Boiler combustion state anomaly early warning method based on graph neural network

Abstract

The application discloses a boiler combustion state abnormal early warning method based on a graph neural network, obtains a pretreated multi-source boiler operation data set, constructs a boiler combustion space-time heterogeneous initial graph structure, obtains a Graph Sparse Training initial model, generates a dynamic sparse graph structure, obtains node embedding representation, obtains a structure reconstruction error scalar value, obtains an overall time sequence reconstruction error scalar value, calculates a node-level initial abnormal score, generates a node-level calibration abnormal score, and forms a boiler combustion state abnormal detection result. The application provides accurate and auditable abnormal diagnosis results for boiler operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of boiler combustion status technology, and in particular to a method for early warning of abnormal boiler combustion status based on graph neural networks. Background Technology

[0002] With the continuous expansion of industrial boiler systems and the ongoing improvement of their automation and intelligence levels, the safe, economical, and environmentally friendly operation of boiler combustion processes places higher demands on the real-time monitoring and accurate diagnosis of abnormal states. Currently, the detection of abnormal boiler combustion states mainly relies on alarm strategies based on single-point thresholds, traditional empirical rule systems, or the use of end-to-end time series models for anomaly identification. Existing technologies have the following shortcomings:

[0003] In large-scale, tightly coupled boiler monitoring networks, existing anomaly detection methods based on high-dimensional fully connected modeling often require a large amount of computing resources and are difficult to run in real time on field edge devices, given limited computing power and strict time delay constraints. Some solutions adapt to actual computing power conditions by downsampling or reducing model complexity, but this leads to a decrease in detection sensitivity and an increase in the risk of false alarms and missed alarms, which seriously affects the safety of boiler operation and the timeliness of early warning.

[0004] Sensor data from boiler combustion processes are susceptible to noise interference, missing data, and operational drift. Traditional methods often rely on univariate thresholds or black-box time series modeling, which struggle to effectively capture the dynamic coupling changes between multiple variables and lack the ability to accurately pinpoint the root causes of anomalies. Existing models often fail to output traceable and auditable physical explanations, limiting their reliability and usability in real-world production environments.

[0005] Existing technologies generally suffer from insufficient system robustness, untimely model updates, and a lack of adaptive capabilities in multi-source data integration, dynamic network structure modeling, and real-time edge inference. Faced with frequent changes in field conditions and sensor anomalies, traditional anomaly detection systems often exhibit delayed reporting, false alarms, and missed alarms, making it difficult to provide maintenance personnel with high-value fault tracing and decision support. Summary of the Invention

[0006] One objective of this invention is to propose a method for early warning of abnormal boiler combustion status based on graph neural networks. This invention provides accurate and auditable abnormal diagnosis results for boiler operation and maintenance.

[0007] A method for early warning of abnormal boiler combustion status based on graph neural networks according to an embodiment of the present invention includes:

[0008] S1. Collect multi-source boiler operation datasets and preprocess them to obtain preprocessed multi-source boiler operation datasets;

[0009] Real-time measurement data from the boiler control system, offline measurement data from the historical database, and actuator control command data are collected and merged to form a multi-source boiler operation dataset. The multi-source boiler operation dataset is then cleaned, time-aligned, has missing measurement interpolation, and undergoes initial outlier suppression to obtain a preprocessed multi-source boiler operation dataset.

[0010] S2. Based on the boiler flow chart and equipment ledger information, map each measuring point and actuator in the preprocessed multi-source boiler operation data to a node in the graph structure, and map the physical adjacency relationship, thermal coupling relationship and control link relationship to the edge in the graph structure to construct the initial graph structure of boiler combustion spatiotemporal heterogeneity.

[0011] S3. Initialize the parameters of the Graph Sparse Training dynamic sparse model using the initial graph structure of the spatiotemporal heterogeneity of boiler combustion and the preprocessed multi-source boiler operation dataset to obtain the initial Graph Sparse Training model;

[0012] S4. On the initial Graph Sparse Training model, pruning and structural growth operations are performed based on structural sensitivity index and temporal mutual information index. Thermal critical path edges are retained to generate a dynamic sparse graph structure. The dynamic sparse graph structure is input into the node representation encoder to obtain the node embedding representation.

[0013] S5. Based on node embedding representation, the adjacency relationship and edge weight distribution of the dynamic sparse graph structure are reconstructed through the structure decoder to obtain the scalar value of the structure reconstruction error;

[0014] S6. Based on node embedding representation, the short-window multivariate observation trajectory of each node in the dynamic sparse graph structure is reconstructed through a temporal decoder to obtain the overall temporal reconstruction error scalar value;

[0015] S7. Calculate the initial anomaly score at the node level based on the scalar value of the structural reconstruction error, the scalar value of the overall temporal reconstruction error, and the short-term prediction residual;

[0016] S8. Perform physical consistency calibration between the node-level initial anomaly score and the industrial boiler thermal balance boundary conditions, air excess coefficient boundary conditions, and flue gas path lag time boundary conditions to generate node-level calibration anomaly scores.

[0017] S9. Perform attribution heat backtracking on the node-level calibration anomaly score on the dynamic sparse graph structure, extract the minimum causal link that leads to the anomaly, and locate the potential root cause node and root cause edge to form the boiler combustion state anomaly detection result.

[0018] Optionally, S2 includes:

[0019] S21. Based on the boiler flow chart and equipment ledger information, a structured component mapping table is constructed. The component mapping table uniquely maps each physical measurement point and each control actuator in the preprocessed multi-source boiler operation dataset to a set of nodes in the graph structure.

[0020] S22. Construct a physical adjacency set based on the physical contact relationships, heat exchange paths, and gas-solid flow paths between boiler equipment;

[0021] S23. Construct thermal coupling edge sets based on the thermal-engineering coupling relationship in the boiler combustion process;

[0022] S24. Based on the boiler control logic, the path through which the control actuator acts on multiple measuring points is defined as the control link, and a control coupling edge set is constructed;

[0023] S25. Merge the physical adjacency edge set, the thermal coupling edge set, and the control coupling edge set to form a complete edge set. Combine this with the node set to construct the initial graph structure of the spatiotemporal heterogeneity of boiler combustion.

[0024] Optionally, S3 includes:

[0025] S31. Input the initial graph structure of the spatiotemporal heterogeneous boiler combustion and the preprocessed multi-source boiler operation dataset into the initialization module of the Graph Sparse Training model. The initialization module generates an initial adjacency matrix for each pair of nodes based on the edge set. The initial adjacency matrix between each pair of connected nodes is assigned the edge weight, and the initial adjacency matrix between each pair of unconnected nodes is assigned zero.

[0026] S32. Initialize the sparse structure mask matrix. If the edge between a pair of nodes is preserved, the corresponding element in the sparse structure mask matrix is ​​set to one. If the edge between a pair of nodes is not preserved, the corresponding element in the sparse structure mask matrix is ​​set to zero.

[0027] S33. By multiplying the structural sparse mask matrix element by element with the initial adjacency matrix, a structural sparse adjacency matrix is ​​obtained. The structural sparse adjacency matrix is ​​then symmetrically normalized to obtain a normalized sparse adjacency matrix.

[0028] S34. Initialize the gradient sparse mask matrix. In each training round, for any row and column of the gradient sparse mask matrix, if the absolute value of the gradient of the edge corresponding to the current position on the loss function is greater than or equal to the dynamic gradient importance threshold corresponding to the current position, then the corresponding position is assigned a value of one in the gradient sparse mask matrix; otherwise, the corresponding position is assigned a value of zero in the gradient sparse mask matrix.

[0029] S35. Perform element-wise multiplication of the normalized sparse adjacency matrix with the gradient sparse mask matrix trained in the current round to obtain the double sparse adjacency matrix.

[0030] S36. Based on the dual sparse adjacency matrix and the feature set corresponding to the preprocessed multi-source boiler operation dataset, the initial configuration of the spatiotemporal heterogeneous initial graph structure of boiler combustion is performed to obtain the Graph Sparse Training initial model.

[0031] Optionally, S4 includes:

[0032] S41. Based on the dual sparse adjacency matrix and node feature set in the initial model of Graph Sparse Training, calculate the structural sensitivity index for each edge;

[0033] S42. Calculate the temporal mutual information index for each pair of nodes, which represents the dynamic response correlation strength of the observation sequences of the two nodes within the selected time window;

[0034] S43. The structural sensitivity index and the temporal mutual information index are fused in a weighted linear combination manner to obtain the retention comprehensive score of each edge;

[0035] S44. Sort all edges according to the overall retention score, remove the edges whose overall retention score is in the reciprocal percentage, and obtain the pruned adjacency matrix;

[0036] S45. Perform structural growth operations based on the pruned adjacency matrix to obtain the growth score;

[0037] S46. Sort all candidate edges according to their growth scores, select the top certain proportion of candidate edges with the highest scores, and add the corresponding edges to the adjacency matrix to form the adjacency matrix after the structure update.

[0038] S47. Input the updated adjacency matrix into the node representation encoder, and perform graph embedding operation together with the node feature set to obtain the node embedding representation set.

[0039] Optionally, S5 includes:

[0040] S51. Input the node embedding representation set into the structure decoder, input the embedding feature vectors of any pair of nodes into the structure decoder to perform similarity operation, and obtain the predicted edge weights between the pair of nodes.

[0041] S52. Arrange the predicted edge weights of all node pairs according to the node order and matrix dimension consistent with the dynamic sparse adjacency matrix to form the structural decoder prediction adjacency matrix.

[0042] S53. Subtract each element in the dynamic sparse adjacency matrix from the predicted adjacency matrix of the structure decoder to obtain the structure reconstruction error matrix;

[0043] S54. Take the absolute value of all error elements in the structural reconstruction error matrix, perform a weighted summation, and sum the absolute errors of all node pairs to obtain the scalar value of the structural reconstruction error.

[0044] Optionally, S6 includes:

[0045] S61. Input the node embedding representation set into the time-series decoder, take the embedding feature vector of any node as the input of the time-series decoder, and output the predicted observation trajectory of the node within a predefined time window;

[0046] S62. Extract the true multivariate observation trajectory of each node within the same time window from the preprocessed multi-source boiler operation dataset;

[0047] S63. For each node in the node set, subtract each corresponding sequence element in the predicted observation trajectory from the actual observation trajectory of the node to obtain the node-level time series reconstruction error sequence;

[0048] S64. Take the absolute value of each error element in the node-level time series reconstruction error sequence, and sum the weighted error elements to obtain the node-level time series reconstruction error scalar value.

[0049] S65. Sum the node-level timing reconstruction error scalar values ​​of all nodes to obtain the overall timing reconstruction error scalar value.

[0050] Optionally, S7 includes:

[0051] S71. For each node in the node set, subtract the actual observation data of the node within the same short window time period from the short-term prediction results corresponding to the node embedding representation to form the short-term prediction residual sequence of the node.

[0052] S72. Take the absolute value of each element in the short-term prediction residual sequence of the node and sum them up by weight to obtain the scalar value of the short-term prediction residual of the node;

[0053] S73. The scalar values ​​of structural reconstruction error, overall temporal reconstruction error, and node-level short-term prediction residuals are weighted and summed to obtain the node-level initial anomaly score.

[0054] Optionally, S8 includes:

[0055] S81. Extract the thermal balance boundary conditions of the industrial boiler and calculate the thermal balance deviation of the nodes within the current time window;

[0056] S82. Extract the boundary conditions of the excess air coefficient of industrial boilers, establish the boundary penalty function of the air coefficient, and calculate the degree of deviation of oxygen content in the observation area where the node is located;

[0057] S83. Extract the boundary conditions of the path lag time of industrial boiler flue gas, establish the path lag penalty function, and calculate the path lag deviation value of the downstream measuring point relative to the upstream measuring point of the path to which the node belongs.

[0058] S84. The physical consistency penalty weight of the node is obtained by weighting and summing the thermal balance deviation value, the oxygen content deviation value and the path lag deviation value.

[0059] S85. The initial anomaly score of the node is fused and calibrated with the integrated physical consistency penalty weight to generate a node-level calibration anomaly score.

[0060] Optionally, S9 includes:

[0061] S91. Take the node-level calibration anomaly score of each node as the initial heat value input, and combine it with the connection relationship between each node in the dynamic sparse graph structure and the normalized sparse adjacency matrix to construct the attribution heat matrix.

[0062] S92. Sort all attribution heat values ​​in the attribution heat matrix from high to low according to the heat value to obtain a connection subgraph, which is used to represent the minimum causal link leading to the high anomaly region;

[0063] S93. By combining the node index, connecting edges and their abnormal attribution strength in the minimum causal link, the attribution heat value of all nodes in the minimum causal link is compared, and the node with the highest attribution heat value is determined as the root cause node. The attribution heat value of all edges in the minimum causal link is compared, and the edge with the highest attribution heat value is determined as the root cause edge.

[0064] S94. When the node-level calibration anomaly score of at least one node is detected to be greater than the adaptive operating condition threshold, it is determined that there is an anomaly in the boiler combustion state.

[0065] Optionally, if the boiler combustion state is abnormal, the following judgment result will be output:

[0066] Abnormal location: Identify the boiler physical components or monitoring points corresponding to all nodes whose node-level calibration abnormality scores are greater than the adaptive operating condition threshold, as the specific location where the abnormality occurred;

[0067] Critical Anomaly Links: List all critical nodes and critical edges contained in the minimum causal link subgraph, indicating the propagation path of the impact of the anomaly and the possible physical and thermal coupling relationships;

[0068] Root cause determination: Mark the node with the highest attribution heat value in the minimum causal link subgraph as the suspected abnormal root cause node, and the edge with the highest attribution heat value as the suspected abnormal root cause link, and give the equipment or measurement point name of the node and edge in the actual process of the boiler system.

[0069] Warning level output: Based on the node-level calibration anomaly score and attribution heat value, and according to the set grading threshold, the anomaly warning level is output, including general anomaly, severe anomaly and emergency anomaly.

[0070] The beneficial effects of this invention are:

[0071] (1) This invention maps the multi-source measuring points and actuators of the boiler system into a spatiotemporal heterogeneous graph structure. It adopts a pruning and structural growth mechanism based on structural sensitivity index and temporal mutual information index to achieve intelligent sparse retention and efficient removal of redundant connections in key thermal links. At the same time, structural sparse mask and gradient sparse mask are introduced in the training stage of graph neural network to form a double sparse adjacency matrix. This enables the model to complete the propagation of high-dimensional coupled information and graph embedding calculation with only a few effective connections under limited computing power and edge computing conditions, which significantly improves the real-time performance of anomaly detection and the throughput under high load conditions.

[0072] (2) This invention constructs a dual-channel reconstruction network based on node embedding structure decoder and time sequence decoder. The structure decoder reconstructs the adjacency relationship and edge weight distribution of the dynamic sparse graph, and the time sequence decoder reconstructs the multivariate observation trajectory of each node within a short window. The structure reconstruction error and the time sequence reconstruction error are calculated respectively. The two types of reconstruction errors are fused with the short-term prediction residual to form the node-level initial anomaly score, providing accurate and auditable anomaly diagnosis results for boiler operation and maintenance. Attached Figure Description

[0073] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0074] Figure 1 This is a flowchart of a boiler combustion state anomaly early warning method based on graph neural networks proposed in this invention. Detailed Implementation

[0075] Example 1: Reference Figure 1 A method for early warning of abnormal boiler combustion status based on graph neural networks includes:

[0076] S1. Collect multi-source boiler operation datasets and preprocess them to obtain preprocessed multi-source boiler operation datasets;

[0077] Real-time measurement data from the boiler control system, offline measurement data from the historical database, and actuator control command data are collected and merged to form a multi-source boiler operation dataset. The multi-source boiler operation dataset is then cleaned, time-aligned, has missing measurement interpolation, and undergoes initial outlier suppression to obtain a preprocessed multi-source boiler operation dataset.

[0078] S2. Based on the boiler flow chart and equipment ledger information, map each measuring point and actuator in the preprocessed multi-source boiler operation data to a node in the graph structure, and map the physical adjacency relationship, thermal coupling relationship and control link relationship to the edge in the graph structure to construct the initial graph structure of boiler combustion spatiotemporal heterogeneity.

[0079] In this embodiment, S2 includes:

[0080] S21. Based on the boiler flow chart and equipment ledger information, a structured component mapping table is constructed. The component mapping table uniquely maps each physical measurement point and each control actuator in the preprocessed multi-source boiler operation dataset to a set of nodes in the graph structure.

[0081] In Example 1, the boiler flow diagram is obtained by retrieving standard equipment process flow diagram files, on-site technical briefing documents, or P&ID built into the automation system provided by the industrial boiler manufacturer; the equipment ledger information is obtained by querying the equipment asset management system at the boiler operation site or the factory equipment archives. The equipment ledger information includes all measuring points, actuators, equipment numbers and their physical attributes, installation locations and measurement / control types in the boiler system.

[0082] The node set includes several nodes, each representing an actual measuring point or an actual actuator device. Each node also includes its corresponding feature vector in the preprocessed multi-source boiler operation dataset, with the feature vector having the same dimension as the node's original feature dimension.

[0083] S22. Construct a physical adjacency set based on the physical contact relationships, heat exchange paths, and gas-solid flow paths between boiler equipment;

[0084] In Example 1, referring to the boiler flow diagram and the actual physical location and interface number of each piece of equipment recorded in the equipment ledger, it is determined whether any two nodes are directly connected through actual pipes, flanges, air ducts, or heating surface structures. If they are directly connected, a physical adjacency edge is established between the two corresponding nodes in the node set. Based on the heat exchange path marked in the boiler system, equipment pairs through which heat exchange media flow are considered as physical adjacencies with direct heat transfer, and a physical adjacency edge is established between the two nodes. Based on the main flow path of gas and solid media in the boiler process, physical adjacency edges are established for nodes that are sequentially connected to equipment in the main flow path. All node pairs that satisfy the relationships of direct physical connection, direct heat transfer, or direct continuity of flow path are established with physical adjacency edges, forming a physical adjacency edge set. Each physical adjacency edge in the physical adjacency edge set represents a direct physical connection between one node and another node in the node set.

[0085] S23. Construct thermal coupling edge sets based on the thermal-engineering coupling relationship in the boiler combustion process;

[0086] In Example 1, the thermal coupling relationship includes the flue gas temperature gradient, the coupled heat exchange between the convection section and the radiation section, and the thermal disturbance information of the burner to the furnace area. Each thermal coupling edge in the thermal coupling edge set represents a significant thermal dynamic response relationship between one node and another node in the node set. Each thermal coupling edge in the thermal coupling edge set is accompanied by a thermal coupling correlation coefficient.

[0087] The thermal coupling correlation coefficient is obtained by the ratio of the product of the covariance and the standard deviation of the flue gas temperature series corresponding to the node in the multi-source boiler operation dataset. Specifically, the covariance of the flue gas temperature series corresponding to the node and the other node is used as the numerator, and the product of the standard deviations of the flue gas temperature series corresponding to the two nodes is used as the denominator. The ratio of these two values ​​is the thermal coupling correlation coefficient.

[0088] S24. Based on the boiler control logic, the path through which the control actuator acts on multiple measuring points is defined as the control link, and a control coupling edge set is constructed;

[0089] In Example 1, each control coupling edge in the control coupling edge set represents a control adjustment relationship between the controller node and the controlled measurement point node. Each control coupling edge is accompanied by a control response strength. The control response strength is obtained by taking the absolute value of the partial derivative of the response variable of the controlled measurement point node with respect to the input variable of the control actuator node. That is, with the input variable of the control actuator node as the independent variable and the response variable of the controlled measurement point node as the dependent variable, the sensitivity of the response variable when the input variable changes is calculated and the absolute value is taken as the control response strength.

[0090] S25. Merge the physical adjacency edge set, the thermal coupling edge set, and the control coupling edge set to form a complete edge set. Combine this with the node set to construct the initial graph structure of the spatiotemporal heterogeneity of boiler combustion.

[0091] The initial graph structure of the spatiotemporal heterogeneous boiler combustion is uniquely determined by the set of nodes and the complete set of edges.

[0092] S3. Initialize the parameters of the Graph Sparse Training dynamic sparse model using the initial graph structure of the spatiotemporal heterogeneity of boiler combustion and the preprocessed multi-source boiler operation dataset to obtain the initial Graph Sparse Training model;

[0093] In this embodiment, S3 includes:

[0094] S31. Input the initial graph structure of the spatiotemporal heterogeneous boiler combustion and the preprocessed multi-source boiler operation dataset into the initialization module of the Graph Sparse Training model. The initialization module generates an initial adjacency matrix for each pair of nodes based on the edge set. The initial adjacency matrix between each pair of connected nodes is assigned the edge weight, and the initial adjacency matrix between each pair of unconnected nodes is assigned zero.

[0095] Each feature vector in the preprocessed multi-source boiler operation dataset is bound one-to-one with the corresponding node in the spatiotemporal heterogeneous initial graph structure of boiler combustion, and the edge weights are derived from the thermal coupling correlation coefficient or the control response intensity.

[0096] S32. Initialize the sparse structure mask matrix. If the edge between a pair of nodes is preserved, the corresponding element in the sparse structure mask matrix is ​​set to one. If the edge between a pair of nodes is not preserved, the corresponding element in the sparse structure mask matrix is ​​set to zero.

[0097] The sparse structural mask matrix is ​​a two-dimensional matrix, where each element takes the value of one or zero. The sparse structural mask matrix is ​​used to identify which connections between nodes are preserved.

[0098] S33. By multiplying the structural sparse mask matrix element by element with the initial adjacency matrix, a structural sparse adjacency matrix is ​​obtained. The structural sparse adjacency matrix is ​​then symmetrically normalized to obtain a normalized sparse adjacency matrix.

[0099] Normalized sparse adjacency matrices are used to explicitly control which connections participate in propagation calculations.

[0100] S34. Initialize the gradient sparse mask matrix. In each training round, for any row and column of the gradient sparse mask matrix, if the absolute value of the gradient of the edge corresponding to the current position on the loss function is greater than or equal to the dynamic gradient importance threshold corresponding to the current position, then the corresponding position is assigned a value of one in the gradient sparse mask matrix; otherwise, the corresponding position is assigned a value of zero in the gradient sparse mask matrix.

[0101] In Example 1, the loss function is used to measure the combined difference between the prediction results of the initial model in Graph Sparse Training and the actual observed data and physical prior constraints. The gradient sparse mask matrix is ​​a square matrix consisting of zero and one values, with the number of rows and columns equal to the size of the node set. The absolute value of the gradient refers to the absolute value of the partial derivative of the loss function with respect to the weight variables of the edges. The dynamic gradient importance threshold is a threshold set for each edge in the current training round to determine the importance of the edge in this round of training.

[0102] S35. Perform element-wise multiplication of the normalized sparse adjacency matrix with the gradient sparse mask matrix trained in the current round to obtain the double sparse adjacency matrix.

[0103] Each element in the double sparse adjacency matrix is ​​equal to the product of the element at the corresponding position in the normalized sparse adjacency matrix and the element at the same position in the gradient sparse mask matrix. The double sparse adjacency matrix represents the effective adjacency relationship that ultimately participates in the graph neural network propagation calculation after taking into account both static structural sparsity and dynamic gradient sparsity in the current training round.

[0104] S36. Based on the dual sparse adjacency matrix and the feature set corresponding to the preprocessed multi-source boiler operation dataset, the initial configuration of the spatiotemporal heterogeneous initial graph structure of boiler combustion is performed to obtain the Graph Sparse Training initial model.

[0105] In Example 1, the dual sparse adjacency matrix is ​​used as the effective connection relationship between nodes at the current time and input into the graph neural network. The feature vector of each node is extracted from the feature set of the preprocessed multi-source boiler operation dataset and assigned to the corresponding node in the spatiotemporal heterogeneous initial graph structure of boiler combustion. The dual sparse adjacency matrix is ​​used as the information transmission weight to control whether nodes participate in information interaction and the interaction intensity in each round of graph neural network propagation calculation. The feature set is used as the node attribute to participate in the initialization of node embedding representation. The dual sparse adjacency matrix and the feature set are applied together to the spatiotemporal heterogeneous initial graph structure of boiler combustion. Through the initialization of graph neural network parameters and the setting of the structure, the Graph Sparse Training initial model that integrates static structural sparsity and dynamic gradient sparsity is obtained.

[0106] S4. On the initial Graph Sparse Training model, pruning and structural growth operations are performed based on structural sensitivity index and temporal mutual information index. Thermal critical path edges are retained to generate a dynamic sparse graph structure. The dynamic sparse graph structure is input into the node representation encoder to obtain the node embedding representation.

[0107] In this embodiment, S4 includes:

[0108] S41. Based on the dual sparse adjacency matrix and node feature set in the initial model of Graph Sparse Training, calculate the structural sensitivity index for each edge;

[0109] The structural sensitivity index refers to the absolute gradient sensitivity of the corresponding edge to the loss function. It is obtained by taking the partial derivatives of the edge weight variables under the loss function and then taking the absolute value as the structural sensitivity index of the edge.

[0110] S42. Calculate the temporal mutual information index for each pair of nodes, which represents the dynamic response correlation strength of the observation sequences of the two nodes within the selected time window;

[0111] The observation subsequence of each pair of nodes within the time window is used as input. The joint probability distribution and the marginal probability distribution of all value combinations are statistically analyzed. The mutual information between the observations of the two nodes is calculated using the joint probability distribution and the marginal probability distribution. The mutual information is used as the time series mutual information index.

[0112] S43. The structural sensitivity index and the temporal mutual information index are fused in a weighted linear combination manner to obtain the retention comprehensive score of each edge;

[0113] S44. Sort all edges according to the overall retention score, remove the edges whose overall retention score is in the reciprocal percentage, and obtain the pruned adjacency matrix;

[0114] S45. Perform structural growth operations based on the pruned adjacency matrix to obtain the growth score;

[0115] The first step in the structure growth operation is to filter out the set of candidate edges that were not retained from the original adjacency matrix. For each candidate edge, its growth score is calculated. The growth score is a score obtained by fusing the structure sensitivity index and the temporal mutual information index in a weighted linear combination.

[0116] S46. Sort all candidate edges according to their growth scores, select the top certain proportion of candidate edges with the highest scores, and add the corresponding edges to the adjacency matrix to form the adjacency matrix after the structure update.

[0117] In Example 1, to prevent thermal critical path edges from being mistakenly pruned or omitted during pruning or structural growth, a set of thermal critical path edges is established, and the weights of all edges belonging to the thermal critical path edge set are forcibly assigned to the corresponding weights in the adjacency matrix after structural update.

[0118] S47. Input the updated adjacency matrix into the node representation encoder, and perform graph embedding operation together with the node feature set to obtain the node embedding representation set.

[0119] In Example 1, the updated adjacency matrix is ​​used as the input to the node representation encoder as the connection relationship between nodes. The feature vector corresponding to each node in the preprocessed multi-source boiler operation dataset is used as the node input attribute. The node representation encoder performs the following operation on each node: Based on the updated adjacency matrix, the node representation encoder performs weighted aggregation on the feature vector of each node and the feature vectors of its neighboring nodes according to the corresponding connection weights. The weighted aggregated feature vectors are then subjected to linear transformation and nonlinear activation to obtain the embedded feature vector of each node. The embedded feature vectors of all nodes form the node embedding representation set. Each node embedding feature vector corresponds one-to-one with a node in the updated adjacency matrix.

[0120] S5. Based on node embedding representation, the adjacency relationship and edge weight distribution of the dynamic sparse graph structure are reconstructed through the structure decoder to obtain the scalar value of the structure reconstruction error;

[0121] In this embodiment, S5 includes:

[0122] S51. Input the node embedding representation set into the structure decoder, input the embedding feature vectors of any pair of nodes into the structure decoder to perform similarity operation, and obtain the predicted edge weights between the pair of nodes.

[0123] S52. Arrange the predicted edge weights of all node pairs according to the node order and matrix dimension consistent with the dynamic sparse adjacency matrix to form the structural decoder prediction adjacency matrix.

[0124] Each element in the structural decoder's predicted adjacency matrix represents the predicted connection relationship and predicted edge weight between the node pairs, and corresponds one-to-one with the corresponding position in the dynamic sparse adjacency matrix.

[0125] S53. Subtract each element in the dynamic sparse adjacency matrix from the predicted adjacency matrix of the structure decoder to obtain the structure reconstruction error matrix;

[0126] Each element in the structure reconstruction error matrix represents the difference between the actual dynamic sparse edge weights of the corresponding node pair and the predicted edge weights of the structure decoder. The structure reconstruction error matrix is ​​used to represent the deviation of the structure decoder when reconstructing the structure of the dynamic sparse graph.

[0127] S54. Take the absolute value of all error elements in the structural reconstruction error matrix, perform a weighted summation, and sum the absolute errors of all node pairs to obtain the scalar value of the structural reconstruction error.

[0128] The structural reconstruction error scalar value represents the overall reconstruction error of the structure decoder for the dynamic sparse graph structure.

[0129] S6. Based on node embedding representation, the short-window multivariate observation trajectory of each node in the dynamic sparse graph structure is reconstructed through a temporal decoder to obtain the overall temporal reconstruction error scalar value;

[0130] In this embodiment, S6 includes:

[0131] S61. Input the node embedding representation set into the time-series decoder, take the embedding feature vector of any node as the input of the time-series decoder, and output the predicted observation trajectory of the node within a predefined time window;

[0132] The predicted observation trajectory represents the reconstruction result of multivariate observation data of a node within a short window time period. The time window length is consistent with that of the short window observation. The predicted observation trajectory corresponds one-to-one with the actual observation trajectory in terms of sequence length and variable dimension.

[0133] S62. Extract the true multivariate observation trajectory of each node within the same time window from the preprocessed multi-source boiler operation dataset;

[0134] The actual observation trajectory represents the actual observed values ​​of each variable at a node within a short window period. The actual observation trajectory and the predicted observation trajectory correspond one-to-one in terms of node order, time order, and variable order.

[0135] S63. For each node in the node set, subtract each corresponding sequence element in the predicted observation trajectory from the actual observation trajectory of the node to obtain the node-level time series reconstruction error sequence;

[0136] Each element in the node-level time series reconstruction error sequence represents the difference between the actual observation and the predicted observation at the corresponding node at the corresponding time point.

[0137] S64. Take the absolute value of each error element in the node-level time series reconstruction error sequence, and sum the weighted error elements to obtain the node-level time series reconstruction error scalar value.

[0138] The node-level time series reconstruction error scalar value represents the overall deviation of the node's observation trajectory reconstruction within a short window period.

[0139] S65. Sum the node-level timing reconstruction error scalar values ​​of all nodes to obtain the overall timing reconstruction error scalar value.

[0140] The overall temporal reconstruction error scalar value measures the overall reconstruction deviation of the temporal decoder for the multivariate observation trajectories of all nodes within a short window time period.

[0141] S7. Calculate the initial anomaly score at the node level based on the scalar value of the structural reconstruction error, the scalar value of the overall temporal reconstruction error, and the short-term prediction residual;

[0142] In this embodiment, S7 includes:

[0143] S71. For each node in the node set, subtract the actual observation data of the node within the same short window time period from the short-term prediction results corresponding to the node embedding representation to form the short-term prediction residual sequence of the node.

[0144] Each element of the short-term prediction residual sequence represents the difference between the actual observation and the predicted observation at a certain moment within the time window.

[0145] S72. Take the absolute value of each element in the short-term prediction residual sequence of the node and sum them up by weight to obtain the scalar value of the short-term prediction residual of the node;

[0146] The node-level short-term prediction residual scalar value represents the magnitude of the prediction error of a node within a short window period.

[0147] S73. The scalar values ​​of structural reconstruction error, overall temporal reconstruction error, and node-level short-term prediction residuals are weighted and summed to obtain the node-level initial anomaly score.

[0148] S8. Perform physical consistency calibration between the node-level initial anomaly score and the industrial boiler thermal balance boundary conditions, air excess coefficient boundary conditions, and flue gas path lag time boundary conditions to generate node-level calibration anomaly scores.

[0149] In this embodiment, S8 includes:

[0150] S81. Extract the thermal balance boundary conditions of the industrial boiler and calculate the thermal balance deviation of the nodes within the current time window;

[0151] The heat input of fuel and the effective heat absorbed by the region per unit time in the statistical node are calculated. The absolute value of the difference between the heat input of fuel and the effective heat absorbed is then divided by the heat input of fuel to obtain the heat balance deviation value of the node. The larger the heat balance deviation value, the greater the heat balance deviation.

[0152] S82. Extract the boundary conditions of the excess air coefficient of industrial boilers, establish the boundary penalty function of the air coefficient, and calculate the degree of deviation of oxygen content in the observation area where the node is located;

[0153] The calculation method of the air coefficient boundary penalty function is as follows: obtain the real-time air excess coefficient of the region corresponding to the node. If the real-time air excess coefficient is within the safe interval boundary, the air coefficient boundary penalty function value is zero; if the real-time air excess coefficient exceeds the safe interval boundary, calculate the difference between the actual air excess coefficient and the upper and lower boundaries of the interval, and take the absolute value of the smaller distance as the degree of deviation of the node's oxygen content.

[0154] S83. Extract the boundary conditions of the path lag time of industrial boiler flue gas, establish the path lag penalty function, and calculate the path lag deviation value of the downstream measuring point relative to the upstream measuring point of the path to which the node belongs.

[0155] The path lag penalty function is calculated as follows: obtain the actual lag time of the flue gas temperature response from the upstream node to the corresponding node, query the standard lag time in the boiler design file, take the absolute value of the difference between the actual lag time and the standard lag time, and then divide it by the standard lag time to obtain the path lag deviation value of the node.

[0156] S84. The physical consistency penalty weight of the node is obtained by weighting and summing the thermal balance deviation value, the oxygen content deviation value and the path lag deviation value.

[0157] S85. The initial anomaly score of the node is fused and calibrated with the integrated physical consistency penalty weight to generate a node-level calibration anomaly score.

[0158] S9. Perform attribution heat backtracking on the node-level calibration anomaly score on the dynamic sparse graph structure, extract the minimum causal link that leads to the anomaly, and locate the potential root cause node and root cause edge to form the boiler combustion state anomaly detection result.

[0159] In this embodiment, S9 includes:

[0160] S91. Take the node-level calibration anomaly score of each node as the initial heat value input, and combine it with the connection relationship between each node in the dynamic sparse graph structure and the normalized sparse adjacency matrix to construct the attribution heat matrix.

[0161] The node-level calibration anomaly score of each node is used as the initial heat value input into the dynamic sparse graph structure. Based on the normalized sparse adjacency matrix, the initial heat value of each node is distributed to its directly connected neighbor nodes according to the connection weight between nodes. During the heat distribution process, the heat value received by each node is equal to the sum of the products of the initial heat values ​​of all its directly connected neighbor nodes and the adjacency weights. After each round of propagation, it is multiplied by a preset attenuation coefficient. The heat propagation process is iterated until the heat value change is lower than the preset convergence threshold or the number of propagation rounds reaches the set maximum value. After the attribution heat propagation iteration, the attribution heat value between each pair of nodes is the corresponding element of the attribution heat matrix. The attribution heat matrix is ​​used to represent the anomaly attribution strength between a pair of nodes.

[0162] S92. Sort all attribution heat values ​​in the attribution heat matrix from high to low according to the heat value to obtain a connection subgraph, which is used to represent the minimum causal link leading to the high anomaly region;

[0163] All attribution heat values ​​in the attribution heat matrix are sorted from high to low according to their heat values. Based on the principle that the cumulative attribution heat reaches the set percentage of the total attribution heat, the connection edge with the highest attribution heat value and its corresponding node are selected in turn until the cumulative attribution heat reaches the set percentage of the total attribution heat. All selected connection edges and corresponding nodes are combined into a connection subgraph. The minimum causal link consists of several key nodes and key edges. The connection subgraph covers the most influential local structure in the abnormal attribution propagation path.

[0164] S93. By combining the node index, connecting edges and their abnormal attribution strength in the minimum causal link, the attribution heat value of all nodes in the minimum causal link is compared, and the node with the highest attribution heat value is determined as the root cause node. The attribution heat value of all edges in the minimum causal link is compared, and the edge with the highest attribution heat value is determined as the root cause edge.

[0165] S94. When the node-level calibration anomaly score of at least one node is detected to be greater than the adaptive operating condition threshold, it is determined that there is an anomaly in the boiler combustion state.

[0166] In this embodiment, an abnormality occurs in the boiler combustion state, and the following judgment result is output:

[0167] Abnormal location: Identify the boiler physical components or monitoring points corresponding to all nodes whose node-level calibration abnormality scores are greater than the adaptive operating condition threshold, as the specific location where the abnormality occurred;

[0168] Critical Anomaly Links: List all critical nodes and critical edges contained in the minimum causal link subgraph, indicating the propagation path of the impact of the anomaly and the possible physical and thermal coupling relationships;

[0169] Root cause determination: Mark the node with the highest attribution heat value in the minimum causal link subgraph as the suspected abnormal root cause node, and the edge with the highest attribution heat value as the suspected abnormal root cause link, and give the equipment or measurement point name of the node and edge in the actual process of the boiler system.

[0170] Warning level output: Based on the node-level calibration anomaly score and attribution heat value, and according to the set grading threshold, the anomaly warning level is output, including general anomaly, severe anomaly and emergency anomaly.

[0171] Example 2: In the boiler operation monitoring center, the system continuously receives data streams from all key measuring points and actuators. During a certain time period, the coal feed rate of the main coal feeder gradually decreases, causing abnormal fluctuations in furnace temperature and some flue gas parameters. The system receives and preprocesses all measuring point and actuator data in real time, including observed values ​​such as temperature, pressure, flow rate, oxygen content, damper opening, and flue gas composition. The samples are numbered from T10001 to T10600, totaling 600 time slices, each containing 17-dimensional observation data from 210 nodes. After standardization, denoising, and missing measurement completion, this data is used to construct a structured node set and a multi-source boiler operation dataset. All nodes have completed feature vector binding.

[0172] Based on the flowchart and equipment register, the system automatically maps equipment such as furnace area measuring points, main flue measuring points, damper actuators at each level, and coal feeder actuators as nodes, connecting them to form a complete dynamic spatiotemporal heterogeneous initial graph structure. The physical adjacency relationships of edges are confirmed by P&ID and historical maintenance data. The thermal coupling weights are dynamically assigned based on historical covariance and temperature correlation, while the control coupling weights are estimated online based on the sensitivity of the main coal feed rate to furnace temperature and oxygen content. For example, the dynamically estimated control edge weight from the main coal feeder to the lower furnace temperature measuring point is 0.82, and the thermal coupling edge weight from the main flue to the outlet pressure measuring point is 0.73.

[0173] Building upon this, the Graph Sparse Training module utilizes a comprehensive scoring system combining the structural sensitivity and temporal mutual information of node pairs to automatically prune edges that do not participate in the main energy flow. After a certain round of pruning, the edge between node v17 (main coal feeder) and node v23 (lower furnace temperature) is forcibly retained due to its high sensitivity, while the edge between node v85 (secondary air blower) and node v57 (boiler sidewall measuring point) is pruned due to redundancy. After normalization of the sparse adjacency matrix, the dynamic propagation weight is reduced to 0.36.

[0174] The system continuously monitors the evolution of node embedding features. The structural decoder calculates the similarity of the connection weights between node v23 and its neighboring nodes. The structural decoder predicts an edge weight of 0.69 between nodes v23 and v25, while the actual value in the dynamic sparse adjacency matrix is ​​0.72, a difference of -0.03, which is recorded in the structural reconstruction error matrix. The temporal decoder reconstructs data such as temperature, oxygen content, and flue gas flow rate for node v23 over the most recent 60 time slices. The weighted sum of the absolute values ​​of the reconstruction error sequence is 7.2.

[0175] The scalar value of the prediction residual within a certain window between the actual observed value and the short-term prediction result of node v23 (lower furnace temperature) is calculated to be 3.1. The system automatically summarizes the structural reconstruction error, temporal reconstruction error, and prediction residual of node v23, and obtains an initial anomaly score of 8.2 through weighting. At the same time, the initial anomaly scores of nodes v17 (main coal feeder), v41 (secondary air outlet pressure), and v63 (main flue gas measuring point) are calculated to be 7.9, 6.5, and 7.2, respectively.

[0176] After the system acquired the thermal balance boundary conditions, the fuel input heat per unit time in the region corresponding to node v23 was 3120 kJ, the absorbed effective heat was 2980 kJ, and the thermal balance deviation was 4.49%. The real-time excess air coefficient measured by node v23 was 1.32, with a safe range of [1.1, 1.3], therefore the oxygen content deviation was 0.02. The measured flue gas path lag time was 7.1 seconds, while the design value was 6.8 seconds, resulting in a lag deviation of 4.4%. The physical consistency penalty weights were calculated using weighted averages of 0.4, 0.3, and 0.3, ultimately reaching 2.27. The calibration anomaly score of node v23 reached 10.5 after fusion calibration, significantly exceeding the adaptive operating condition threshold of 7.0.

[0177] At this point, the nodes whose calibration anomaly scores exceed the threshold are v17, v23, v41, and v63, respectively. The set of nodes in the anomaly region is determined. The system uses the calibration anomaly scores of all nodes as the initial heat value and employs a normalized sparse adjacency matrix for attribution heat propagation. The attribution heat values ​​between nodes v17, v23, v41, and v63 converge to 0.28, 0.33, 0.24, and 0.18 respectively under the propagation model, forming a high attribution heat value link. When the cumulative attribution heat value reaches 80% of the total heat value after sorting, the required critical nodes are v17, v23, and v41, and the critical edges are v17—v23 (main coal feeder—lower furnace temperature) and v23—v41 (lower furnace temperature—secondary tuyer pressure).

[0178] The minimum causal link identifies the above three nodes and two key edges. After comparison, the node with the highest heat in the link is v23, which is determined to be the suspected root cause node. The edge with the highest heat in the minimum causal link is v17-v23, which is determined to be the suspected root cause link. The system automatically generates a boiler combustion status anomaly detection report, which includes: the anomaly type is abnormal temperature in the lower part of the furnace caused by a sudden reduction in fuel input, the main cause is located in the linkage link between the main coal feeder and the furnace temperature, it is recommended to check the linkage control logic between the coal feeder and the secondary air outlet pressure, and the warning level is marked as severe.

[0179] To further verify the effectiveness, the system evaluated 50 abnormal windows and 50 normal windows randomly selected from 600 time slices using both traditional methods and the method of this invention. The statistical results are as follows:

[0180] The method of this invention detected 47 abnormal windows (3 missed detections), while the traditional threshold method detected only 33 (17 missed detections) and LSTM detected 42 (8 missed detections).

[0181] The false alarm rate of the method of this invention is 4%, while that of the traditional threshold method is 21% and that of LSTM is 11%.

[0182] Regarding lead time, the average advance detection time of the method of this invention is 22 seconds, LSTM is 13 seconds, and the threshold method is 2 seconds.

[0183] The root cause localization accuracy is 90% for this invention, while LSTM and thresholding methods both achieve 0% (unable to provide link attribution).

[0184] The average CPU resource usage during the detection process was 14% for this invention, 29% for LSTM, and 2% for the threshold method.

[0185] Furthermore, the system archives the training samples for each node and each edge throughout the entire process. For example, the reconstruction errors of the actual observed trajectory and reconstructed trajectory samples (temperature: 684, 689, 701, 706, 710; oxygen content: 3.2, 3.1, 3.3, 3.0, 2.9; flue gas flow: 28, 30, 27, 29, 28) for node v23 in the training set are 3, 2, 5, 3, and 2 (all absolute values), respectively. The average structural reconstruction error is 0.032, and the average temporal reconstruction error is 0.031. The node-level initial anomaly score and physical consistency calibration process are complete and traceable. The model automatically completes a small-step structural adjustment every thirty minutes, and the model sparsity rate is maintained at around 70%.

[0186] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of abnormal boiler combustion status based on graph neural networks, characterized in that, include: S1. Collect multi-source boiler operation datasets and preprocess them to obtain preprocessed multi-source boiler operation datasets; S2. Based on the boiler flow chart and equipment ledger information, map each measuring point and actuator in the preprocessed multi-source boiler operation data to a node in the graph structure, and map the physical adjacency relationship, thermal coupling relationship and control link relationship to the edge in the graph structure to construct the initial graph structure of boiler combustion spatiotemporal heterogeneity. S3. Initialize the parameters of the Graph Sparse Training dynamic sparse model using the initial graph structure of the spatiotemporal heterogeneity of boiler combustion and the preprocessed multi-source boiler operation dataset to obtain the initial Graph Sparse Training model; S4. On the initial Graph Sparse Training model, pruning and structural growth operations are performed based on structural sensitivity index and temporal mutual information index. Thermal critical path edges are retained to generate a dynamic sparse graph structure. The dynamic sparse graph structure is input into the node representation encoder to obtain the node embedding representation. S5. Based on node embedding representation, the adjacency relationship and edge weight distribution of the dynamic sparse graph structure are reconstructed through the structure decoder to obtain the scalar value of the structure reconstruction error; S6. Based on node embedding representation, the short-window multivariate observation trajectory of each node in the dynamic sparse graph structure is reconstructed through a temporal decoder to obtain the overall temporal reconstruction error scalar value; S7. Calculate the initial anomaly score at the node level based on the scalar value of the structural reconstruction error, the scalar value of the overall temporal reconstruction error, and the short-term prediction residual; S8. Perform physical consistency calibration between the node-level initial anomaly score and the industrial boiler thermal balance boundary conditions, air excess coefficient boundary conditions, and flue gas path lag time boundary conditions to generate node-level calibration anomaly scores. S9. Perform attribution heat backtracking on the node-level calibration anomaly score on the dynamic sparse graph structure, extract the minimum causal link that leads to the anomaly, and locate the potential root cause node and root cause edge to form the boiler combustion state anomaly detection result.

2. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, The S2 includes: S21. Based on the boiler flow chart and equipment ledger information, a structured component mapping table is constructed. The component mapping table uniquely maps each physical measurement point and each control actuator in the preprocessed multi-source boiler operation dataset to a set of nodes in the graph structure. S22. Construct a physical adjacency set based on the physical contact relationships, heat exchange paths, and gas-solid flow paths between boiler equipment; S23. Construct thermal coupling edge sets based on the thermal-engineering coupling relationship in the boiler combustion process; S24. Based on the boiler control logic, the path through which the control actuator acts on multiple measuring points is defined as the control link, and a control coupling edge set is constructed; S25. Merge the physical adjacency edge set, the thermal coupling edge set, and the control coupling edge set to form a complete edge set. Combine this with the node set to construct the initial spatiotemporal heterogeneous graph structure for boiler combustion.

3. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, S3 includes: S31. Input the initial graph structure of the spatiotemporal heterogeneous boiler combustion and the preprocessed multi-source boiler operation dataset into the initialization module of the GraphSparse Training model. The initialization module generates an initial adjacency matrix for each pair of nodes based on the edge set. The initial adjacency matrix between each pair of connected nodes is assigned the edge weight, and the initial adjacency matrix between each pair of unconnected nodes is assigned zero. S32. Initialize the sparse structure mask matrix. If the edge between a pair of nodes is preserved, the corresponding element in the sparse structure mask matrix is ​​set to one. If the edge between a pair of nodes is not preserved, the corresponding element in the sparse structure mask matrix is ​​set to zero. S33. By multiplying the structural sparse mask matrix element by element with the initial adjacency matrix, a structural sparse adjacency matrix is ​​obtained. The structural sparse adjacency matrix is ​​then symmetrically normalized to obtain a normalized sparse adjacency matrix. S34. Initialize the gradient sparse mask matrix. In each training round, for any row and column of the gradient sparse mask matrix, if the absolute value of the gradient of the edge corresponding to the current position on the loss function is greater than or equal to the dynamic gradient importance threshold corresponding to the current position, then the corresponding position is assigned a value of one in the gradient sparse mask matrix; otherwise, the corresponding position is assigned a value of zero in the gradient sparse mask matrix. S35. Perform element-wise multiplication of the normalized sparse adjacency matrix with the gradient sparse mask matrix trained in the current round to obtain the double sparse adjacency matrix. S36. Based on the dual sparse adjacency matrix and the feature set corresponding to the preprocessed multi-source boiler operation dataset, the initial configuration of the spatiotemporal heterogeneous initial graph structure of boiler combustion is performed to obtain the Graph Sparse Training initial model.

4. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, S4 includes: S41. Based on the dual sparse adjacency matrix and node feature set in the initial model of Graph Sparse Training, calculate the structural sensitivity index for each edge; S42. Calculate the temporal mutual information index for each pair of nodes, which represents the dynamic response correlation strength of the observation sequences of the two nodes within the selected time window; S43. The structural sensitivity index and the temporal mutual information index are fused in a weighted linear combination manner to obtain the retention comprehensive score of each edge; S44. Sort all edges according to the overall retention score, remove the edges whose overall retention score is in the reciprocal percentage, and obtain the pruned adjacency matrix; S45. Perform structural growth operations based on the pruned adjacency matrix to obtain the growth score; S46. Sort all candidate edges according to their growth scores, select the top certain proportion of candidate edges with the highest scores, and add the corresponding edges to the adjacency matrix to form the adjacency matrix after the structure update. S47. Input the updated adjacency matrix into the node representation encoder, and perform graph embedding operation together with the node feature set to obtain the node embedding representation set.

5. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, S5 includes: S51. Input the node embedding representation set into the structure decoder, input the embedding feature vectors of any pair of nodes into the structure decoder to perform similarity operation, and obtain the predicted edge weights between the pair of nodes. S52. Arrange the predicted edge weights of all node pairs according to the node order and matrix dimension consistent with the dynamic sparse adjacency matrix to form the structural decoder prediction adjacency matrix. S53. Subtract each element in the dynamic sparse adjacency matrix from the predicted adjacency matrix of the structure decoder to obtain the structure reconstruction error matrix; S54. Take the absolute value of all error elements in the structural reconstruction error matrix, perform a weighted summation, and sum the absolute errors of all node pairs to obtain the scalar value of the structural reconstruction error.

6. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, S6 includes: S61. Input the node embedding representation set into the time-series decoder, take the embedding feature vector of any node as the input of the time-series decoder, and output the predicted observation trajectory of the node within a predefined time window; S62. Extract the true multivariate observation trajectory of each node within the same time window from the preprocessed multi-source boiler operation dataset; S63. For each node in the node set, subtract each corresponding sequence element in the predicted observation trajectory from the actual observation trajectory of the node to obtain the node-level time series reconstruction error sequence; S64. Take the absolute value of each error element in the node-level time series reconstruction error sequence, and sum the weighted error elements to obtain the node-level time series reconstruction error scalar value. S65. Sum the node-level timing reconstruction error scalar values ​​of all nodes to obtain the overall timing reconstruction error scalar value.

7. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, S7 includes: S71. For each node in the node set, subtract the actual observation data of the node within the same short window time period from the short-term prediction results corresponding to the node embedding representation to form the short-term prediction residual sequence of the node. S72. Take the absolute value of each element in the short-term prediction residual sequence of the node and sum them up by weight to obtain the scalar value of the short-term prediction residual of the node; S73. The scalar values ​​of structural reconstruction error, overall temporal reconstruction error, and node-level short-term prediction residuals are weighted and summed to obtain the node-level initial anomaly score.

8. The method for early warning of abnormal boiler combustion status based on graph neural networks according to claim 1, characterized in that, S8 includes: S81. Extract the thermal balance boundary conditions of the industrial boiler and calculate the thermal balance deviation of the nodes within the current time window; S82. Extract the boundary conditions of the excess air coefficient of industrial boilers, establish the boundary penalty function of the air coefficient, and calculate the degree of deviation of oxygen content in the observation area where the node is located; S83. Extract the boundary conditions of the path lag time of industrial boiler flue gas, establish the path lag penalty function, and calculate the path lag deviation value of the downstream measuring point relative to the upstream measuring point of the path to which the node belongs. S84. The physical consistency penalty weight of the node is obtained by weighting and summing the thermal balance deviation value, the oxygen content deviation value and the path lag deviation value. S85. The initial anomaly score of the node is fused and calibrated with the integrated physical consistency penalty weight to generate a node-level calibration anomaly score.

9. The boiler combustion state anomaly early warning method based on graph neural network according to claim 1, characterized in that, S9 includes: S91. Take the node-level calibration anomaly score of each node as the initial heat value input, and combine it with the connection relationship between each node in the dynamic sparse graph structure and the normalized sparse adjacency matrix to construct the attribution heat matrix. S92. Sort all attribution heat values ​​in the attribution heat matrix from high to low according to the heat value to obtain a connection subgraph, which is used to represent the minimum causal link leading to the high anomaly region; S93. By combining the node index, connecting edges and their abnormal attribution strength in the minimum causal link, the attribution heat value of all nodes in the minimum causal link is compared, and the node with the highest attribution heat value is determined as the root cause node. The attribution heat value of all edges in the minimum causal link is compared, and the edge with the highest attribution heat value is determined as the root cause edge. S94. When the node-level calibration anomaly score of at least one node is detected to be greater than the adaptive operating condition threshold, it is determined that there is an anomaly in the boiler combustion state.

10. The boiler combustion state anomaly early warning method based on graph neural network according to claim 1, characterized in that, The boiler combustion state is abnormal, and the following judgment result is output: Anomaly location: Identify the boiler physical components or monitoring points corresponding to all nodes whose node-level calibration anomaly scores are greater than the adaptive operating condition threshold, as the specific location where the anomaly occurred; Critical Anomaly Links: List all critical nodes and critical edges contained in the minimum causal link subgraph, indicating the propagation path of the impact of the anomaly and the possible physical and thermal coupling relationships; Root cause determination: Mark the node with the highest attribution heat value in the minimum causal link subgraph as the suspected abnormal root cause node, and the edge with the highest attribution heat value as the suspected abnormal root cause link, and give the equipment or measurement point name of the node and edge in the actual process of the boiler system. Warning level output: Based on the node-level calibration anomaly score and attribution heat value, and according to the set grading threshold, the anomaly warning level is output, including general anomaly, severe anomaly and emergency anomaly.

Images