Desulfurization equipment temperature abnormity diagnosis method and system using thermal imaging technology

Abstract

The invention provides a desulfurization equipment temperature anomaly diagnosis method and system using a thermal imaging technology, and relates to the technical field of temperature monitoring, and the method comprises the steps: obtaining thermal imaging data, and carrying out the normalization; temperature distribution characteristics are extracted, working condition parameters are obtained, and a mapping relation is established through position time sequence coding; modeling and calculating an abnormal metric value by using a variational auto-encoder; determining an abnormal source and a propagation path based on the functional region directed graph; and fusing the abnormal metric value and the abnormal propagation information to output a diagnosis result. According to the method, temperature abnormity can be accurately identified, an abnormal source and a propagation rule can be accurately positioned, and the fault prediction accuracy of the desulfurization equipment is improved.

Description

Technical Field

[0001] This invention relates to the field of temperature monitoring technology, and in particular to a method and system for diagnosing temperature anomalies in desulfurization equipment using thermal imaging technology. Background Technology

[0002] As a key component of the environmental protection system in thermal power plants, the operating status of desulfurization equipment directly affects environmental emission standards and the safe and stable operation of generator units. Temperature, as an important parameter reflecting the health status of desulfurization equipment, often indicates potential malfunctions such as scaling, corrosion, and blockage due to abnormal temperature changes. Traditional temperature monitoring methods primarily rely on installing fixed temperature sensors on the equipment surface to determine equipment status through point-based temperature data. With the development of thermal imaging technology, a non-contact, large-area, and high-precision temperature field monitoring method has been provided, capable of comprehensively capturing the temperature distribution on the surface of desulfurization equipment, offering a new technical approach for the early detection of equipment anomalies.

[0003] Current methods for diagnosing temperature anomalies in desulfurization equipment using thermal imaging technology have the following shortcomings: The temperature distribution of desulfurization equipment is closely related to its operating parameters, and existing diagnostic methods often overlook this coupling relationship, leading to incorrect identification of temperature changes under different operating conditions as anomalies. Most existing technologies rely on simple threshold judgments or statistical models for anomaly detection, lacking the ability to extract and model deep features from temperature distribution data, thus failing to effectively capture complex temperature anomaly patterns. Furthermore, physical connections and heat conduction exist between the functional areas of the desulfurization system, meaning temperature anomalies may propagate from one area to adjacent areas. Existing methods struggle to trace the source of the anomaly and analyze its propagation path, resulting in inaccurate and incomplete diagnostic results, failing to provide maintenance personnel with effective fault location and handling suggestions. Summary of the Invention

[0004] This invention provides a method and system for diagnosing temperature anomalies in desulfurization equipment using thermal imaging technology, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for diagnosing temperature anomalies in desulfurization equipment using thermal imaging technology, comprising: Acquire thermal imaging data of the desulfurization equipment and perform normalization processing to obtain standard thermal imaging data; The temperature distribution features of the standard thermal imaging data are extracted using multi-level feature processing, and the real-time operating parameters of the desulfurization equipment are obtained at the same time. The temperature distribution features and the real-time operating parameters are encoded by a position timing encoder, and a mapping relationship between temperature distribution and operating status is established in the semantic representation space to generate operating condition coupled temperature features. The temperature features coupled with operating conditions in historical normal operation data are input into a variational autoencoder for encoding and modeling. The temperature baseline probability distribution is obtained by modulation based on real-time operating condition parameters. The probability divergence between the latent variable distribution corresponding to the current temperature feature and the temperature baseline probability distribution is calculated to obtain the anomaly measurement value. Based on the physical connection relationship and heat conduction path of each functional area of ​​the desulfurization equipment, a directed graph of the functional area is constructed, and the abnormal measurement value is mapped to the graph node feature. The node role type is determined based on the abnormal measurement relationship between upstream and downstream graph nodes, and the abnormal source area and abnormal propagation path are determined through multi-hop message passing of graph convolutional network. By integrating the abnormal measurement value with the abnormal source region and abnormal propagation path, the temperature abnormality diagnosis result is output.

[0006] In one optional embodiment, extracting the temperature distribution features of the standard thermal imaging data using multi-level feature processing includes: Adaptive downsampling is performed on the standard thermal imaging data to establish multiple feature extraction levels at different scales. In each feature extraction level, local temperature change information is extracted by a temperature gradient operator, and regional temperature distribution information is extracted by nonlinear transformation. The temperature change information and the temperature distribution information are combined to obtain a hierarchical feature representation. The initial multi-level temperature features are determined by a progressive approach. Based on the initial multi-level temperature features, feature transmission paths are established between different levels and information is exchanged to obtain hierarchically associated temperature features. The hierarchical temperature features are decomposed into temperature gradient features and temperature distribution features. Based on the spatial distribution and temporal variation of the temperature features, the weighting factors of each feature are calculated to generate a feature weighting mapping table. The temperature gradient features and temperature distribution features are then combined in a weighted manner to form the optimized temperature distribution features.

[0007] In one optional embodiment, the temperature distribution features and the real-time operating parameters are encoded separately using a position-time encoder to establish a mapping relationship between temperature distribution and operating state in the semantic representation space, generating operating condition-coupled temperature features including: The temperature distribution features are spatially encoded to mark the location information of the device functional area corresponding to each pixel in the thermal imaging image, thereby generating location-aware temperature features. The real-time operating parameters are time-series encoded to extract dynamic trend information of operating condition changes and generate time-series-aware operating condition features. By using bidirectional projection transformation, the location-aware temperature features and the time-aware operating condition features are mapped to a unified semantic representation space. In the semantic representation space, the similarity matrix between the temperature feature vector and the operating condition feature vector is calculated to determine the temperature region distribution affected by the operating condition and the combination of operating condition parameters that affect the temperature distribution. Based on the temperature region distribution and the combination of operating parameters, by backtracking historical data, the normal temperature distribution range of the corresponding temperature region under different operating conditions is extracted. The temperature characteristics sensed at the current location are compared with the normal temperature distribution range to obtain the temperature deviation characteristics that eliminate the influence of operating conditions. The temperature deviation feature is spliced ​​and fused with the time-series-aware operating condition feature to generate an operating condition-coupled temperature feature.

[0008] In one optional embodiment, the temperature features coupled with operating conditions in historical normal operation data are input into a variational autoencoder for encoding modeling. A temperature baseline probability distribution is obtained by modulating real-time operating condition parameters. The probability divergence between the latent variable distribution corresponding to the current temperature feature and the temperature baseline probability distribution is calculated to obtain anomaly metrics, including: The temperature features coupled with the operating conditions in the historical normal operation data are input into the variational autoencoder for encoding, and the latent variable distribution is obtained in the latent space. The mean vector and variance vector of the latent variable distribution are extracted to obtain the latent space distribution benchmark. For the real-time operating parameters of the desulfurization equipment, the response changes of the mean vector and variance vector in the implicit space distribution benchmark are calculated when the values ​​of each operating parameter change. A mapping relationship between each operating parameter and the response changes of the implicit space distribution benchmark is established to obtain the operating condition-implicit space response mapping table. Extract the values ​​of the current real-time operating condition parameters, input them into the operating condition-latent space response mapping table, and query the offset of the mean vector and the scaling factor of the variance vector. In the latent space distribution benchmark, superimpose the offset on the mean vector and multiply the variance vector by the scaling factor to obtain the distribution of the modulation latent variable. The modulation latent variable distribution is determined as a temperature reference probability distribution that dynamically changes with operating conditions; The temperature characteristics coupled with the operating conditions at the current moment are obtained, encoded into the latent space to obtain the current latent variable distribution, and the probability divergence between the current latent variable distribution and the temperature baseline probability distribution is calculated. The probability divergence is used as a deviation measure to obtain an anomaly measure value.

[0009] In an optional embodiment, for the real-time operating parameters of the desulfurization equipment, the response changes of the mean vector and variance vector in the implicit space distribution benchmark are calculated when the values ​​of each operating parameter change, and a mapping relationship between each operating parameter and the response changes of the implicit space distribution benchmark is established, resulting in an operating condition-implicit space response mapping table including: The real-time operating parameters of the desulfurization equipment are divided into multiple subsets of operating parameters according to the physical action mechanism. The operating parameters in each subset are combined in pairs to obtain the operating parameter combination pairs. For each pair of operating condition parameters, extract sample data of synchronous changes of the pair of operating condition parameters from historical normal operation data, calculate the joint response change of the mean vector and variance vector in the latent space distribution benchmark when the pair of operating condition parameters changes jointly, and identify the coupling effect type between the pair of operating condition parameters. A synergistic enhancement coefficient and a mutual inhibition coefficient are preset, and the corresponding coupling coefficient is selected according to the coupling interaction type. When the coupling interaction type is synergistic enhancement, the synergistic enhancement coefficient is selected; when the coupling interaction type is mutual inhibition, the mutual inhibition coefficient is selected. The individual response changes of each operating parameter in the operating parameter combination pair are summed, and the summation result is multiplied by the coupling coefficient to obtain the joint response correction amount of the operating parameter combination pair. Calculate the independent response change of a single operating condition parameter and the joint response correction of a combination of operating condition parameters, and store the independent response change and the joint response correction in the operating condition-latent space response mapping table to establish a mapping relationship that includes single-parameter response and multi-parameter coupled response.

[0010] In one optional embodiment, a directed graph of functional regions is constructed based on the physical connection relationships and heat conduction paths of each functional region of the desulfurization equipment. Anomaly metrics are mapped to graph node features. Node role types are determined based on the anomaly metric relationships between upstream and downstream graph nodes. The anomaly source region and anomaly propagation path are determined through multi-hop message passing in a graph convolutional network, including: Obtain the physical connection relationship and heat conduction path between the functional areas of the desulfurization equipment, construct a directed graph of the functional areas according to the heat transfer direction, and map the abnormal metric values ​​of each functional area to the initial features of the corresponding graph nodes; For each graph node in the directed graph of the functional area, the abnormal metric value of the upstream neighboring node is extracted by tracing back along the directed edge to determine the upstream abnormal input value, and the abnormal metric value of the downstream neighboring node is extracted by extending forward to determine the downstream abnormal received value. Calculate the anomaly measurement relationship between the current graph node and the upstream and downstream graph nodes to obtain the anomaly source intensity and anomaly propagation attenuation rate. Based on the numerical relationship, determine the anomaly generation type node, the anomaly propagation type node, and the anomaly termination type node, and construct the anomaly propagation feature vector. The anomaly propagation feature vector is input into a graph convolutional network for multi-hop message passing. Anomaly-generating nodes are identified as potential anomaly source nodes, and directed edges connecting anomaly-terminating nodes are identified as anomaly propagation blocking edges. The region corresponding to the node with the highest anomaly intensity and no upstream anomaly input among the potential anomaly source nodes is identified as the anomaly source region. Starting from the graph node corresponding to the anomaly source region, traverse the directed edges that are not marked as anomaly propagation blocking edges along the positive direction of the directed edges to determine the anomaly propagation path.

[0011] In one optional embodiment, the anomaly measurement relationship between the current graph node and its upstream and downstream graph nodes is calculated to obtain the anomaly source intensity and anomaly propagation attenuation rate. Based on the numerical relationship, anomaly-generating nodes, anomaly-propagating nodes, and anomaly-terminating nodes are determined, and an anomaly propagation feature vector is constructed, including: The difference between the current graph node's anomaly metric value and the upstream anomaly input value is calculated to obtain the upstream anomaly increment. The ratio between the upstream anomaly increment and the current graph node's anomaly metric value is calculated to obtain the anomaly source intensity. The anomaly source intensity represents the proportion of the current graph node's autonomous contribution to the anomaly relative to the upstream input. The difference between the current graph node's anomaly metric value and the downstream anomaly received value is calculated to obtain the downstream anomaly attenuation. The ratio between the downstream anomaly attenuation and the current graph node's anomaly metric value is calculated to obtain the anomaly conduction attenuation rate. The anomaly conduction attenuation rate represents the proportion of energy loss as the anomaly propagates downstream. Based on the numerical relationship between the abnormal source intensity and the abnormal conduction attenuation rate, the role type of the current graph node is determined, including: when the abnormal source intensity is greater than the abnormal conduction attenuation rate, it is marked as an abnormal generation node; when the abnormal source intensity is less than the abnormal conduction attenuation rate, it is marked as an abnormal conduction node; when both the abnormal source intensity and the abnormal conduction attenuation rate are lower than the preset role determination threshold, it is marked as an abnormal termination node. The abnormal source intensity, abnormal propagation attenuation rate, and role type are combined to form an abnormal propagation feature vector.

[0012] A second aspect of the present invention provides a temperature anomaly diagnosis system for desulfurization equipment utilizing thermal imaging technology, comprising: The data preprocessing module is used to acquire thermal imaging data from desulfurization equipment and perform normalization processing to obtain standard thermal imaging data. The feature encoding module is used to extract the temperature distribution features of the standard thermal imaging data using multi-level feature processing, and at the same time obtain the real-time operating parameters of the desulfurization equipment. The temperature distribution features and the real-time operating parameters are encoded by the position timing encoder, and a mapping relationship between temperature distribution and operating status is established in the semantic representation space to generate operating condition coupled temperature features. The modeling and evaluation module is used to input the temperature features coupled with the operating conditions in the historical normal operation data into the variational autoencoder for encoding and modeling. It obtains the temperature reference probability distribution based on the real-time operating condition parameters, calculates the probability divergence between the latent variable distribution corresponding to the current temperature feature and the temperature reference probability distribution, and obtains the anomaly measurement value. The graph propagation analysis module is used to construct a directed graph of functional areas based on the physical connection relationship and heat conduction path of each functional area of ​​the desulfurization equipment, map abnormal measurement values ​​to graph node features, determine the node role type based on the abnormal measurement relationship between upstream and downstream graph nodes, and determine the abnormal source area and abnormal propagation path through multi-hop message passing of graph convolutional network. The result output module is used to integrate the abnormal measurement value with the abnormal source region and abnormal propagation path to output the temperature abnormality diagnosis result.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this embodiment of the invention, a positional timing encoder is used to encode temperature distribution features and real-time operating parameters, establishing a mapping relationship between temperature distribution and operating status in the semantic representation space. This effectively solves the problem of large differences in temperature patterns under different operating conditions and improves the accuracy of anomaly detection. A variational autoencoder is used to encode and model historical normal operation data, and a temperature baseline probability distribution is obtained by modulating real-time operating parameters. This can dynamically adapt to changes in operating conditions, achieve more accurate anomaly measurement, and reduce the false alarm rate. Based on the multi-hop message passing mechanism of functional region directed graphs and graph convolutional networks, it can not only detect temperature anomalies but also accurately locate the anomaly source area and track the anomaly propagation path, providing decision support for rapid diagnosis and handling of equipment faults. By organically combining thermal imaging data with operating parameters and through probabilistic modeling and graph structure analysis, a shift from passive response to proactive early warning is achieved, significantly improving the early detection capability and diagnostic accuracy of temperature anomalies in desulfurization equipment. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of a method for diagnosing temperature anomalies in desulfurization equipment using thermal imaging technology, according to an embodiment of the present invention. Figure 2Flowchart for temperature baseline modeling and anomaly measurement. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the temperature anomaly diagnosis method for desulfurization equipment using thermal imaging technology according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire thermal imaging data of the desulfurization equipment and perform normalization processing to obtain standard thermal imaging data; The temperature distribution features of the standard thermal imaging data are extracted using multi-level feature processing, and the real-time operating parameters of the desulfurization equipment are obtained at the same time. The temperature distribution features and the real-time operating parameters are encoded by a position timing encoder, and a mapping relationship between temperature distribution and operating status is established in the semantic representation space to generate operating condition coupled temperature features. The temperature features coupled with operating conditions in historical normal operation data are input into a variational autoencoder for encoding and modeling. The temperature baseline probability distribution is obtained by modulation based on real-time operating condition parameters. The probability divergence between the latent variable distribution corresponding to the current temperature feature and the temperature baseline probability distribution is calculated to obtain the anomaly measurement value. Based on the physical connection relationship and heat conduction path of each functional area of ​​the desulfurization equipment, a directed graph of the functional area is constructed, and the abnormal measurement value is mapped to the graph node feature. The node role type is determined based on the abnormal measurement relationship between upstream and downstream graph nodes, and the abnormal source area and abnormal propagation path are determined through multi-hop message passing of graph convolutional network. By integrating the abnormal measurement value with the abnormal source region and propagation path, a temperature anomaly diagnosis result is output.

[0020] In one optional implementation, extracting the temperature distribution features of the standard thermal imaging data using multi-level feature processing includes: Adaptive downsampling is performed on the standard thermal imaging data to establish multiple feature extraction levels at different scales. In each feature extraction level, local temperature change information is extracted by a temperature gradient operator, and regional temperature distribution information is extracted by nonlinear transformation. The temperature change information and the temperature distribution information are combined to obtain a hierarchical feature representation. The initial multi-level temperature features are determined by a progressive approach. Based on the initial multi-level temperature features, feature transmission paths are established between different levels and information is exchanged to obtain hierarchically associated temperature features. The hierarchical temperature features are decomposed into temperature gradient features and temperature distribution features. Based on the spatial distribution and temporal variation of the temperature features, the weighting factors of each feature are calculated to generate a feature weighting mapping table. The temperature gradient features and temperature distribution features are then combined in a weighted manner to form the optimized temperature distribution features.

[0021] In one optional embodiment, an adaptive downsampling algorithm is used to reduce the dimensionality of the original thermal imaging data based on its resolution and information density. In practical applications, three downsampling levels can be set, representing 1 / 2, 1 / 4, and 1 / 8 of the original resolution. For example, for a 320×240 resolution thermal imaging dataset, three feature extraction levels of different scales—160×120, 80×60, and 40×30—will be generated sequentially. During adaptive downsampling, the sampling density is dynamically adjusted based on the intensity of local temperature changes, preserving more detailed information in areas of drastic temperature fluctuations.

[0022] In each feature extraction level, local temperature change information is extracted using a temperature gradient operator, and the improved Sobel operator is used to calculate the horizontal and vertical temperature gradients. For the temperature value T(i,j) at location (i,j) in the thermal imaging data, its horizontal gradient Gx and vertical gradient Gy are calculated as follows: Apply the horizontal and vertical Sobel operators to the 3×3 neighborhood of the current point, and calculate the total gradient intensity G = sqrt(Gx) 2 +Gy 2 The gradient information, θ = arctan(Gy / Gx), effectively captures edges and temperature abrupt changes in the heatmap, providing a foundation for subsequent feature extraction.

[0023] Simultaneously, nonlinear transformations are used to extract regional temperature distribution information, and an improved local binary pattern descriptor is employed to capture temperature texture information. The neighborhood temperature value of each pixel is compared with the center temperature value; if the neighborhood temperature is higher than the center temperature, the corresponding position is encoded as 1, otherwise as 0. The resulting binary code is converted into a decimal value, which serves as the temperature distribution feature value for that pixel. Furthermore, a Gaussian weighting function can be used to assign different weights to the temperature values ​​at different locations within the neighborhood, enhancing the robustness of the descriptor.

[0024] The hierarchical feature representation is obtained by combining temperature change information and temperature distribution information. The combination method employs feature channel overlay, that is, merging the temperature gradient feature map and the temperature distribution feature map as different channels. For a certain level, assuming the temperature gradient feature is Fg and the temperature distribution feature is Fd, then the feature representation of that level is F = [Fg, Fd], where [·, ·] represents the concatenation operation of channel dimensions. In this way, a multi-channel feature representation is generated for each level, preserving both the complementary information of temperature gradient change and regional distribution.

[0025] The initial multi-level temperature features are determined through a progressive layer-by-layer approach. Starting with the highest resolution level, the feature representations of each level are calculated sequentially, forming a feature pyramid containing information at multiple scales. In this process, lower-level features capture the global temperature distribution patterns, while higher-level features contain more local details.

[0026] Based on initial multi-level temperature features, feature transfer pathways are established between different levels to facilitate information exchange, including a bidirectional information flow mechanism from top to bottom and bottom to top. In the top-down process, higher-level features are upsampled and transferred to lower levels to enhance their semantic information. In the bottom-up process, lower-level features are transferred to higher levels via skip connections to supplement their detailed information. In practice, 1×1 convolutions are used to adjust the feature channels, ensuring dimensionality matching across different levels. Feature fusion is achieved through element-level addition or concatenation operations. Through this bidirectional information exchange, hierarchically correlated temperature features are obtained, with each level's features incorporating multi-scale information.

[0027] The hierarchical temperature features are decomposed into temperature gradient features and temperature distribution features. Principal component analysis is then used to further decompose the fused features and extract the principal feature components. Specifically, the covariance matrix of the features is calculated, and its eigenvalues ​​and eigenvectors are solved. Based on the magnitude of the eigenvalues, the principal components are selected as the temperature gradient features and temperature distribution features.

[0028] Based on the spatial distribution and temporal variation of temperature features, weighting factors for each feature are calculated. For static thermal imaging, the spatial saliency of the features is mainly considered; for dynamic thermal sequences, the magnitude of change in the temporal dimension also needs to be considered. The saliency of the features is quantified by calculating the local variance or entropy value of the features, and regions with high saliency are assigned greater weight.

[0029] Based on the above calculation results, a feature weighted mapping table is generated, which records the weight coefficients of features at different locations and of different types. The mapping table can be represented as a two-dimensional matrix, where the row and column coordinates correspond to the spatial location of the thermal imaging data, and the element values ​​correspond to the weight coefficients of that location.

[0030] Finally, the temperature gradient features and temperature distribution features are weighted and combined to form the optimized temperature distribution features. The weighted combination uses a linear fusion method, i.e., F_final = α·F_gradient + β·F_distribution, where α and β are weight coefficients obtained from the feature weighting mapping table. This adaptive weighting method allows for flexible adjustment of the feature combination ratio according to the characteristics of different regions, improving the accuracy and adaptability of temperature feature extraction.

[0031] In practical applications, such as anomaly detection in industrial equipment, the optimized temperature distribution features can effectively identify abnormal hot spots on the equipment surface, providing a reliable basis for fault early warning. Similarly, in the field of medical diagnosis, the temperature features extracted by this method can be used to assist in the identification of inflammatory areas, improving the accuracy of disease screening.

[0032] In one optional implementation, the temperature distribution features and the real-time operating parameters are encoded separately using a position-time encoder, establishing a mapping relationship between temperature distribution and operating state in the semantic representation space, and generating operating condition-coupled temperature features including: The temperature distribution features are spatially encoded to mark the location information of the device functional area corresponding to each pixel in the thermal imaging image, thereby generating location-aware temperature features. The real-time operating parameters are time-series encoded to extract dynamic trend information of operating condition changes and generate time-series-aware operating condition features. By using bidirectional projection transformation, the location-aware temperature features and the time-aware operating condition features are mapped to a unified semantic representation space. In the semantic representation space, the similarity matrix between the temperature feature vector and the operating condition feature vector is calculated to determine the temperature region distribution affected by the operating condition and the combination of operating condition parameters that affect the temperature distribution. Based on the temperature region distribution and the combination of operating parameters, by backtracking historical data, the normal temperature distribution range of the corresponding temperature region under different operating conditions is extracted. The temperature characteristics sensed at the current location are compared with the normal temperature distribution range to obtain the temperature deviation characteristics that eliminate the influence of operating conditions. The temperature deviation feature is spliced ​​and fused with the time-series-aware operating condition feature to generate an operating condition-coupled temperature feature.

[0033] In one specific implementation, the temperature distribution characteristics of the acquired thermal imaging image of the desulfurization equipment are spatially encoded. The encoding process uses region partitioning technology, dividing the thermal imaging image into multiple blocks according to the equipment's functional areas, such as the bottom area of ​​the absorption tower, the spray layer area, and the demister area. For each pixel in the thermal imaging image, a location encoding vector is assigned. This vector contains the pixel's relative position within the equipment and the identifier of its functional area. For example, for a pixel in the demister area of ​​the desulfurization tower, its location encoding vector may contain the pixel's relative coordinates within the demister area and the demister area's identifier. The location encoding uses a Gaussian kernel function to process the relative position information, ensuring similarity in encoding adjacent positions while significantly different encodings for distant positions. The location encoding vector is combined with the original temperature feature vector to generate a location-aware temperature feature. This feature not only contains temperature information but also the location information of the equipment's functional areas, providing spatial context for subsequent analysis.

[0034] Real-time acquired operating parameters are time-series encoded, including key operational parameters such as flue gas flow rate, flue gas temperature, slurry pH, and slurry density. The time-series encoding employs a sliding window mechanism, processing the historical data sequence for each operating parameter. For each parameter, its variation characteristics at different time scales are extracted, including short-term fluctuation characteristics, medium-term trends, and long-term operating patterns. For each operating parameter, a three-layer time-series feature is constructed: the first layer is the parameter change rate over the most recent 30 minutes; the second layer is the parameter mean change trend over the most recent 4 hours; and the third layer is the parameter fluctuation range over the most recent 24 hours. For example, for slurry pH, its 30-minute slope, 4-hour mean change curve, and 24-hour fluctuation range can be extracted. Through this multi-scale time-series encoding, time-aware operating feature is generated, which not only includes the current operating parameter value but also the dynamic trend information of parameter changes.

[0035] A bidirectional projection transformation maps location-aware temperature features and time-series-aware operating condition features to a unified semantic representation space. The bidirectional projection transformation consists of two transformation modules: a temperature feature projection module and an operating condition feature projection module. The temperature feature projection module maps location-aware temperature features to a 256-dimensional semantic vector through a multilayer perceptron network; the operating condition feature projection module similarly maps time-series-aware operating condition features to a semantic vector of the same dimension through a multilayer perceptron network. In the semantic representation space, a similarity matrix is ​​calculated between the temperature feature vector and the operating condition feature vector. Cosine similarity is used for similarity calculation; for each pair of temperature regions and operating condition parameters, the cosine of the angle between their feature vectors in the semantic space is calculated. The similarity value ranges from -1 to 1, with a higher value indicating a stronger correlation. By setting a similarity threshold, such as 0.7, highly correlated temperature regions and operating condition parameter combinations are selected. For example, it might be identified that the temperature distribution in the demister region is highly correlated with flue gas flow rate and inlet temperature, while the temperature distribution in the bottom region of the absorber tower is closely related to slurry density and the operating status of the circulating pump.

[0036] Based on the determined temperature region distribution and operating parameter combinations, the normal temperature distribution range of the corresponding temperature region under different operating conditions is extracted through historical data backtracking. The historical data backtracking process consists of two steps: operating condition clustering and temperature distribution extraction. Operating condition clustering classifies historical operating condition data according to different value ranges of the main influencing parameters, forming multiple operating condition sets. For example, when the flue gas flow rate is between 10000-12000 m³ / h... 3 Within a given range ( / h), an inlet temperature between 120-130℃ can be defined as a single operating condition. For each operating condition, temperature distribution data for each functional area during normal equipment operation is extracted from the historical database. The mean and standard deviation of the normal temperature distribution are calculated using statistical analysis methods to form a normal temperature distribution range. The temperature characteristics sensed at the current location are compared with the normal temperature distribution range under the corresponding operating condition to calculate the temperature deviation. The deviation calculation uses a standardized difference method: subtracting the mean normal temperature from the actual temperature value and then dividing by the standard deviation of the normal temperature value yields a standardized temperature deviation index. This method eliminates the normal influence of operating condition changes on the temperature distribution, highlights abnormal temperature changes, and forms a temperature deviation characteristic that eliminates the influence of operating conditions.

[0037] Temperature deviation features are spliced ​​and fused with time-series-aware operating condition features to generate operating condition-coupled temperature features. The splicing and fusion process employs a weighted feature combination method, assigning different weight coefficients to temperature deviation features and operating condition features based on the characteristics of different functional areas. For example, operating condition features may have a higher weight for areas sensitive to operating conditions, such as the spray layer; while temperature deviation features may have a higher weight for relatively stable areas, such as the tower wall. Through weighted combination, a fused feature vector that comprehensively considers the impact of temperature anomalies and operating conditions is generated. This fused feature retains information about temperature anomalies and also includes the changing trends of relevant operating condition parameters, providing a more comprehensive feature representation for subsequent anomaly diagnosis. In practical applications, in a desulfurization system, a 5°C temperature increase in the demister area under a sudden 15% increase in flue gas flow is considered a normal operating condition response. However, under the same operating condition changes, if the temperature increase exceeds 8°C, the temperature deviation features can effectively identify this anomaly, and combined with the operating condition features, it can be determined that it may be a localized overheating phenomenon caused by demister blockage.

[0038] In one optional implementation, the temperature features coupled with operating conditions in historical normal operation data are input into a variational autoencoder for encoding modeling. A temperature baseline probability distribution is obtained by modulating real-time operating condition parameters. The probability divergence between the latent variable distribution corresponding to the current temperature feature and the temperature baseline probability distribution is calculated to obtain anomaly metrics, including: The temperature features coupled with the operating conditions in the historical normal operation data are input into the variational autoencoder for encoding, and the latent variable distribution is obtained in the latent space. The mean vector and variance vector of the latent variable distribution are extracted to obtain the latent space distribution benchmark. For the real-time operating parameters of the desulfurization equipment, the response changes of the mean vector and variance vector in the implicit space distribution benchmark are calculated when the values ​​of each operating parameter change. A mapping relationship between each operating parameter and the response changes of the implicit space distribution benchmark is established to obtain the operating condition-implicit space response mapping table. Extract the values ​​of the current real-time operating condition parameters, input them into the operating condition-latent space response mapping table, and query the offset of the mean vector and the scaling factor of the variance vector. In the latent space distribution benchmark, superimpose the offset on the mean vector and multiply the variance vector by the scaling factor to obtain the distribution of the modulation latent variable. The modulation latent variable distribution is determined as a temperature reference probability distribution that dynamically changes with operating conditions; The temperature characteristics coupled with the operating conditions at the current moment are obtained, encoded into the latent space to obtain the current latent variable distribution, and the probability divergence between the current latent variable distribution and the temperature baseline probability distribution is calculated. The probability divergence is used as a deviation measure to obtain an anomaly measure value.

[0039] In one specific implementation, during the operation of the desulfurization equipment, temperature anomalies coupled with operating conditions are monitored by collecting temperature characteristics of operating conditions from historical normal operation data, including temperature values ​​at each measuring point, temperature gradients and their time change rates, and at the same time recording corresponding operating parameters such as load rate, fuel composition, and ambient temperature.

[0040] The collected temperature features coupled with the operating conditions are input into a pre-designed variational autoencoder (VAE) for encoding. The VAE consists of an encoder and a decoder. The encoder maps the high-dimensional temperature features to a low-dimensional latent space, while the decoder reconstructs the original features from the latent space representation. The encoder is composed of a multi-layer fully connected neural network and outputs two vectors: a mean vector μ and a log-variance vector logσ. 2 , representing the probability distribution in the latent space.

[0041] During the training phase, a variational autoencoder is trained using normal operating data, with optimization objectives including reconstruction error and KL divergence. After training, all historical normal operating data are input into the encoder to obtain the corresponding latent variable distributions. The mean vector and variance vector are extracted from these distributions, and their statistical average is calculated as the benchmark for the latent space distribution.

[0042] For various operating parameters of the desulfurization equipment, such as load rate, inlet SO2 concentration, and slurry pH, the influence of their numerical changes on the implicit space distribution baseline is analyzed. Specifically, while keeping other parameters constant, a single operating parameter is changed, and the changes in the mean vector and variance vector in the implicit space distribution baseline are recorded. Through a large amount of experimental data, a mapping relationship between the operating parameters and the changes in the implicit space response is established, forming an operating condition-implicit space response mapping table.

[0043] This mapping table consists of two parts: a mean offset mapping and a variance scaling mapping. The mean offset mapping represents the offset of the mean vector caused by changes in operating parameters, while the variance scaling mapping represents the scaling factor of the variance vector caused by changes in operating parameters. For example, when the load factor increases from 80% to 90%, the mean of the first dimension in the latent space may increase by 0.2, while the variance may decrease by 10%.

[0044] During the real-time monitoring phase, the current operating parameters are first extracted, such as load rate of 85% and inlet SO2 concentration of 1200 mg / m³. 3 These parameter values ​​are input into the operating condition-latent space response mapping table to retrieve the corresponding mean vector offset and variance vector scaling factor. An offset is added to the mean vector in the latent space distribution baseline, and the variance vector is multiplied by the scaling factor to obtain the modulated latent variable distribution. This distribution is the temperature baseline probability distribution under the current operating condition.

[0045] The actual operating condition coupled temperature characteristics at the current moment are obtained and input into the encoder to obtain the current latent variable distribution. The probability divergence between the current latent variable distribution and the temperature baseline probability distribution is calculated. Commonly used metrics include KL divergence or Wasserstein distance. The calculated probability divergence is used as a deviation metric, i.e., the anomaly metric.

[0046] Specifically, if the current latent variable distribution is P and the temperature baseline probability distribution is Q, both of which are multidimensional Gaussian distributions, then the KL divergence can be represented as the distance between the two distributions. When the KL divergence exceeds a preset threshold, the current temperature characteristic is determined to be abnormal, and the system issues an alarm.

[0047] In practical applications, different thresholds can be set for different equipment areas. For example, a lower threshold can be set for critical components such as the bottom of the absorption tower to improve sensitivity; for areas with large fluctuations, the threshold can be appropriately increased to reduce false alarms.

[0048] This anomaly detection method based on variational autoencoders and operating condition modulation can adapt to normal temperature fluctuations in equipment under different operating conditions, reducing false alarms caused by changes in operating conditions and improving the accuracy and reliability of temperature anomaly detection in desulfurization equipment. By calculating anomaly metrics in real time, equipment temperature anomalies can be detected promptly, providing decision support for operation and maintenance personnel and avoiding production losses caused by equipment failures.

[0049] like Figure 2 As shown, a flowchart illustrating temperature baseline modeling and anomaly measurement is presented.

[0050] In one optional implementation, for the real-time operating parameters of the desulfurization equipment, the response changes of the mean vector and variance vector in the implicit space distribution benchmark are calculated when the values ​​of each operating parameter change, and a mapping relationship between each operating parameter and the response changes of the implicit space distribution benchmark is established, resulting in an operating condition-implicit space response mapping table including: The real-time operating parameters of the desulfurization equipment are divided into multiple subsets of operating parameters according to the physical action mechanism. The operating parameters in each subset are combined in pairs to obtain the operating parameter combination pairs. For each pair of operating condition parameters, extract sample data of synchronous changes of the pair of operating condition parameters from historical normal operation data, calculate the joint response change of the mean vector and variance vector in the latent space distribution benchmark when the pair of operating condition parameters changes jointly, and identify the coupling effect type between the pair of operating condition parameters. A synergistic enhancement coefficient and a mutual inhibition coefficient are preset, and the corresponding coupling coefficient is selected according to the coupling interaction type. When the coupling interaction type is synergistic enhancement, the synergistic enhancement coefficient is selected; when the coupling interaction type is mutual inhibition, the mutual inhibition coefficient is selected. The individual response changes of each operating parameter in the operating parameter combination pair are summed, and the summation result is multiplied by the coupling coefficient to obtain the joint response correction amount of the operating parameter combination pair. Calculate the independent response change of a single operating condition parameter and the joint response correction of a combination of operating condition parameters, and store the independent response change and the joint response correction in the operating condition-latent space response mapping table to establish a mapping relationship that includes single-parameter response and multi-parameter coupled response.

[0051] In one specific implementation, the real-time operating parameters of the desulfurization equipment are divided into multiple subsets based on their physical mechanisms. For example, the operating parameters can be divided into subsets of flue gas system parameters, absorption system parameters, and slurry circulation system parameters. The flue gas system parameter subset includes inlet flue gas flow rate, inlet sulfur dioxide concentration, and inlet flue gas temperature; the absorption system parameter subset includes the absorber liquid-to-gas ratio, absorbent concentration, and pH value; and the slurry circulation system parameter subset includes slurry circulation flow rate, slurry density, and gypsum discharge rate.

[0052] For each subset of operating parameters, pairwise combinations are made to obtain operating parameter combination pairs. For example, for a subset of flue gas system parameters, combinations such as inlet flue gas flow rate and inlet sulfur dioxide concentration, inlet flue gas flow rate and inlet flue gas temperature, and inlet sulfur dioxide concentration and inlet flue gas temperature can be obtained.

[0053] For each pair of operating parameters, sample data showing synchronous changes in that pair are extracted from historical normal operating data. For example, historical data points showing simultaneous changes in inlet flue gas flow rate and inlet sulfur dioxide concentration are selected as samples. Using these sample data, the joint change in the mean vector and variance vector in the latent space distribution benchmark is calculated through a pre-trained deep autoencoder model.

[0054] Specifically, the encoder portion of a depth autoencoder maps the original operating parameters to the latent space. Assume that under normal operating conditions, the mean vector of the latent space distribution baseline is μ0, and the variance vector is σ0. 2 When the values ​​of the combination of operating parameters change, a new latent space representation is calculated by the encoder, and the corresponding mean vector becomes μ1 and the variance vector becomes σ1. 2 The change in the joint response can be expressed as the change in the mean vector Δμ = μ1 - μ0 and the change in the variance vector Δσ. 2 =σ1 2 -σ0 2 .

[0055] By analyzing the changes in the combined response, the coupling type between pairs of operating parameters is identified. When the combined response change caused by the simultaneous change of two parameters is greater than the sum of their individual responses, it is determined to be a synergistic enhancement type; when the combined response change is less than the sum of their individual responses, it is determined to be a mutual inhibition type.

[0056] The synergistic enhancement coefficient α and the mutual inhibition coefficient β are preset. Based on actual operating experience and data analysis results, α can be set to 1.2 and β to 0.8. The corresponding coupling coefficient is selected according to the identified coupling type: when the coupling type is synergistic enhancement, the synergistic enhancement coefficient α is selected; when the coupling type is mutual inhibition, the mutual inhibition coefficient β is selected.

[0057] The individual response changes of each operating parameter in the operating parameter combination pair are summed, and the summation result is multiplied by the coupling coefficient to obtain the joint response correction amount of the operating parameter combination pair. For example, for the operating parameter combination pair consisting of inlet flue gas flow rate and inlet sulfur dioxide concentration, if their individual response changes are Δμ1 and Δμ2 respectively, and the coupling effect type is synergistic enhancement, then the joint response correction amount is α×(Δμ1+Δμ2).

[0058] To establish a complete operating condition-latent space response mapping table, it is necessary to calculate the independent response changes of individual operating condition parameters and the joint response corrections for combinations of operating condition parameters. The independent response changes are obtained by changing the value of a single operating condition parameter while keeping other parameters constant, and observing the changes in the mean and variance vectors in the latent space distribution baseline. For example, by changing only the inlet flue gas flow rate, the changes in the latent space mean and variance vectors are calculated to obtain the independent response changes of the inlet flue gas flow rate.

[0059] All independent response changes and joint response corrections are stored together in a load condition-implicit space response mapping table, establishing a mapping relationship that includes single-parameter responses and multi-parameter coupled responses. This mapping table uses load parameters or combinations of load parameters as indices and the corresponding response changes as values, forming a complete mapping structure.

[0060] In practical applications, when the operating parameters of the desulfurization system change, the changes in the latent space distribution can be quickly estimated by querying this mapping table, thereby determining whether the system is in a normal state. For example, when a 20% increase in inlet flue gas flow and a 15% increase in inlet sulfur dioxide concentration are detected, the mapping table is queried to obtain the independent response changes of these two parameters and their combined joint response correction. This allows for a comprehensive assessment of whether the latent space distribution corresponding to the current operating condition deviates from the normal range, thus enabling early identification and warning of abnormal states of the desulfurization equipment.

[0061] Using the above methods, a mapping relationship between operating parameters and implicit space distributed responses was established. This approach considers not only the influence of individual operating parameters but also the coupling effects between multiple parameters, improving the accuracy and adaptability of anomaly detection in desulfurization equipment. This method is particularly suitable for scenarios where parameters in industrial processes exhibit complex interactions, and can more accurately reflect the impact of parameter changes on the system state.

[0062] In one optional implementation, a directed graph of functional regions is constructed based on the physical connection relationships and heat conduction paths of each functional region of the desulfurization equipment. Anomaly metrics are mapped to graph node features. Node role types are determined based on the anomaly metric relationships between upstream and downstream graph nodes. The anomaly source region and anomaly propagation path are determined through multi-hop message passing in a graph convolutional network, including: Obtain the physical connection relationship and heat conduction path between the functional areas of the desulfurization equipment, construct a directed graph of the functional areas according to the heat transfer direction, and map the abnormal metric values ​​of each functional area to the initial features of the corresponding graph nodes; For each graph node in the directed graph of the functional area, the abnormal metric value of the upstream neighboring node is extracted by tracing back along the directed edge to determine the upstream abnormal input value, and the abnormal metric value of the downstream neighboring node is extracted by extending forward to determine the downstream abnormal received value. Calculate the anomaly measurement relationship between the current graph node and the upstream and downstream graph nodes to obtain the anomaly source intensity and anomaly propagation attenuation rate. Based on the numerical relationship, determine the anomaly generation type node, the anomaly propagation type node, and the anomaly termination type node, and construct the anomaly propagation feature vector. The anomaly propagation feature vector is input into a graph convolutional network for multi-hop message passing. Anomaly-generating nodes are identified as potential anomaly source nodes, and directed edges connecting anomaly-terminating nodes are identified as anomaly propagation blocking edges. The region corresponding to the node with the highest anomaly intensity and no upstream anomaly input among the potential anomaly source nodes is identified as the anomaly source region. Starting from the graph node corresponding to the anomaly source region, traverse the directed edges that are not marked as anomaly propagation blocking edges along the positive direction of the directed edges to determine the anomaly propagation path.

[0063] In one specific implementation, by acquiring the physical connections and heat conduction paths between the functional areas of the desulfurization equipment, the direction of heat transfer within the equipment can be clearly identified. For example, in a wet desulfurization unit, the absorption tower contains functional areas such as a spray layer, a demister, and an oxidation zone. Flue gas and slurry flow between these areas, forming heat transfer paths. Based on these physical connections and heat conduction directions, a directed graph G=(V, E) of the functional areas is constructed, where V represents the set of functional area nodes and E represents the directed connections between areas. For each functional area vi, its abnormal metrics such as temperature, pressure, and concentration are collected, and these values ​​are mapped to the initial feature vector Fi of the corresponding graph nodes.

[0064] For each node *vi* in the directed graph, trace back along the directed edges to its upstream neighboring node set U(vi) = {vj | (vj, vi) ∈ E}, and calculate the upstream abnormal input value Iin(vi). Specifically, this is achieved by extracting the weighted average of the abnormal metrics of all upstream nodes; the weights can be determined based on the heat conduction coefficient of the connecting edges. Similarly, by extending forward, determine the downstream neighboring node set D(vi) = {vk | (vi, vk) ∈ E}, and calculate the downstream abnormal reception value Iout(vi). When a functional region node has no upstream connection, its upstream abnormal input value is set to zero; similarly, when there is no downstream connection, its downstream abnormal reception value is also set to zero.

[0065] After obtaining the upstream and downstream anomaly metrics of a node, the anomaly metric relationship between the current node vi and its upstream and downstream nodes is calculated. The anomaly source intensity S(vi) is defined as the difference between the current node's anomaly metric Fi and the upstream anomaly input value Iin(vi), i.e., S(vi) = Fi - Iin(vi); the anomaly propagation attenuation rate A(vi) is defined as the ratio of the downstream anomaly received value Iout(vi) to the current node's anomaly metric Fi, i.e., A(vi) = Iout(vi) / Fi. Based on the numerical relationship between these two indicators, node roles are determined: when S(vi) is greater than a preset threshold Ts and Fi is greater than Iin(vi), node vi is determined as an anomaly generating node; when |Fi - Iin(vi)| is less than a threshold Tc and A(vi) is close to 1, it is determined as an anomaly propagating node; when A(vi) is less than a threshold Ta and a downstream node exists, it is determined as an anomaly terminating node. Based on the above determination results, an anomaly propagation feature vector P(vi) = [S(vi), A(vi), TypeCode] is constructed, where TypeCode is the encoding value of the node type.

[0066] The constructed anomaly propagation feature vectors are input into a graph convolutional network for multi-hop message passing. The graph convolutional layer is defined as H(l+1) = σ(ÃH(l)W(l)), where à is the normalized adjacency matrix, H(l) is the node feature matrix of the l-th layer, W(l) is the weight matrix, and σ is the activation function. The initial feature matrix H(0) consists of the anomaly propagation feature vectors P(vi) of each node. By using the designed graph convolutional network for message passing with more than two hops, the multi-step propagation pattern of anomalies between functional regions can be captured. Furthermore, the set of potential anomaly-generating nodes S is identified, and the directed edges connecting anomaly-terminating nodes are labeled as the set of anomaly propagation blocking edges B.

[0067] In the set S of potential anomaly source nodes, the node that meets the following conditions is identified as the final anomaly source node: the anomaly source intensity S(vi) is the largest and there is no upstream anomaly input (Iin(vi)=0). The functional area corresponding to this node is the anomaly source area. For example, in a desulfurization system, if the oxidation fan area is identified as the anomaly source area, it indicates that the equipment failure in this area is the root cause of the abnormal temperature in the entire system.

[0068] Starting from the graph node vi corresponding to the identified anomaly source region, a depth-first traversal is performed along the positive direction of the directed edges. During the traversal, only directed edges not marked as anomaly propagation blocking edges are used to construct one or more paths from the source to the anomaly's termination region; these paths are the anomaly propagation paths. For example, an anomaly might propagate from the oxidation blower area to the slurry circulation area, then to the spray layer and flue gas outlet area, but be blocked in the demister area. This information is of great guiding significance for fault diagnosis and maintenance strategy formulation.

[0069] In one optional implementation, the anomaly measurement relationship between the current graph node and its upstream and downstream graph nodes is calculated to obtain the anomaly source intensity and anomaly propagation attenuation rate. Based on the numerical relationship, anomaly-generating nodes, anomaly-propagating nodes, and anomaly-terminating nodes are determined, and an anomaly propagation feature vector is constructed, including: The difference between the current graph node's anomaly metric value and the upstream anomaly input value is calculated to obtain the upstream anomaly increment. The ratio between the upstream anomaly increment and the current graph node's anomaly metric value is calculated to obtain the anomaly source intensity. The anomaly source intensity represents the proportion of the current graph node's autonomous contribution to the anomaly relative to the upstream input. The difference between the current graph node's anomaly metric value and the downstream anomaly received value is calculated to obtain the downstream anomaly attenuation. The ratio between the downstream anomaly attenuation and the current graph node's anomaly metric value is calculated to obtain the anomaly conduction attenuation rate. The anomaly conduction attenuation rate represents the proportion of energy loss as the anomaly propagates downstream. Based on the numerical relationship between the abnormal source intensity and the abnormal conduction attenuation rate, the role type of the current graph node is determined, including: when the abnormal source intensity is greater than the abnormal conduction attenuation rate, it is marked as an abnormal generation node; when the abnormal source intensity is less than the abnormal conduction attenuation rate, it is marked as an abnormal conduction node; when both the abnormal source intensity and the abnormal conduction attenuation rate are lower than the preset role determination threshold, it is marked as an abnormal termination node. The abnormal source intensity, abnormal propagation attenuation rate, and role type are combined to form an abnormal propagation feature vector.

[0070] In one specific implementation, when analyzing abnormal nodes in a network graph structure, the abnormality metric values ​​of the current graph node and its upstream and downstream nodes are obtained. The abnormal node type is identified by calculating the abnormality source intensity and abnormal propagation attenuation rate, and an abnormal propagation feature vector is constructed.

[0071] The anomaly metric for each node in the current network is obtained. The anomaly metric can be calculated based on various methods, such as the degree of deviation of node attribute data, the magnitude of time-series fluctuations, or the degree of difference between node behavior characteristics and normal patterns. For node v, its anomaly metric is denoted as A(v), which represents the quantitative indicator of the node's abnormal state.

[0072] For the target node v in the current analysis, identify its upstream node set U(v) and downstream node set D(v). Upstream nodes are those nodes that directly transmit information or influence the target node v, while downstream nodes are those that receive the output of the target node v. In a directed graph, upstream nodes are nodes that point to v, and downstream nodes are nodes that v points to.

[0073] The upstream anomaly input value I(v) is calculated by taking a weighted average of the anomaly metrics of all upstream nodes: Upstream abnormal input value = weighted average of upstream node abnormality metrics; The weights can be determined based on the connection strength or influence between the upstream node and the current node. If there is no upstream node, I(v) defaults to 0.

[0074] Calculate the anomaly source intensity S(v) of the current node v, which represents the proportion of the node's autonomous contribution to the anomaly relative to the upstream input: If I(v) is 0 and A(v) is greater than 0, then S(v) = 1, indicating that node v is a complete source of anomalies; If I(v) is greater than 0, then S(v) = (A(v) - I(v)) / A(v), which represents the proportion of anomalies generated by the node itself to the total anomalies.

[0075] The range of the anomaly source intensity S(v) is [-∞, 1]. A positive value indicates that the node enhances the anomaly signal, a negative value indicates that the node suppresses the anomaly signal to a certain extent, and a value of 0 indicates that the node only transmits the anomaly without changing its intensity.

[0076] The downstream abnormal received value R(v) is calculated in a similar manner to the upstream abnormal input value: Downstream abnormal received value = weighted average of downstream node abnormal measurement values; Calculate the anomalous conduction attenuation rate D(v), which represents the proportion of energy lost during the downstream propagation of the anomalous event: If A(v) is 0, then D(v) = 0; If A(v) is greater than 0 and R(v) = 0, then D(v) = 1, indicating that the anomaly has been completely absorbed by the current node; In other cases, D(v) = (A(v) - R(v)) / A(v) quantifies the degree of attenuation of the anomaly during propagation.

[0077] The range of the anomaly conduction attenuation rate D(v) is [-∞, 1]. A positive value indicates that the anomaly weakens during propagation, a negative value indicates that the anomaly strengthens during propagation, and a value of 0 indicates that the anomaly intensity remains unchanged.

[0078] Based on the numerical relationship between the anomaly source intensity S(v) and the anomaly conduction attenuation rate D(v), determine the role type of the current node v: When S(v)>D(v) and S(v)>θ (θ is a preset role determination threshold, such as 0.3), it is marked as an abnormal generation node. This type of node generates anomalies and propagates them outward. When D(v)>S(v) and D(v)>θ, it is marked as an anomaly propagation type node. This type of node is mainly responsible for the propagation and amplification of anomalies. When S(v) < θ and D(v) < θ, it is marked as an anomalous termination node. This type of node neither generates anomalies significantly nor propagates them.

[0079] The anomaly source intensity S(v), the anomaly conduction attenuation rate D(v), and the role type Type(v) are combined to form the anomaly propagation feature vector F(v)=[S(v), D(v), Type(v)].

[0080] The anomaly propagation feature vector constructed in this way can clearly depict the generation, propagation, and dissipation of anomalies in the network, providing strong support for anomaly propagation path tracing and key node identification.

[0081] The temperature anomaly diagnosis system for desulfurization equipment utilizing thermal imaging technology, as described in this embodiment of the invention, includes: The data preprocessing module is used to acquire thermal imaging data from desulfurization equipment and perform normalization processing to obtain standard thermal imaging data. The feature encoding module is used to extract the temperature distribution features of the standard thermal imaging data using multi-level feature processing, and at the same time obtain the real-time operating parameters of the desulfurization equipment. The temperature distribution features and the real-time operating parameters are encoded by the position timing encoder, and a mapping relationship between temperature distribution and operating status is established in the semantic representation space to generate operating condition coupled temperature features. The modeling and evaluation module is used to input the temperature features coupled with the operating conditions in the historical normal operation data into the variational autoencoder for encoding and modeling. It obtains the temperature reference probability distribution based on the real-time operating condition parameters, calculates the probability divergence between the latent variable distribution corresponding to the current temperature feature and the temperature reference probability distribution, and obtains the anomaly measurement value. The graph propagation analysis module is used to construct a directed graph of functional areas based on the physical connection relationship and heat conduction path of each functional area of ​​the desulfurization equipment, map abnormal measurement values ​​to graph node features, determine the node role type based on the abnormal measurement relationship between upstream and downstream graph nodes, and determine the abnormal source area and abnormal propagation path through multi-hop message passing of graph convolutional network. The result output module is used to integrate the abnormal measurement value with the abnormal source region and abnormal propagation path to output the temperature abnormality diagnosis result.

[0082] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0083] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0084] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing temperature abnormality of a desulfurization apparatus using thermography, characterized by, The method comprises the following steps: acquiring thermal imaging data of a desulfurization device and performing normalization processing to obtain standard thermal imaging data; extracting temperature distribution features of the standard thermal imaging data by using multi-level feature processing, simultaneously acquiring real-time working condition parameters of the desulfurization device, respectively encoding the temperature distribution features and the real-time working condition parameters by using a position time sequence encoder, establishing a mapping relationship between the temperature distribution and the working condition state in a semantic representation space, and generating temperature features coupled with working conditions; inputting the temperature features coupled with working conditions in historical normal operation data into a variational autoencoder for encoding modeling, obtaining a temperature reference probability distribution according to real-time working condition parameters, calculating a probability divergence between an implicit variable distribution corresponding to the current temperature features and the temperature reference probability distribution, and obtaining an abnormality measurement value; according to a physical connection relationship and a heat conduction path of each functional region of the desulfurization device, constructing a functional region directed graph, mapping the abnormality measurement value to a graph node feature, determining a node role type based on an abnormality measurement relationship between upstream and downstream graph nodes, determining an abnormal source region and an abnormal propagation path through multi-hop message passing of a graph convolution network; and fusing the abnormality measurement value, the abnormal source region and the abnormal propagation path, and outputting a temperature abnormality diagnosis result.

2. The method of claim 1, wherein, The method for extracting the temperature distribution features of the standard thermal imaging data by using multi-level feature processing comprises the following steps: performing adaptive down-sampling processing on the standard thermal imaging data to establish a plurality of feature extraction levels with different scales, in each feature extraction level, extracting local temperature change information by using a temperature gradient operator, simultaneously extracting regional temperature distribution information by using a nonlinear transformation, combining the temperature change information and the temperature distribution information to obtain a level feature representation, and determining an initial multi-level temperature feature through a layer-by-layer progressive manner; based on the initial multi-level temperature feature, establishing a feature transmission path between different levels and performing information interaction to obtain a level-associated temperature feature; performing feature decomposition on the level-associated temperature feature to separate the temperature feature into a temperature gradient feature and a temperature distribution feature, calculating a weighting factor of each feature according to a spatial distribution and a time sequence change of the temperature feature, generating a feature weighting mapping table, and performing weighted combination on the temperature gradient feature and the temperature distribution feature to form an optimized temperature distribution feature.

3. The method of claim 1, wherein, The method for respectively encoding the temperature distribution features and the real-time working condition parameters by using a position time sequence encoder, establishing a mapping relationship between the temperature distribution and the working condition state in a semantic representation space, and generating temperature features coupled with working conditions comprises the following steps: performing spatial position coding on the temperature distribution features to mark position information of a functional region of a device corresponding to each pixel point in a thermal imaging image, and generating position-aware temperature features; performing time sequence coding on the real-time working condition parameters to extract dynamic trend information of working condition changes, and generating time sequence-aware working condition features; through bidirectional projection transformation, respectively mapping the position-aware temperature features and the time sequence-aware working condition features to a unified semantic representation space, calculating a similarity matrix between a temperature feature vector and a working condition feature vector in the semantic representation space, and determining a temperature region distribution affected by a working condition and a working condition parameter combination affecting the temperature distribution. According to the temperature region distribution and the working condition parameter combination, through historical data backtracking, the normal temperature distribution range of the corresponding temperature region under different working condition states is extracted, the current position-aware temperature feature is compared with the normal temperature distribution range, and a temperature deviation feature eliminating the working condition influence is obtained; The temperature deviation feature is spliced and fused with the time sequence-aware working condition feature to generate a working condition-coupled temperature feature.

4. The method of claim 1, wherein, The working condition-coupled temperature feature in the historical normal operation data is input into a variational autoencoder for encoding modeling, a temperature reference probability distribution is obtained according to real-time working condition parameters modulation, a probability divergence between the current temperature feature corresponding to the hidden variable distribution and the temperature reference probability distribution is calculated, and an abnormality measurement value is obtained, including: The working condition-coupled temperature feature in the historical normal operation data is input into a variational autoencoder for encoding, and a hidden variable distribution is obtained in the hidden space. The mean vector and the variance vector of the hidden variable distribution are extracted to obtain a hidden space distribution reference. For real-time working condition parameters of the desulfurization equipment, the response change amount of the mean vector and the variance vector in the hidden space distribution reference when each working condition parameter value changes is calculated, a mapping relationship between each working condition parameter and the response change amount of the hidden space distribution reference is established, and a working condition-hidden space response mapping table is obtained. The value of the current real-time working condition parameter is extracted and input into the working condition-hidden space response mapping table to query the offset of the mean vector and the scaling coefficient of the variance vector. In the hidden space distribution reference, the mean vector is superimposed with the offset, and the variance vector is multiplied by the scaling coefficient to obtain a modulated hidden variable distribution. The modulated hidden variable distribution is determined as a temperature reference probability distribution that dynamically changes with the working condition. For real-time working condition parameters of the desulfurization equipment, the response change amount of the mean vector and the variance vector in the hidden space distribution reference when each working condition parameter value changes is calculated, a mapping relationship between each working condition parameter and the response change amount of the hidden space distribution reference is established, and a working condition-hidden space response mapping table is obtained, including:

5. The method of claim 4, wherein, The real-time working condition parameters of the desulfurization equipment are divided into multiple working condition parameter subsets according to physical action mechanisms. For the working condition parameters in each working condition parameter subset, a working condition parameter combination pair is obtained by two-by-two combination. For each working condition parameter combination pair, sample data of synchronous changes of the working condition parameter combination pair is extracted from the historical normal operation data, the joint response change amount of the mean vector and the variance vector in the hidden space distribution reference when the working condition parameter combination pair jointly changes is calculated, and the coupling effect type between the working condition parameter combination pair is identified. The synergistic enhancement coefficient and the mutual inhibition coefficient are set in advance. According to the coupling effect type, the corresponding coupling coefficient is selected. When the coupling effect type is synergistic enhancement, the synergistic enhancement coefficient is selected. When the coupling effect type is mutual inhibition, the mutual inhibition coefficient is selected. ​ Summing up the individual response change amounts of each working condition parameter to obtain a combined response correction amount of the working condition parameter combination by multiplying the combined response correction amount by the coupling coefficient; The independent response change amount of a single working condition parameter and the combined response correction amount of the working condition parameter combination are calculated respectively, and the independent response change amount and the combined response correction amount are stored in a working condition-hidden space response mapping table to establish a mapping relationship including single-parameter responses and multi-parameter coupled responses.

6. The method of claim 1, wherein, According to the physical connection relationship and heat conduction path of each functional region of the desulfurization equipment, a directed graph of the functional regions is constructed, the abnormality metric value is mapped to the graph node feature, the node role type is determined based on the abnormality metric relationship between the upstream and downstream graph nodes, and the abnormal source region and the abnormal propagation path are determined through multi-hop message passing of the graph convolution network, including: Obtaining the physical connection relationship and heat conduction path between each functional region of the desulfurization equipment, constructing a directed graph of the functional regions according to the heat transfer direction, and mapping the abnormality metric value of each functional region to the initial feature of the corresponding graph node; For each graph node in the directed graph of the functional regions, the upstream adjacent node abnormality metric value is extracted by reverse tracing along the directed edge to determine the upstream abnormal input value, and the downstream adjacent node abnormality metric value is extracted by forward extension to determine the downstream abnormal receiving value; The abnormality metric relationship between the current graph node and the upstream and downstream graph nodes is calculated to obtain the abnormal source intensity and the abnormal conduction attenuation rate, and the abnormal generation type node, the abnormal conduction type node and the abnormal termination type node are determined according to the numerical relationship, and the abnormal propagation feature vector is constructed; The abnormal generation type node is determined as the potential abnormal source node, and the directed edge connecting the abnormal termination type node is determined as the abnormal propagation blocking edge through multi-hop message passing of the graph convolution network; The region corresponding to the node with the maximum abnormal source intensity and no upstream abnormal input in the potential abnormal source node is determined as the abnormal source region; Starting from the graph node corresponding to the abnormal source region, the directed edges that are not marked as abnormal propagation blocking edges are traversed in the forward direction of the directed edges to determine the abnormal propagation path.

7. The method of claim 6, wherein, The abnormality metric relationship between the current graph node and the upstream and downstream graph nodes is calculated to obtain the abnormal source intensity and the abnormal conduction attenuation rate, and the abnormal generation type node, the abnormal conduction type node and the abnormal termination type node are determined according to the numerical relationship, and the abnormal propagation feature vector is constructed, including: The difference between the abnormality metric value of the current graph node and the upstream abnormal input value is calculated to obtain the upstream abnormal increment, and the proportion relationship between the upstream abnormal increment and the abnormality metric value of the current graph node is calculated to obtain the abnormal source intensity, which represents the abnormal autonomous contribution proportion of the current graph node relative to the upstream input; The difference between the abnormality metric value of the current graph node and the downstream abnormal receiving value is calculated to obtain the downstream abnormal attenuation amount, and the proportion relationship between the downstream abnormal attenuation amount and the abnormality metric value of the current graph node is calculated to obtain the abnormal conduction attenuation rate, which represents the energy loss proportion of the abnormality propagation to the downstream; According to the numerical relationship between the abnormal source intensity and the abnormal conduction attenuation rate, the role type of the current graph node is judged, including: when the abnormal source intensity is greater than the abnormal conduction attenuation rate, it is marked as an abnormal generation type node, when the abnormal source intensity is less than the abnormal conduction attenuation rate, it is marked as an abnormal conduction type node, and when the abnormal source intensity and the abnormal conduction attenuation rate are both lower than a preset role determination threshold, it is marked as an abnormal termination type node; The abnormal source intensity, the abnormal conduction attenuation rate and the role type are combined to form an abnormal propagation feature vector.

8. A temperature anomaly diagnostic system for desulphurization plants using thermography, for implementing the method according to any one of the preceding claims 1-7, characterized in that, Comprise: A data preprocessing module is configured to acquire thermal imaging data of a desulfurization device and perform normalization processing to obtain standard thermal imaging data; A feature encoding module is configured to extract temperature distribution features of the standard thermal imaging data by using multi-level feature processing, acquire real-time working condition parameters of the desulfurization device, encode the temperature distribution features and the real-time working condition parameters by using a position time encoder, establish a mapping relationship between temperature distribution and working condition state in a semantic representation space, and generate working condition-coupled temperature features; A modeling and evaluation module is configured to input the working condition-coupled temperature features in historical normal operation data into a variational autoencoder to perform encoding modeling, obtain a temperature reference probability distribution according to real-time working condition parameters, calculate a probability divergence between an implicit variable distribution corresponding to a current temperature feature and the temperature reference probability distribution, and obtain an abnormality measurement value; A graph propagation analysis module is configured to construct a functional region directed graph according to a physical connection relationship and a heat conduction path of each functional region of the desulfurization device, map the abnormality measurement value to graph node features, determine a node role type based on an abnormality measurement relationship between upstream and downstream graph nodes, determine an abnormal source region and an abnormal propagation path by using multi-hop message passing of a graph convolution network, and A result output module is configured to fuse the abnormality measurement value, the abnormal source region and the abnormal propagation path, and output a temperature abnormality diagnosis result.

9. An electronic device, comprising: Comprise: A processor; A memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 7.

Images