Infrared temperature measurement intelligent diagnosis system for blockage of indirect cooling fin

Abstract

The invention belongs to the field of power station intelligent air cooling, and discloses an indirect cooling fin blockage infrared temperature measurement intelligent diagnosis system, which comprises a temperature detection module, an infrared temperature measurement module and an infrared temperature measurement module, the data processing module is used for constructing a calibration rule to calibrate the infrared thermal image sequence, and feature fusion is carried out to generate an enhanced thermal image feature vector set; the intelligent identification module is used for calculating the thermodynamic correlation degree and the infrared feature similarity to construct an infrared thermogram feature correlation map, outputting the temperature anomaly probability by constructing an infrared intelligent judgment model to judge the blockage risk and the smudginess level, and introducing a local anomaly propagation coefficient to be combined with reverse heat conduction simulation to judge an omission blockage smudginess area; and the diagnosis optimization module is used for judging the comprehensive evaluation coefficient and spatial similarity, eliminating false alarms to generate a preliminary verification result, constructing an early warning constraint to judge a comprehensive early warning level, dynamically adjusting judgment logic to form a diagnosis optimization result, and realizing intelligent diagnosis of fin blockage and smudginess states.

Description

Technical Field

[0001] This invention belongs to the field of intelligent air-cooling technology for power plants, and relates to an intelligent diagnostic system for intercooler fin blockage using infrared temperature measurement. Background Technology

[0002] In the operation of indirect air-cooled systems in power plants, finned tube bundle blockage is the core problem leading to a decrease in heat exchange efficiency. Traditional methods rely on manual periodic inspections and differential pressure assessments, which suffer from poor timeliness, vague location, and inability to quantify the degree of blockage and fouling. Although existing research uses fixed infrared thermal imagers for monitoring, the fixed locations and limited viewing angles make it difficult to comprehensively capture local blockage hotspots in the air-cooled tower. Furthermore, the analysis of thermal images is mostly limited to manual interpretation, lacking intelligent diagnostic and early warning capabilities from temperature field to blockage level, resulting in delayed maintenance decisions and affecting the economic efficiency and safety of unit operation.

[0003] A Chinese patent with authorization announcement number CN112905947B discloses a method for real-time monitoring of the fouling degree of an indirect air-cooled tower finned tube heat exchanger. The method includes a real-time operation data acquisition module, a data preprocessing module, an online fouling degree monitoring module, a parameter storage module, and an output display module. The operation data acquisition module acquires and records the required real-time operation data of the unit and sends the data to the data preprocessing module. The data preprocessing module cleans the data, removing data that exceeds temperature or limits. The online fouling degree monitoring module calculates the fouling coefficient of the corresponding indirect air-cooled tower heat exchanger tube bundle. , thermal resistance due to dirt on the outside of the pipe Overall heat transfer coefficient of clean pipe Actual overall heat transfer coefficient and in each calculation formula , , , Parameter values. The beneficial effects of this invention are: all the basic data of this invention come from actual field measurements, and rigorous formula derivation is used to ensure that the physical concepts are clear, yet suitable for on-site calculation.

[0004] Although there is an existing method for real-time monitoring of the fouling degree of finned tube heat exchangers in indirect air-cooled towers, which calculates the fouling coefficient and the thermal resistance of fouling on the outside of the tubes based on field operation data, and realizes real-time monitoring and data visualization output of the fouling degree of finned tube heat exchangers, it has the problems of difficulty in covering the temperature distribution of large fin areas, easy to miss local blocked and dirty areas, and inability to accurately determine the degree of fin blockage and fouling through temperature measurement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent diagnostic system for indirect cooling fin blockage using infrared thermography. This system collects infrared thermogram sequences, combines environmental parameter correction and image enhancement processing to generate an enhanced thermogram feature vector set, utilizes spatial clustering and thermal correlation analysis to construct a dynamic spectral structure, builds an infrared intelligent judgment model to learn the propagation patterns of neighborhood anomalies, and achieves refined classification of blockage risk and contamination levels. It introduces a local anomaly propagation coefficient combined with reverse heat conduction simulation to determine missed blockage and contamination areas, incorporates expert rules for false alarm verification, and constructs a dual early warning mechanism combined with reinforcement learning to dynamically optimize the diagnostic strategy. This achieves intelligent and high-precision diagnosis of the indirect cooling fin status, improving the operational reliability and maintenance efficiency of power plant air-cooling systems.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The infrared temperature measurement intelligent diagnostic system for intercooled fin blockage includes: a temperature detection module, a data processing module, an intelligent identification module, and a diagnostic optimization module;

[0008] The temperature detection module is used to acquire infrared thermal image sequences in real time;

[0009] The data processing module is used to construct environmental temperature calibration rules and humidity calibration rules, calibrate infrared thermal image sequences, extract contour features to generate contour feature vectors, calculate gray-level co-occurrence matrix to generate texture feature vectors, and weighted fuse to generate enhanced thermal image feature vector sets.

[0010] The intelligent recognition module is used to calculate the thermal correlation degree and infrared feature similarity to obtain a dynamic edge weight matrix, and construct an infrared thermal image feature correlation map; construct an infrared intelligent judgment model, output the temperature anomaly probability, and determine the blockage risk level and dirt level; introduce the local anomaly propagation coefficient, and combine it with reverse heat conduction simulation to determine the missed blockage and dirt areas.

[0011] The diagnostic optimization module is used to determine the comprehensive evaluation coefficient and spatial similarity, exclude suspected false alarm areas, and generate preliminary verification results; it constructs congestion warning constraints and temperature warning constraints to determine the comprehensive warning level, and dynamically adjusts the judgment logic to form diagnostic optimization results.

[0012] Specifically, the steps for generating the enhanced heatmap feature vector set include:

[0013] Establish an ambient temperature calibration rule: if the actual ambient temperature is higher than the reference ambient temperature, perform temperature calibration and update the infrared thermal image sequence with the first calibration temperature as the replacement value; otherwise, maintain the original infrared thermal image sequence.

[0014] A humidity calibration rule is established: if the ambient humidity exceeds the preset humidity threshold, the second calibration temperature is used as the replacement value to update the infrared thermal image sequence; otherwise, the infrared thermal image sequence after ambient temperature calibration is maintained.

[0015] Extract the contour features of the fin edge, mark the contour break points, calculate the contour integrity ratio, and calculate the break point density.

[0016] Set the fin spacing, calculate the fin spacing deviation, and combine the contour integrity ratio and the break point density to generate a contour feature vector;

[0017] By combining the radial and tangential directions, the gray-level co-occurrence matrix is ​​calculated, and texture feature vectors are generated;

[0018] By weighted fusion, an enhanced heatmap feature vector set is generated.

[0019] Specifically, the intelligent recognition module includes a map construction unit, a model construction unit, and a level determination unit;

[0020] The graph construction unit is used to form a set of heat map nodes through K-means clustering algorithm, calculate the thermal correlation between adjacent fins and the infrared feature similarity between nodes to obtain a dynamic edge weight matrix, and construct an infrared heat map feature correlation graph.

[0021] The model building unit is used to build an infrared intelligent judgment model based on graph neural network. It inputs the infrared thermal image feature association spectrum and, combined with a multi-layer message passing mechanism, outputs the probability of temperature anomalies at each node.

[0022] The level determination unit is used to construct blockage level determination conditions to determine the blockage risk level; construct a historical clean status benchmark value table, calculate the pollution sensitivity in combination with the temperature deviation formula, and construct dirt level determination conditions to determine the dirt level.

[0023] The omission area determination unit is used to introduce the local abnormal propagation coefficient, calculate the comprehensive thermal conduction influence value, determine the omission, blockage and dirt areas, and verify the determination results by combining reverse thermal conduction simulation.

[0024] Specifically, the steps for constructing an infrared thermal image feature correlation map include:

[0025] The spatial coordinates of the enhanced heatmap feature vector set are extracted, and the K-means clustering algorithm is used to divide the clusters and determine the adjacency relationship between the clusters to form a heatmap node set with a topological structure.

[0026] Based on the node topology, thermal resistance is calculated by combining physical spacing, contact area and thermal conductivity coefficient. The maximum thermal resistance of all adjacent node pairs is screened in descending order and the thermal resistance ratio with the current thermal resistance is calculated. The thermal correlation degree is calculated by combining the basic coefficient.

[0027] Extract the mean feature vector of each node, and combine it with the standardized mean feature vector of adjacent nodes to calculate the infrared feature similarity.

[0028] Set and dynamically adjust the thermal correlation threshold and the infrared feature similarity threshold, and obtain a dynamic edge weight matrix through weighted calculation and fusion.

[0029] Using heatmap nodes as vertices and the dynamic edge weight matrix as edges, an infrared heatmap feature association map with dynamic edge weights is constructed.

[0030] Specifically, the steps for constructing the infrared intelligent judgment model include:

[0031] An infrared intelligent judgment model is constructed based on graph neural networks, including an input layer, a perception enhancement layer, a diffusion attention layer, and an output layer;

[0032] The input layer extracts the temperature gradient, calculates the deviation value of each node, and generates a list of temperature deviation values.

[0033] The perception enhancement layer determines temperature anomalies based on a preset temperature difference threshold, introduces enhancement coefficients to strengthen features for temperature anomaly nodes, and forms a perception enhancement feature matrix.

[0034] The diffusion attention layer calculates the diffusion correlation degree between each node and its neighboring nodes, constructs a spatial diffusion attention weight matrix, and fuses it with the perception enhancement feature matrix in a weighted manner to generate a diffusion perception feature matrix;

[0035] The output layer calculates the initial temperature anomaly probability and combines it with the neighborhood average temperature anomaly probability to determine isolated nodes and non-isolated nodes. It reduces the initial temperature anomaly probability of isolated nodes that exceed a preset limit and outputs the temperature anomaly probability of each node. ;

[0036] Using binary cross-entropy as the loss function, a heat map sample set is formed based on the infrared heat map feature association spectrum, and the set is divided into a training set, a validation set, and a test set. The infrared intelligent judgment model is then trained using the Adam optimizer.

[0037] Specifically, the steps for determining the level of blockage risk and the level of contamination include:

[0038] Set the maximum probability threshold With minimum probability threshold ;

[0039] Construct congestion level determination criteria, if If so, then it is determined to be a high-risk candidate; if If so, then it is determined to be a medium-risk candidate; if If so, it is directly judged as low risk;

[0040] For high-risk candidate nodes, count the total number of high-risk and medium-risk candidate nodes among all adjacent nodes. If the total number exceeds the preset first-level threshold, it is judged as high-risk; otherwise, it is downgraded to medium-risk.

[0041] For medium-risk candidate nodes, count the total number of medium-risk candidate nodes among all adjacent nodes. If the total number exceeds the preset secondary threshold, the node is classified as medium-risk; otherwise, it is downgraded to low-risk.

[0042] Specifically, the steps for determining the level of blockage risk and the level of contamination also include:

[0043] Construct a historical cleanliness baseline table and calculate the probability of temperature anomalies. Compared with historical cleanliness baseline values Difference;

[0044] A deviation coefficient is introduced, and a baseline compensation value is obtained by combining the historical cleanliness baseline value. The pollution sensitivity of each node is determined by the ratio of the difference to the baseline compensation value. This ultimately leads to the temperature deviation formula;

[0045] like ,but ;otherwise ;

[0046] Set the maximum deviation threshold With the lowest deviation threshold ;

[0047] Construct conditions for determining the level of dirtiness, if If so, it is classified as Level 1 pollution; if If so, it is classified as Level II pollution; if If so, it is classified as Level 3 pollution.

[0048] Specifically, the steps for determining areas that have been missed or blocked by dirt include:

[0049] Extract the probability of temperature anomalies at each node and set a baseline threshold. ,like If the condition is met, then the node is considered a high-abnormal node; otherwise, it is considered a low-abnormal node.

[0050] By introducing a local anomaly propagation coefficient, and combining the anomaly degree coefficient of high anomaly nodes, the comprehensive thermal conduction impact value is calculated for each low anomaly node.

[0051] If the combined thermal conduction influence value exceeds the preset propagation threshold, it is determined to be a node to be verified. The temperature gradient change rate is calculated, and the deviation rate between the temperature gradient change rate and the tube bundle heat dissipation curve is compared.

[0052] If the deviation rate is greater than the preset deviation rate threshold, it is determined to be a suspected missing area node. The reverse heat conduction model is used to simulate the theoretical normal temperature, and the residual is calculated in combination with the actual detection temperature.

[0053] If the residual is greater than the preset residual threshold, it is determined that the suspected missing area node is blocked or dirty. All suspected missing area nodes with blockage or dirt are integrated to form a missing blocked or dirty area.

[0054] Specifically, the steps for generating preliminary verification results include:

[0055] The blockage risk level is mapped to a risk quantification value, and the dirt level is mapped to a pollution quantification value. The comprehensive evaluation coefficient for each monitoring area is calculated by weighting the values.

[0056] Use the historical comprehensive evaluation coefficients of the monitored area to calculate the absolute difference between the coefficients and the historical comprehensive evaluation coefficients.

[0057] If the absolute difference is lower than the preset difference threshold, the comprehensive evaluation coefficients are considered similar; otherwise, they are considered dissimilar.

[0058] Similar candidate cases are selected from the historical diagnostic database, and the Euclidean distance is calculated.

[0059] If the Euclidean distance is lower than a preset distance threshold, the spaces are considered similar; otherwise, they are considered dissimilar.

[0060] Screen similar candidate cases that simultaneously meet the requirements of comprehensive evaluation coefficient similarity and spatial similarity to form a similar case set, calculate the blocking ratio and false alarm ratio, and form a historical matching tendency label table;

[0061] An expert rule base is constructed. Based on the matching rules of the historical matching tendency label table, the preliminary verification categories are determined as severely abnormal areas, abnormal areas to be observed, and suspected false alarm areas. Suspected false alarm areas are filtered out, and preliminary verification results are generated.

[0062] Specifically, the steps to generate diagnostic optimization results include:

[0063] For severely abnormal areas and areas to be observed for abnormality, calculate the proportion of blocked area. Temperature deviation from dirty areas ;

[0064] Set a threshold for the percentage of blocked area. , ;

[0065] Construct congestion early warning constraints, if If so, a Level 1 congestion warning is issued; if If so, a level 2 congestion warning is issued; if If so, a Level 3 congestion warning is issued;

[0066] Set temperature deviation threshold , ;

[0067] Construct temperature early warning constraints, if If so, a Level 1 temperature warning is issued; if If so, a level two temperature warning is issued; if If so, a Level III temperature warning will be issued;

[0068] If both the Level 1 congestion warning and the Level 1 temperature warning are met simultaneously, it is determined to be a Level 1 comprehensive warning;

[0069] If both a Level 3 congestion warning and a Level 3 temperature warning are met simultaneously, it is determined to be a Level 3 comprehensive warning;

[0070] Otherwise, it will be classified as a Level II comprehensive warning;

[0071] Based on the comprehensive early warning level, and combined with reinforcement learning to output the optimal diagnostic optimization strategy, the judgment logic of congestion warning and temperature warning is dynamically adjusted to form a diagnostic optimization result.

[0072] The beneficial effects of this invention are:

[0073] 1. This invention is based on an enhanced thermal image feature vector set, uses a spatial clustering algorithm to form a set of thermal image nodes, and integrates physical thermal resistance characteristics and infrared feature similarity to construct a feature association map with dynamic edge weights; it adopts a graph neural network model to aggregate temperature anomaly information of neighboring nodes through a multi-layer message passing mechanism, effectively learning the local thermal accumulation and diffusion laws, and avoiding misjudgment of isolated noise points; after outputting the temperature anomaly probability, it combines the probability threshold and the historical cleanliness benchmark value for dual discrimination, realizing a refined division of blockage risk level and dirt level, and introduces a local anomaly propagation coefficient combined with the reverse heat conduction simulation residual to determine the missed blockage and dirt areas, realizing intelligent perception from local anomalies to regional patterns, and improving diagnostic accuracy and robustness.

[0074] 2. Based on preliminary identification, this invention integrates a historical diagnostic database and constructs an expert rule base. Through comprehensive evaluation coefficient matching and spatial distribution similarity analysis, it constructs historical matching tendency labels to effectively eliminate isolated noise and false alarm areas. Combining the proportion of blocked area and temperature deviation, it constructs a dual early warning constraint to determine the comprehensive early warning level. With the comprehensive early warning level, the proportion of blocked area, temperature deviation, and real-time operating parameters as states, the diagnostic strategy adjustment is the action, and the fault matching degree is the reward. Through the DQN algorithm, it learns the optimal optimization strategy and dynamically adjusts the early warning judgment logic, enabling the system to have self-learning and self-adaptive capabilities. Based on actual operation feedback, it optimizes the diagnostic results, improves the timeliness and accuracy of early warnings, and enhances the operational safety and maintenance efficiency of the power plant air-cooled system. Attached Figure Description

[0075] Figure 1 Structural diagram of an infrared temperature measurement intelligent diagnostic system for intercooled fin blockage;

[0076] Figure 2 This is a flowchart of the process for generating the enhanced heatmap feature vector set in this invention;

[0077] Figure 3 This is a flowchart illustrating the determination of blockage risk level and dirt level in this invention;

[0078] Figure 4 This is a flowchart for determining the areas of missing blockage and dirt in this invention;

[0079] Figure 5 This is a schematic diagram of the infrared thermal image feature correlation spectrum in this invention. Detailed Implementation

[0080] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0081] refer to Figures 1 to 5 As shown in the figure, this embodiment introduces an infrared temperature measurement intelligent diagnostic system for intercooled fin blockage, including a temperature detection module, a data processing module, an intelligent identification module, and a diagnostic optimization module;

[0082] The temperature detection module uses an infrared thermal imager deployed at the end of the flexible robot, combined with infrared thermal imaging technology, to continuously scan the surface of the finned tube heat exchanger in the power plant's indirect air-cooled system along a preset gridded inspection path, acquiring high-resolution infrared thermal image sequences in real time. The flexible robot possesses multiple degrees of freedom of bending and extension capabilities, adapting to the narrow gaps and irregular curved surfaces between fins in the power plant's indirect air-cooled system. It moves along the fin surface via magnetic adsorption or wheel-driven propulsion. The gridded inspection path is based on the fin array size of the power plant's indirect air-cooled system's intermediate cold finned tube heat exchanger, dividing the entire fin area into... The sampling units are equally distributed, and during path optimization, priority is given to covering areas with dense fins and areas with a high incidence of historical blockages, while avoiding obstacles such as pipelines, to ensure that each sampling unit can be scanned and covered.

[0083] The data processing module is used to preprocess and extract features from the received infrared thermal image sequence. Based on the spatial coordinates of the sampling units, invalid temperature data that deviates from the normal operating temperature range of the fins, is repeatedly collected, or is caused by signal interference is removed. Missing data is filled by interpolation of the mean temperature of the neighboring sampling units. Combined with real-time environmental parameters, such as ambient temperature, ambient temperature calibration rules and humidity calibration rules are constructed to calibrate the infrared thermal image sequence to obtain a calibrated infrared thermal image sequence. A Gaussian filtering algorithm is used to perform noise reduction filtering on the calibrated infrared thermal image sequence to reduce image noise interference and obtain an enhanced infrared thermal image sequence. The Canny edge detection algorithm is used to extract the contour features of the fin edges from the enhanced infrared thermal image sequence. After feature quantization and dimension standardization, the contour feature vector is obtained. The energy, entropy, and contrast of the thermal image are calculated through the gray-level co-occurrence matrix to quantize the texture feature vector of the thermal image. The contour feature vector and texture feature vector of each frame of the thermal image are weighted and fused to generate an enhanced thermal image feature vector set containing spatial coordinate information.

[0084] The intelligent identification module is used to analyze the spatial distribution of the enhanced thermal image feature vector set based on the distribution pattern of the finned tube bundles in the power plant, forming a set of thermal image nodes with a topological structure. Based on the physical characteristics and thermal conductivity coefficients of the tube bundles between nodes, it calculates the thermal correlation between adjacent fins and, combined with the enhanced thermal image feature vectors of each node, calculates the infrared feature similarity between nodes. By weighted fusion of thermal correlation and infrared feature similarity, a dynamic edge weight matrix is ​​obtained, and an infrared thermal image feature correlation map is constructed. A machine learning algorithm, such as the graphical neural network GCN, is used to construct an infrared intelligent judgment model. Taking the infrared thermal image feature correlation map as input, a multi-layer message passing mechanism is used to aggregate temperature anomaly information of neighboring nodes, learn the thermal accumulation pattern of local areas, and achieve full connectivity. The system employs a layer-by-layer probability mapping to output the temperature anomaly probability for each node. Based on probability thresholds, it classifies blockage risk levels and calculates the contamination sensitivity between the temperature anomaly probability and historical cleanliness baseline values. Based on deviation thresholds, it classifies contamination levels. A local anomaly propagation coefficient is introduced to calculate the comprehensive thermal conduction impact of low-anomaly nodes on neighboring high-anomaly nodes. If the comprehensive thermal conduction impact exceeds a preset propagation threshold and the node's temperature gradient change rate deviates from the tube bundle heat dissipation curve, a missed blockage / contamination area is identified. The determination results are verified using reverse heat conduction simulation and residual analysis. The local anomaly propagation coefficient is calculated based on the thermal correlation between a node and its neighboring high-anomaly nodes, weighted by the temperature anomaly probability of neighboring nodes, thus quantifying the intensity of the thermal conduction impact of high-anomaly nodes on low-anomaly nodes.

[0085] The diagnostic optimization module is used to perform multi-source data verification. Based on the blockage risk level and the level of contamination, it performs correlation analysis with the historical blockage diagnosis database, builds an expert rule base for logical verification, eliminates false alarms, generates preliminary verification results, classifies warning levels by the proportion of blockage area and the degree of temperature deviation of contaminated area, triggers the reinforcement learning time-series optimization mechanism, and forms diagnostic optimization results, including but not limited to the spatial distribution of blockage area and the optimized warning level.

[0086] Specifically, the steps for generating the enhanced heatmap feature vector set include:

[0087] Based on the infrared thermal image sequence, each frame of the thermal image is traversed. According to the coordinates of the sampling unit, pixels whose temperature values ​​deviate from the normal operating range of the fins are marked as invalid data, and thermal image segments with duplicate acquisition and signal interference are removed.

[0088] For missing data, such as blank sampling units caused by obstacles, calculate the average temperature of the effective sampling units in the neighborhood and use bilinear interpolation to fill in the missing data;

[0089] Since changes in ambient temperature can cause the original fin temperature detection value to deviate from the actual operating conditions, and humidity can also interfere with the temperature value due to the absorption of infrared radiation, environmental interference needs to be eliminated through temperature and humidity calibration.

[0090] Set a humidity threshold based on real-time environmental parameters, such as ambient temperature and humidity. ;

[0091] An ambient temperature calibration rule is established based on the collected original fin temperature. If the actual ambient temperature is higher than the reference ambient temperature, the difference between the actual ambient temperature and the reference ambient temperature is calculated. The difference is multiplied by a preset first correction value and summed with the original fin temperature to obtain the first calibration temperature. The first calibration temperature is used as the replacement value to update the infrared thermal image sequence; otherwise, the original infrared thermal image sequence is maintained, and the original fin temperature is determined as the current calibration temperature. In this embodiment, the reference ambient temperature is set to 25°C.

[0092] Build humidity calibration rules; if the ambient humidity exceeds the humidity threshold... If the water vapor in the air absorbs infrared radiation, causing the first calibration temperature to decrease, the first calibration temperature is added to a preset second correction value to obtain the second calibration temperature. The second calibration temperature is then used as the replacement value to update the infrared thermal image sequence, and the second calibration temperature is determined as the final calibration temperature. Otherwise, the infrared thermal image sequence after ambient temperature calibration is maintained, and the first calibration temperature is determined as the final calibration temperature.

[0093] The final calibration temperature of all sampling units is mapped to the pixels of the infrared thermal image sequence to obtain the calibration infrared thermal image sequence;

[0094] For each frame of calibration heatmap, the following is used: Convolution operations are performed using Gaussian filter kernels of varying sizes to smooth image noise, such as thermal image spots caused by power plant dust, resulting in an enhanced infrared thermal image sequence.

[0095] The Canny edge detection algorithm is used to extract the contour features of the fin edges from each frame of enhanced heat map, including contour integrity, break point density and fin spacing deviation, and the contour break points are marked.

[0096] The contour integrity ratio is obtained by calculating the ratio of the actual extracted contour length to the theoretical contour length; the fracture point density is obtained by calculating the ratio of the number of fracture points to the actual extracted contour length; the fin spacing is set according to the standard spacing of the cold finned tube heat exchanger in the power plant, and the difference between the measured actual spacing and the fin spacing is calculated. The difference is divided by the fin spacing to obtain the fin spacing deviation; the contour integrity ratio, fracture point density, and fin spacing deviation are normalized using Min-Max dimensionality to obtain the contour feature vector.

[0097] For each frame of enhanced heatmap, regions are divided according to sampling units. For each region, radial and tangential directions are added based on traditional angles, and gray-level co-occurrence matrix is ​​calculated at a preset offset distance. The traditional angles include 0°, 45°, 90°, and 135°. The radial direction is used to capture the thermal gradient texture along the fin length direction, and the tangential direction is used to capture the lateral texture differences between adjacent fins caused by uneven blockage.

[0098] Energy, entropy, and contrast are extracted from the gray-level co-occurrence matrix in each direction. The feature mean of the enhanced infrared thermal image sequence is calculated, and the feature mean is integrated into the texture feature vector of the thermal image.

[0099] Based on the fin blockage mechanism of power plants, contour feature weights and texture feature weights are set, the contour feature vector and texture feature vector are weighted and summed, and arranged in spatial coordinate order to generate an enhanced heat map feature vector set.

[0100] Specifically, the intelligent recognition module includes a map construction unit, a model construction unit, a level determination unit, and a missing region determination unit;

[0101] The graph construction unit is used to analyze the spatial distribution of the received enhanced thermal image feature vector set and the distribution pattern of the power plant finned tube bundles using the K-means spatial coordinate clustering algorithm to form a set of thermal image nodes with a topological structure. Based on the physical spacing, contact area and thermal conductivity coefficient of the tube bundles between nodes, the thermal resistance correlation formula is used to calculate the thermal correlation degree between adjacent fins. Combining the enhanced thermal image feature vectors of each node, the cosine similarity is used to calculate the infrared feature similarity between nodes. By weighted fusion of thermal correlation degree and infrared feature similarity, a dynamic edge-weighted infrared thermal image feature correlation graph is constructed.

[0102] The model building unit employs machine learning algorithms to construct an infrared intelligent judgment model based on a graph neural network (GCN). Using infrared thermal image feature correlation maps as input, it calculates the deviation between the final calibration temperature of a node and the normal operating temperature range of the fins. If the deviation exceeds a preset temperature difference threshold, it is judged as a temperature anomaly. A multi-layer message passing mechanism aggregates temperature anomaly information from neighboring nodes, learns the thermal accumulation patterns in local areas, and outputs the temperature anomaly probability for each node through a fully connected layer and softmax probability mapping, combined with the spatial diffusion pattern of the congested area. ;in, , The total number of nodes;

[0103] The level determination unit is used to receive the temperature anomaly probability of each node output by the infrared intelligent judgment model, construct the blockage level determination conditions, and divide the blockage risk level into high risk, medium risk and low risk based on the preset probability threshold; construct a historical clean state benchmark value table, construct a temperature deviation formula, calculate the pollution sensitivity of temperature anomaly probability and historical clean state benchmark value, construct the dirt level determination conditions, and divide the dirt level into first-level pollution, second-level pollution and third-level pollution based on the deviation threshold.

[0104] The omission area determination unit is used to introduce the local anomaly propagation coefficient. Based on the dynamic edge weight matrix in the infrared thermal image feature correlation spectrum, it calculates the comprehensive thermal conduction influence value of each low-anomaly node on the neighboring high-anomaly nodes. If the comprehensive thermal conduction influence value exceeds the preset propagation threshold and the node temperature gradient change rate deviates from the tube bundle heat dissipation curve, then the omission blockage and contamination area is determined. The thermal conductivity coefficient of the fin material is modeled, and the reverse thermal conduction simulation is performed using the finite element method combined with the actual detection value to obtain the simulated temperature. The residual analysis is performed by calculating the deviation between the simulated temperature and the actual detection value to verify the determination result.

[0105] Specifically, the steps for constructing an infrared thermal image feature correlation map include:

[0106] The spatial coordinates of the enhanced heatmap feature vector set are extracted as the basis for clustering, and the total number of finned tube bundles is counted to determine the number of clusters for K-means clustering. The K-means clustering algorithm is used to divide the enhanced feature vectors with similar spatial coordinates into the same cluster. Each cluster corresponds to a fin region. The adjacency relationship between clusters is determined according to the row and column distribution pattern of the fin bundle, forming a set of heat map nodes with topological structure.

[0107] Based on the topological relationships between nodes, such as physical proximity, all adjacent node pairs in the heatmap node set are traversed. The center coordinates of each pair of nodes are extracted to calculate the physical distance. The product of the contact area and the thermal conductivity coefficient is calculated. The ratio of the physical distance to the product is calculated using the thermal resistance correlation formula to obtain the thermal resistance. The maximum thermal resistance among all adjacent node pairs is filtered in descending order. The thermal resistance ratio between the current thermal resistance and the maximum thermal resistance is calculated. The thermal resistance ratio is subtracted from the base coefficient to obtain the thermal correlation degree. Here, the current thermal resistance represents the thermal resistance value calculated for the currently processed adjacent node pair during the calculation process of traversing the heatmap node set. In this embodiment, the base coefficient is set to 1.

[0108] Extract the mean feature vector of each node, including but not limited to the average temperature, standardize the mean feature vectors of adjacent node pairs, use the cosine similarity formula, and combine the standardized mean feature vectors of adjacent node pairs to calculate the infrared feature similarity of adjacent nodes.

[0109] Set thermal correlation threshold and infrared feature similarity threshold, and dynamically adjust the thermal correlation threshold and infrared feature similarity threshold based on the historical operating data of the power plant fins and the real-time detection environment. For example, reduce the thermal correlation threshold under high temperature environment and reduce the infrared feature similarity threshold when the dust concentration is high. Through weighted calculation, fuse the thermal correlation and infrared feature similarity to obtain a dynamic edge weight matrix.

[0110] Using heatmap nodes as vertices and dynamic edge weight matrices as edges, an adjacency list is used to store vertices and edges to construct an infrared heatmap feature association map with dynamic edge weights. Each vertex attribute includes node number, center coordinates, and mean feature vector, and the edge attribute includes real-time edge weight. Each edge connects adjacent nodes.

[0111] Specifically, the steps for constructing the infrared intelligent judgment model include:

[0112] Based on the Graph Neural Network (GCN) algorithm in machine learning, an infrared intelligent judgment model is constructed, including an input layer, a perception enhancement layer, a diffusion attention layer, and an output layer.

[0113] The input layer receives the infrared thermal image feature correlation spectrum, extracts the temperature gradient from the infrared thermal image feature correlation spectrum, calculates the deviation value between the final calibration temperature and the normal operating range of the fin for each node, and generates a list of temperature deviation values.

[0114] The perception enhancement layer iterates through the list of temperature deviation values. For each node, if the deviation value exceeds the preset temperature difference threshold, it is determined to be a temperature anomaly, the temperature anomaly node is marked, and an enhancement coefficient is introduced to enhance the feature values ​​of the temperature anomaly node, forming a perception enhancement feature matrix; otherwise, it is determined to be a normal temperature, and the feature values ​​remain unchanged. The enhancement coefficient is set based on the ratio of the deviation value of the temperature anomaly node to the preset temperature difference threshold. The larger the deviation value, the larger the enhancement coefficient, so as to enhance the degree of feature enhancement.

[0115] The diffusion attention layer is based on the perception enhancement feature matrix. It calculates the cosine similarity between each node and its neighboring nodes to obtain the diffusion correlation. For temperature anomaly nodes, the correlation weight of the neighboring nodes is increased to strengthen the transmission of anomaly diffusion. The initial attention weight matrix is ​​constructed by combining the diffusion correlation and the spatial diffusion attention weight matrix is ​​obtained by softmax normalization. The perception enhancement feature matrix and the spatial diffusion attention weight matrix are weighted and fused to obtain the diffusion perception feature matrix, which is used to capture spatial diffusion patterns.

[0116] The output layer performs dimensionality reduction on the diffusion-sensing feature matrix through a fully connected layer to form a comprehensive feature vector. This comprehensive feature vector is then used to calculate the initial temperature anomaly probability of each node using a sigmoid function. The average initial temperature anomaly probability of all neighboring nodes of each node is also calculated to obtain the neighborhood average temperature anomaly probability. For each node, if the neighborhood average temperature anomaly probability is lower than a preset average probability threshold, it is determined to be an isolated node; otherwise, it is determined to be a non-isolated node. The initial temperature anomaly probability of isolated nodes is corrected using the spatial diffusion attention weight matrix. For isolated nodes whose initial temperature anomaly probability exceeds the preset limit, the initial temperature anomaly probability is reduced to avoid misjudging isolated noise points. The initial temperature anomaly probability of non-isolated nodes remains unchanged. Finally, the temperature anomaly probability of each node is obtained and output. ;

[0117] The binary cross-entropy loss function is set to extract the node feature data and edge topology association data of the infrared heat map feature association map, and the node labels are marked to form a heat map sample set. The heat map sample set is divided into training set, validation set and test set according to a preset ratio. The infrared intelligent judgment model is trained by combining the training set as input and the Adam optimizer.

[0118] Specifically, the steps for determining the level of blockage risk and the level of contamination include:

[0119] Iterate through the temperature anomaly probabilities of all nodes and set a maximum probability threshold based on expert experience. With minimum probability threshold Congestion level determination criteria are constructed to classify congestion risk levels. For each node, if... This indicates a significant abnormality in node temperature, suggesting severe congestion, and classifies it as a high-risk candidate; if This indicates an abnormal node temperature and congestion, classifying it as a medium-risk candidate; if This indicates that the node temperature is fluctuating slightly, and it is directly judged as low risk;

[0120] For high-risk candidate nodes, the total number of high-risk and medium-risk candidate nodes among all adjacent nodes is counted. Combined with the historical characteristics of multiple nodes clustering in blocking, if the total number of nodes exceeds the preset first-level threshold, it is judged as high-risk; otherwise, it is downgraded to medium-risk to avoid isolated nodes being misjudged as high-risk.

[0121] For nodes that are candidates for medium risk, the total number of medium-risk candidate nodes among all adjacent nodes is counted. If the total number of medium-risk candidate nodes exceeds the preset secondary threshold, it is determined to be medium risk; otherwise, it is downgraded to low risk to avoid isolated nodes being mistakenly judged as medium risk.

[0122] Temperature data of the power plant's finned tube bundles under clean and normal operating conditions were collected over f cycles using an infrared thermal imager. Statistical analysis was employed to calculate the average temperature at each monitoring node, and a historical cleanliness baseline table was constructed based on the normal fluctuation range. For each node, the probability of temperature anomalies was analyzed. Subtract the historical cleaning status baseline value from the historical cleaning status baseline value table. The difference is obtained, and a deviation coefficient is introduced. Historical cleanliness baseline values coefficient of variation Summing the values ​​yields the baseline compensation value. The ratio of the difference to the baseline compensation value is then calculated to determine the pollution sensitivity of each node based on its current temperature anomaly probability and historical cleanliness baseline value. This ultimately leads to a temperature deviation formula; in this embodiment, f is set... The temperature deviation formula is as follows:

[0123]

[0124] like ,but This indicates that the node is uncontaminated; if ,but , A higher value indicates more severe node contamination;

[0125] Integrating the contamination sensitivity of all nodes based on the temperature deviation formula This forms a list of pollution sensitivities.

[0126] Iterate through each node in the pollution sensitivity list and set the maximum deviation threshold based on the power plant's historical operation and maintenance data. With the lowest deviation threshold The criteria for determining the level of dirtiness are constructed and classified into three levels. This indicates that the degree of node fouling has led to a severe decrease in heat exchange efficiency, and is classified as Level 1 fouling; if This indicates significant fouling at the node, resulting in a moderate decrease in heat exchange efficiency, and is classified as secondary fouling; if If the node is slightly dirty and the heat exchange efficiency is slightly reduced, it is judged as a level three pollution.

[0127] Specifically, the steps for determining areas that have been missed or blocked by dirt include:

[0128] Based on the dynamic edge weight matrix in the infrared thermal image feature correlation map, the temperature anomaly probability of each node is extracted, and a baseline threshold is set. ,like If the condition is met, it is determined to be a high-anomaly node; otherwise, it is determined to be a low-anomaly node. Among them, high-anomaly nodes are potential blockage nodes that have a thermal impact on the neighborhood.

[0129] Filter out neighboring nodes whose edge weights exceed a preset weight threshold from the dynamic edge weight matrix to form a sequence of highly abnormal neighboring nodes;

[0130] A local anomaly propagation coefficient is introduced. For each low-anomaly node, the sequence of neighboring high-anomaly nodes is traversed. Combining the anomaly degree coefficient of the high-anomaly nodes, the initial thermal conduction influence value of the low-anomaly node on the neighboring high-anomaly nodes is calculated. The initial thermal conduction influence values ​​of all neighboring high-anomaly nodes on the low-anomaly node are accumulated and normalized to eliminate the magnitude deviation of the influence value caused by the difference in the number of neighbors of different nodes, so as to obtain the comprehensive thermal conduction influence value of each low-anomaly node. Among them, the anomaly degree coefficient is obtained by normalizing the initial temperature anomaly probability of the high-anomaly node, reflecting the relative intensity of the node anomaly.

[0131] If the combined thermal conduction influence value exceeds the preset propagation threshold, it is determined to be a node to be verified; otherwise, it is determined to be a node not to be verified.

[0132] Under clean and normal operating conditions of the power plant equipment, historical time-series temperature data is collected. The temperature change over time is fitted using nonlinear regression to construct a tube bundle heat dissipation curve. For each node to be verified, the temperature gradient change rate is calculated using the time-series temperature data. The deviation rate between the temperature gradient change rate and the tube bundle heat dissipation curve is compared. If the deviation rate is greater than a preset deviation rate threshold, it indicates that the temperature gradient of the node to be verified deviates from the normal heat dissipation pattern and is identified as a suspected missing area node. Otherwise, it indicates that the temperature gradient of the node to be verified conforms to the normal heat dissipation pattern and is identified as a non-suspected missing area node.

[0133] For each suspected missing node, a reverse heat conduction model is used, with the actual temperature of the neighboring high-anomaly nodes as the heat source boundary. The thermal conductivity coefficient of the fin material, the ambient temperature, and the heat dissipation rate are input to simulate the theoretical normal temperature of the suspected missing node. The residual is calculated by combining the actual detection temperature of the suspected missing node.

[0134] Set a residual threshold. If the residual is greater than the residual threshold, it is determined that the suspected missing area node is blocked or dirty; otherwise, it is determined that the suspected missing area node is a false anomaly caused by environmental fluctuations, and the suspected missing area node with false anomaly is removed.

[0135] Integrate all suspected missing area nodes with blockages and dirt to form a missing blockage and dirt area.

[0136] Specifically, the steps for generating preliminary verification results include:

[0137] For each monitoring area, the high-risk, medium-risk, and low-risk levels of congestion risk are mapped to quantitative risk values, respectively. , , The pollution levels of Level 1, Level 2, and Level 3 are mapped to quantitative pollution values, respectively. , , ;in, ,and ;

[0138] Risk weights and pollution weights are set based on expert experience, and the comprehensive evaluation coefficient for each monitoring area is calculated by combining the risk quantification value and the pollution quantification value.

[0139] The historical comprehensive evaluation coefficient of the monitoring area is retrieved, and the absolute difference between the coefficient and the comprehensive evaluation coefficient is calculated. If the absolute difference is lower than the preset difference threshold, the comprehensive evaluation coefficients are determined to be similar; otherwise, the comprehensive evaluation coefficients are determined to be dissimilar.

[0140] By collecting historical anomaly diagnosis case data of power plant finned tube bundles, such as spatial coordinates of anomaly areas and maintenance records, a historical diagnosis database is constructed. Candidate cases with similar comprehensive evaluation coefficients are selected from the historical diagnosis database, and the spatial coordinates of the anomaly areas in each candidate case are extracted. The Euclidean distance is calculated by combining the spatial coordinates of the current monitoring area. If the Euclidean distance is lower than the preset distance threshold, the space is determined to be similar; otherwise, the space is determined to be dissimilar.

[0141] Historical cases that simultaneously meet the criteria of similarity in comprehensive evaluation coefficient and spatial similarity are selected to form a similar case set. The total number of cases in the similar case set, the number of cases ultimately determined to be blocked, and the number of false positive cases that were initially determined to be abnormal but were ultimately confirmed to be free of blockage are counted. The blockage ratio is obtained by calculating the ratio of the number of cases ultimately determined to be blocked to the total number of cases, and the false positive ratio is obtained by calculating the ratio of the number of false positive cases to the total number of cases.

[0142] If the proportion of congestion exceeds the proportion of false alarms, the current monitoring area is determined to have a high historical matching congestion tendency and is marked as having a high historical matching congestion tendency; otherwise, the current monitoring area is determined to have a high historical matching false alarm tendency and is marked as having a high historical matching false alarm tendency, thus forming a historical matching tendency label table.

[0143] An expert rule base is established, including: if the monitored area is high-risk and its adjacent nodes are also high-risk, it is prioritized as a severely abnormal area; if the monitored area is an isolated node and is high-risk, it is determined as a suspected false alarm area; if the monitored area has a pollution level of level three and a risk level of low risk, it is determined as an abnormal area to be observed.

[0144] Based on the historical matching tendency label table, expert rules are matched one by one to determine the preliminary verification categories as severely abnormal areas, abnormal areas to be observed, and suspected false alarm areas. Since suspected false alarm areas have noise interference in historical data and lack spatial clustering and persistence characteristics, suspected false alarm areas are filtered out, and the preliminary verification results are finally generated.

[0145] Specifically, the steps to generate diagnostic optimization results include:

[0146] For the severely abnormal areas and the areas to be observed in the preliminary verification results, the actual blockage area of ​​each area is calculated using spatial coordinate range. The actual blockage area is then divided by the total heat dissipation area of ​​the fins to obtain the blockage area percentage of each area. ;

[0147] The average temperature is extracted from each contaminated area. The historical normal operating temperature is retrieved from the database, and the normal reference temperature is extracted. The absolute difference between the average temperature and the normal reference temperature is calculated to obtain the temperature deviation for each contaminated area. ;

[0148] Set a threshold for the percentage of blocked area. , and Based on the safety operation specifications of power plant finned tube bundles, a blockage early warning constraint is constructed. If it is, it is judged as a Level 1 congestion warning; if If it is, it is judged as a level 2 congestion warning; if If so, it is determined to be a Level 3 congestion warning;

[0149] Set temperature deviation threshold , and Based on the thermal performance of the equipment, a temperature early warning constraint is constructed. If so, it is determined to be a Level 1 temperature warning; if If so, it is determined to be a Level II temperature warning; if If so, it is determined to be a Level III temperature warning;

[0150] If both a Level 1 congestion warning and a Level 1 temperature warning are met simultaneously, it is determined to be a Level 1 comprehensive warning; if both a Level 3 congestion warning and a Level 3 temperature warning are met simultaneously, it is determined to be a Level 3 comprehensive warning; otherwise, it is determined to be a Level 2 comprehensive warning.

[0151] Based on the comprehensive early warning level, a reinforcement learning timing optimization mechanism is triggered, using the comprehensive early warning level, the proportion of congested area, temperature deviation, and real-time operating parameters as state features to form a state space for reinforcement learning; among which, real-time operating parameters include, but are not limited to, load.

[0152] The action space executes diagnostic strategies to adjust actions, such as optimizing the similarity threshold for matching historical data;

[0153] Define a reward function: if the actual fault matches the warning, such as a severe congestion occurring after a Level 1 comprehensive warning, a positive reward is given; if a false alarm or missed alarm occurs, a negative reward is given.

[0154] By retrieving historical comprehensive early warning cases and combining them with actual fault time-series data, the DQN algorithm is used to select the optimal action in the state space. Through multiple rounds of iterative learning, the optimal diagnostic optimization strategy for different comprehensive early warning levels and operating states is obtained, such as increasing the similarity threshold for historical data matching.

[0155] Based on the optimal diagnostic optimization strategy output by reinforcement learning, the judgment logic of congestion warning and temperature warning is dynamically adjusted. For example, a compensation coefficient of real-time load is introduced to increase the temperature deviation threshold, forming a diagnostic optimization result, including but not limited to the spatial distribution of congestion area and the optimized warning level. Among them, the various thresholds of the present invention are set by those skilled in the art.

[0156] In summary, this invention constructs an intelligent diagnostic system for finned tube heat exchanger blockage in power plant indirect air-cooled systems by integrating four modules: fin temperature detection, data processing, intelligent identification, and diagnostic optimization. This system enables intelligent diagnosis and optimization of blockage and fouling anomalies in the finned tube heat exchanger. The temperature detection module uses an infrared thermal imager deployed at the end of a flexible robot to perform continuous gridded scanning of the power plant indirect air-cooled fin surface, acquiring real-time infrared thermal image sequences. Dynamic calibration is performed using ambient temperature and humidity parameters, and Gaussian filtering for noise reduction and Canny edge detection are used to extract fin contour features. The system then integrates thermal image texture features calculated using the gray-level co-occurrence matrix to generate an enhanced thermal image feature vector set. The intelligent identification module, based on K-means spatial clustering and the distribution patterns of finned tube bundles, forms a set of thermal image nodes with a topological structure. This is combined with a weighted fusion of thermal correlation and infrared feature similarity to construct a dynamically weighted infrared thermal image feature correlation map. Finally, a graph neural network (GCN) is used to construct an infrared intelligent judgment model, aggregating neighborhood anomaly information through a multi-layer message passing mechanism. The system outputs the probability of temperature anomalies at each node and, combined with probability thresholds and historical cleanliness baselines, classifies blockage risk levels and contamination levels. It also introduces a local anomaly propagation coefficient and reverse heat conduction simulation to identify any missed blockage or contamination areas. The diagnostic optimization module, based on a multi-source data verification mechanism, maps blockage risk levels and contamination levels to quantifiable values. It then uses a historical diagnostic database for comprehensive evaluation coefficient matching and spatial distribution similarity comparison, filters similar cases, and statistically analyzes the blockage and false alarm rates. Logical verification is performed using an expert rule base to eliminate isolated noise and suspected false alarm areas, generating preliminary verification results. A dual early warning constraint is constructed to determine the comprehensive early warning level and triggers a reinforcement learning temporal optimization mechanism to dynamically adjust the diagnostic strategy and generate optimized diagnostic results. This invention achieves high-precision, intelligent, and full-process diagnosis of blockage status in indirect cooling fins, improving the operational safety and maintenance efficiency of power plant air-cooled systems.

[0157] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An infrared temperature measurement intelligent diagnostic system for intercooled fin blockage, characterized in that, include: Temperature detection module, data processing module, intelligent recognition module, diagnostic optimization module; The temperature detection module is used to acquire infrared thermal image sequences in real time; The data processing module is used to construct environmental temperature calibration rules and humidity calibration rules, calibrate infrared thermal image sequences, extract contour features to generate contour feature vectors, calculate gray-level co-occurrence matrix to generate texture feature vectors, and weighted fuse to generate enhanced thermal image feature vector sets. The intelligent recognition module is used to calculate the thermal correlation degree and infrared feature similarity to obtain a dynamic edge weight matrix, and to construct an infrared thermal image feature correlation map. An infrared intelligent judgment model is constructed to output the probability of temperature anomalies and determine the level of blockage risk and dirt level; a local anomaly propagation coefficient is introduced and combined with reverse heat conduction simulation to determine the areas of missed blockage and dirt. The diagnostic optimization module is used to determine the comprehensive evaluation coefficient and spatial similarity, exclude suspected false alarm areas, and generate preliminary verification results; it constructs congestion warning constraints and temperature warning constraints to determine the comprehensive warning level, and dynamically adjusts the judgment logic to form diagnostic optimization results.

2. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 1, characterized in that, The specific steps for generating the enhanced heatmap feature vector set include: Establish an ambient temperature calibration rule: if the actual ambient temperature is higher than the reference ambient temperature, perform temperature calibration and update the infrared thermal image sequence with the first calibration temperature as the replacement value; otherwise, maintain the original infrared thermal image sequence. A humidity calibration rule is established: if the ambient humidity exceeds the preset humidity threshold, the second calibration temperature is used as the replacement value to update the infrared thermal image sequence; otherwise, the infrared thermal image sequence after ambient temperature calibration is maintained. Extract the contour features of the fin edge, mark the contour break points, calculate the contour integrity ratio, and calculate the break point density. Set the fin spacing, calculate the fin spacing deviation, and combine the contour integrity ratio and the break point density to generate a contour feature vector; By combining the radial and tangential directions, the gray-level co-occurrence matrix is ​​calculated, and texture feature vectors are generated; By weighted fusion, an enhanced heatmap feature vector set is generated.

3. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 2, characterized in that, The intelligent recognition module includes a map construction unit, a model construction unit, a level determination unit, and a missing region determination unit; The graph construction unit is used to form a set of heat map nodes through K-means clustering algorithm, calculate the thermal correlation between adjacent fins and the infrared feature similarity between nodes to obtain a dynamic edge weight matrix, and construct an infrared heat map feature correlation graph. The model building unit is used to build an infrared intelligent judgment model based on graph neural network. It inputs the infrared thermal image feature association spectrum and, combined with a multi-layer message passing mechanism, outputs the probability of temperature anomalies at each node. The level determination unit is used to construct blockage level determination conditions to determine the blockage risk level; construct a historical clean status benchmark value table, calculate the pollution sensitivity in combination with the temperature deviation formula, and construct dirt level determination conditions to determine the dirt level. The omission area determination unit is used to introduce the local abnormal propagation coefficient, calculate the comprehensive thermal conduction influence value, determine the omission, blockage and dirt areas, and verify the determination results by combining reverse thermal conduction simulation.

4. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 3, characterized in that, The specific steps for constructing an infrared thermal image feature correlation map include: The spatial coordinates of the enhanced heatmap feature vector set are extracted, and the K-means clustering algorithm is used to divide the clusters and determine the adjacency relationship between the clusters to form a heatmap node set with a topological structure. Based on the node topology, thermal resistance is calculated by combining physical spacing, contact area and thermal conductivity coefficient. The maximum thermal resistance of all adjacent node pairs is screened in descending order and the thermal resistance ratio with the current thermal resistance is calculated. The thermal correlation degree is calculated by combining the basic coefficient. Extract the mean feature vector of each node, and combine it with the standardized mean feature vector of adjacent nodes to calculate the infrared feature similarity. Set and dynamically adjust the thermal correlation threshold and the infrared feature similarity threshold, and obtain a dynamic edge weight matrix through weighted calculation and fusion. Using heatmap nodes as vertices and the dynamic edge weight matrix as edges, an infrared heatmap feature association map with dynamic edge weights is constructed.

5. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 4, characterized in that, The specific steps for constructing an infrared intelligent judgment model include: An infrared intelligent judgment model is constructed based on graph neural networks, including an input layer, a perception enhancement layer, a diffusion attention layer, and an output layer; The input layer extracts the temperature gradient, calculates the deviation value of each node, and generates a list of temperature deviation values. The perception enhancement layer determines temperature anomalies based on a preset temperature difference threshold, introduces enhancement coefficients to strengthen features for temperature anomaly nodes, and forms a perception enhancement feature matrix. The diffusion attention layer calculates the diffusion correlation degree between each node and its neighboring nodes, constructs a spatial diffusion attention weight matrix, and fuses it with the perception enhancement feature matrix in a weighted manner to generate a diffusion perception feature matrix; The output layer calculates the initial temperature anomaly probability and combines it with the neighborhood average temperature anomaly probability to determine isolated nodes and non-isolated nodes. It reduces the initial temperature anomaly probability of isolated nodes that exceed a preset limit and outputs the temperature anomaly probability of each node. ; Using binary cross-entropy as the loss function, a heat map sample set is formed based on the infrared heat map feature association spectrum, and the set is divided into a training set, a validation set, and a test set. The infrared intelligent judgment model is then trained using the Adam optimizer.

6. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 5, characterized in that, The specific steps for determining the level of blockage risk and the level of dirt include: Set the maximum probability threshold With minimum probability threshold ; Construct congestion level determination criteria, if If so, then it is determined to be a high-risk candidate; if If so, then it is determined to be a medium-risk candidate; if If so, it is directly judged as low risk; For high-risk candidate nodes, count the total number of high-risk and medium-risk candidate nodes among all adjacent nodes. If the total number exceeds the preset first-level threshold, it is judged as high-risk; otherwise, it is downgraded to medium-risk. For medium-risk candidate nodes, count the total number of medium-risk candidate nodes among all adjacent nodes. If the total number exceeds the preset secondary threshold, the node is classified as medium-risk; otherwise, it is downgraded to low-risk.

7. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 6, characterized in that, The specific steps for determining the level of blockage risk and the level of dirt also include: Construct a historical cleanliness baseline table and calculate the probability of temperature anomalies. Compared with historical cleanliness baseline values Difference; A deviation coefficient is introduced, and a baseline compensation value is obtained by combining the historical cleanliness baseline value. The pollution sensitivity of each node is determined by the ratio of the difference to the baseline compensation value. This ultimately leads to the temperature deviation formula; like ,but ;otherwise ; Set the maximum deviation threshold With the lowest deviation threshold ; Construct conditions for determining the level of dirtiness, if If so, it is classified as Level 1 pollution; if If so, it is classified as Level II pollution; if If so, it is classified as Level 3 pollution.

8. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 7, characterized in that, The specific steps for determining missed, clogged, or dirty areas include: Extract the probability of temperature anomalies at each node and set a baseline threshold. ,like If the condition is met, then the node is considered a high-abnormal node; otherwise, it is considered a low-abnormal node. By introducing a local anomaly propagation coefficient, and combining the anomaly degree coefficient of high anomaly nodes, the comprehensive thermal conduction impact value is calculated for each low anomaly node. If the combined thermal conduction influence value exceeds the preset propagation threshold, it is determined to be a node to be verified. The temperature gradient change rate is calculated, and the deviation rate between the temperature gradient change rate and the tube bundle heat dissipation curve is compared. If the deviation rate is greater than the preset deviation rate threshold, it is determined to be a suspected missing area node. The reverse heat conduction model is used to simulate the theoretical normal temperature, and the residual is calculated in combination with the actual detection temperature. If the residual is greater than the preset residual threshold, it is determined that the suspected missing area node is blocked or dirty. All suspected missing area nodes with blockage or dirt are integrated to form a missing blocked or dirty area.

9. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 8, characterized in that, The specific steps for generating preliminary verification results include: The blockage risk level is mapped to a risk quantification value, and the dirt level is mapped to a pollution quantification value. The comprehensive evaluation coefficient for each monitoring area is calculated by weighting the values. Use the historical comprehensive evaluation coefficients of the monitored area to calculate the absolute difference between the coefficients and the historical comprehensive evaluation coefficients. If the absolute difference is lower than the preset difference threshold, the comprehensive evaluation coefficients are considered similar; otherwise, they are considered dissimilar. Similar candidate cases are selected from the historical diagnostic database, and the Euclidean distance is calculated. If the Euclidean distance is lower than a preset distance threshold, the spaces are considered similar; otherwise, they are considered dissimilar. Screen similar candidate cases that simultaneously meet the requirements of comprehensive evaluation coefficient similarity and spatial similarity to form a similar case set, calculate the blocking ratio and false alarm ratio, and form a historical matching tendency label table; An expert rule base is constructed. Based on the matching rules of the historical matching tendency label table, the preliminary verification categories are determined as severely abnormal areas, abnormal areas to be observed, and suspected false alarm areas. Suspected false alarm areas are filtered out, and preliminary verification results are generated.

10. The intelligent diagnostic system for intercooled fin blockage using infrared temperature measurement according to claim 9, characterized in that, The specific steps to generate diagnostic optimization results include: For severely abnormal areas and areas to be observed for abnormality, calculate the proportion of blocked area. Temperature deviation from dirty areas ; Set a threshold for the percentage of blocked area. , ; Construct congestion early warning constraints, if If so, a Level 1 congestion warning is issued; if If so, a level 2 congestion warning is issued; if If so, a Level 3 congestion warning is issued; Set temperature deviation threshold , ; Construct temperature early warning constraints, if If so, a Level 1 temperature warning is issued; if If so, a level two temperature warning is issued; if If so, a Level III temperature warning will be issued; If both the Level 1 congestion warning and the Level 1 temperature warning are met simultaneously, it is determined to be a Level 1 comprehensive warning; If both a Level 3 congestion warning and a Level 3 temperature warning are met simultaneously, it is determined to be a Level 3 comprehensive warning; Otherwise, it will be classified as a Level II comprehensive warning; Based on the comprehensive early warning level, and combined with reinforcement learning to output the optimal diagnostic optimization strategy, the judgment logic of congestion warning and temperature warning is dynamically adjusted to form a diagnostic optimization result.

Images