely high industrial robustness.

[0102] Example 3: This example is based on the same principles as Example 1, but differs in that it interfaces with the power plant's intelligent operation and maintenance platform to achieve coordinated control of the blockage detection and soot blowing systems. The blockage location coordinates and severity signals output by the detection system are transmitted to the power plant's intelligent operation and maintenance platform via industrial Ethernet. The platform then makes comprehensive decisions based on multi-source information, including unit operating parameters, historical soot blowing records, and coal quality data.

[0103] When a minor blockage is detected, the maintenance platform issues an early warning, records the blockage location and its development trend, and temporarily suspends soot blowing operations. Operators can then schedule inspections to confirm the blockage based on the actual situation. When a moderate blockage is detected, the maintenance platform automatically generates targeted soot blowing instructions, adjusting the blowing area and duration of the steam soot blower. This ensures precise soot blowing is performed only on the blocked area, avoiding steam waste and component wear caused by blowing the entire area.

[0104] When a severe blockage is detected, the operations and maintenance platform triggers a high-level alarm and simultaneously adjusts the output of the induced draft fan to prevent excessive ventilation resistance from causing fan stall. After the soot blowing operation is completed, the detection system automatically evaluates the soot blowing effect, compares the change in blockage confidence before and after soot blowing, and generates a soot blowing effect evaluation report. If the blockage is not effectively relieved after soot blowing, the system automatically upgrades the blockage level and issues a secondary alarm, prompting operators to take further measures.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for locating and detecting local blockages in an air preheater by fusing infrared images, characterized in that, Includes the following steps: The initial infrared video stream signal of the heated surface of the air preheater is collected, and the initial infrared video stream signal is preprocessed with emissivity dynamic correction to obtain a standardized infrared thermal spatiotemporal signal. Based on standardized infrared thermal spatiotemporal signals, a spatiotemporal graph convolutional network is used to perform spatiotemporal graph convolutional mapping processing to extract the evolution features of the time dimension and the correlation features of the spatial dimension, thereby obtaining the dynamic reference signal of the heated surface. The actual temperature grid signal is extracted from the standardized infrared thermal spatiotemporal signal. The actual temperature grid signal is compared with the dynamic reference signal of the heated surface and subjected to first-order differential operation in multiple spatial directions to obtain the temperature residual characteristic signal. Adaptive density clustering is performed on the temperature residual feature signal. The set of connected high-density anomalous gradient grids is searched by calculating the core distance to obtain the spatial abrupt boundary signal. At the same time, local outlier feature extraction is performed on the temperature residual feature signal. The degree of outlier of the grid relative to the neighborhood is evaluated by comparing the actual distance with the nearest neighbor reachable distance to obtain the local outlier heat source signal. The spatial mutation boundary signal is fed back to the spatiotemporal graph convolutional network for cross-correction until the interaction termination condition is met. After the interaction termination condition is met, the spatial mutation boundary signal and the local outlier heat source signal are subjected to evidence chain fusion processing. The orthogonal combination rule is used to eliminate conflicts and contradictions and calculate the joint basic probability allocation value to obtain the blockage situation confidence signal. The blockage situation confidence signal is graded and judged according to its amplitude range. When the amplitude exceeds the judgment threshold, the local blockage location coordinate signal and severity signal of the air preheater are output.

2. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The initial infrared video stream signal is preprocessed with dynamic emissivity correction, specifically including: extracting the reflectivity constant of the heated surface material and the atmospheric transmittance attenuation constant, and constructing the Planck correction function; using the Planck correction function to perform nonlinear inversion on the apparent radiance recorded by the initial infrared video stream signal, and reconstructing the standardized infrared thermal spatiotemporal signal characterizing the physical absolute temperature.

3. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: Spatiotemporal graph convolutional mapping is performed using a spatiotemporal graph convolutional network, specifically including: dividing the heating surface of the air preheater into discrete temperature grids; constructing a topological undirected graph structure for the corresponding heating surface using the discrete temperature grids as nodes and the physically adjacent heat conduction paths of the grids as edges; and performing forward propagation on the topological undirected graph structure using Chebyshev polynomial approximation graph convolution kernels with self-loop adjacency matrices.

4. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The actual temperature grid signal is compared with the dynamic reference signal of the heated surface. Specifically, this includes: registering nodes according to the same physical coordinate system and calculating the absolute difference between the actual temperature grid signal value and the predicted value of the dynamic reference signal of the heated surface at each coordinate node.

5. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The spatial abrupt boundary signal is fed back to the spatiotemporal graph convolutional network for cross-correction. Specifically, this includes: extracting the set of grid nodes contained in the spatial abrupt boundary signal and calculating the average density confidence corresponding to the set of grid nodes; feeding the average density confidence as a penalty weight back to the loss function of the spatiotemporal graph convolutional network in real time for joint correction of model parameters; and the interaction termination condition is when the first derivative value of the interactive correction loss function output in two consecutive calculation cycles is lower than a preset convergence threshold.

6. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 5, characterized in that: The interactive correction loss function, upon which the joint correction of model parameters depends, is constructed according to the following rules: The interactive correction loss function includes a basic mean squared error loss function and an additional feedback penalty term; The feedback penalty term is obtained by calculating the squared difference between the predicted baseline temperature and the historical unblocked real physical temperature within the grid node set, and multiplying it by the global penalty hyperparameter and the average density confidence value.

7. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The classification is based on the amplitude range of the congestion situation confidence signal, specifically including: establishing a mapping lookup table from continuous confidence space to discrete severity intervals, performing inverse matrix transformation on the coordinates of grid nodes exceeding the judgment threshold, and outputting two-dimensional geometric positioning parameters in Cartesian coordinate system.

8. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The core distance is calculated by searching for a set of high-density anomaly gradient grids that are connected. The following logic is executed when calculating the core distance: The difference in the physical space span of the horizontal and vertical coordinates of the grid nodes in the heated surface plane coordinate system is comprehensively considered, as well as the difference in the numerical space span of the temperature residual characteristic value adjusted by introducing a temperature difference weighting coefficient.

9. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The assessment of the outlier degree of the grid relative to its neighborhood specifically includes: extracting the local reachability density value of the assessment node; calculating the ratio of the sum of the reachability density values ​​of all neighbors in the nearest neighborhood of the assessment node to the corresponding reachability density value of the assessment node, and generating a local anomaly factor score characterizing independent thermal anomalies.

10. The method for locating and detecting local blockages in an air preheater by fusing infrared images according to claim 1, characterized in that: The calculation of the joint basic probability assignment includes: extracting the first basic probability value for assigning congestion events based on the results of adaptive density clustering; extracting the second basic probability value for assigning congestion events based on the local anomaly factor score; calculating the orthogonal product of the first and second basic probability values, and dividing by the normalized denominator after conflict coefficient correction.

Images