A method and device for detecting and locating multi-parameter correlation abnormalities of a power plant thermal system

Abstract

The application discloses a power plant thermal system multi-parameter correlation anomaly detection and positioning method, and belongs to the technical field of industrial process state monitoring and fault diagnosis. By acquiring historical normal operation data, a multi-source fusion graph is constructed according to a physical connection relationship and statistical correlation; a graph neural network is trained by using normal working condition data to obtain a health benchmark model. Real-time data are acquired, input into the health benchmark model to calculate reconstruction error, and it is determined whether the system is abnormal. When it is determined that the system is abnormal, a real-time fine graph representing a current abnormal state is subjected to multi-stage graph coarsening, coarse graph sequences of different resolutions are generated, abnormal distribution is analyzed layer by layer to position a fault starting node from coarse to fine, and in combination with the correlation change characteristics of the fault starting node and neighbor nodes, the fault type is determined. The application solves the problem that multi-parameter correlation anomalies are difficult to quickly and accurately position.

Description

Technical Field

[0001] This invention relates to the field of industrial process condition monitoring and fault diagnosis technology, and in particular to a method and device for detecting and locating multi-parameter correlation anomalies in power plant thermal systems. Background Technology

[0002] The thermal system is the core carrier of energy conversion and transfer in a power plant, and its operational reliability directly determines the safety and economy of the unit. Equipment within the system is tightly coupled through the flow of the working fluid, and parameters influence each other. When component performance deteriorates, anomalies propagate along the working fluid path, triggering a cascading change at multiple measuring points. Therefore, system-level anomaly detection and localization based on multi-parameter correlation is crucial for achieving intelligent operation and maintenance of thermal systems.

[0003] Currently, data-driven fault early warning methods have become a research hotspot. However, existing models generally suffer from drawbacks when dealing with complex thermodynamic systems, such as isolated modeling dimensions, lack of topological information, and difficulty in anomaly localization. Specifically, this manifests in two technical approaches:

[0004] The first type focuses on single-point threshold alarms or time-series modeling, such as setting upper and lower limits or using long short-term memory networks to predict individual parameters. This method isolates the coupled system into information silos, ignoring physical connections and parameter correlations. In practice, minute leaks are difficult to detect and are prone to being missed; while transient interferences are prone to false alarms, and when a fault occurs, it generates an "alarm storm," making it impossible for operators to quickly locate the source.

[0005] The second approach attempts to introduce multivariate statistical analysis (such as principal component analysis) or machine learning (such as multilayer perceptron) for multi-parameter state estimation, recognizing the correlation between measurement points. However, traditional methods treat measurement points as independent features, ignoring the inherent topological structure of the thermodynamic system. Although some studies have used graph convolutional networks to extract spatial dependencies between variables, achieving better results than traditional methods in fault classification, no effective solution has yet been provided for fault source localization.

[0006] Even if anomalies can be detected, quickly and accurately locating specific devices remains a significant challenge. Existing fault isolation methods largely rely on expert rules or causal models. The former involves cumbersome rule formulation and incomplete coverage, while the latter involves complex modeling and poor generalization. Furthermore, the large scale of thermal system graphs (tens to hundreds of nodes) makes direct analysis computationally complex. While existing research has proposed reducing computational costs through graph coarsening, it has primarily been applied to classification tasks and not to step-by-step localization.

[0007] Furthermore, fault propagation exhibits clear physical laws: equipment performance degradation affects downstream processes along the working fluid flow, forming a propagation pattern; while instrument failures are typically isolated and do not trigger cascading reactions in neighboring nodes. Existing methods have not fully utilized this characteristic to differentiate fault types.

[0008] In summary, existing technologies suffer from inaccurate detection due to fragmented correlations, low diagnostic efficiency due to a lack of localization mechanisms, difficulty in type identification due to failure to consider fault propagation characteristics, or excessive computational complexity that hinders deployment. There is an urgent need for an integrated intelligent diagnostic framework that unifies the representation of physical topology, deeply mines the spatiotemporal correlations of parameters, considers fault propagation characteristics, and enables system-level detection and component-level localization. Summary of the Invention

[0009] The present invention aims to at least partially solve one of the technical problems in the related art.

[0010] To address this, this invention proposes a method for detecting and locating multi-parameter correlation anomalies in power plant thermal systems. By acquiring historical normal operation data, a multi-source fusion graph is constructed based on physical connections and statistical correlations. This multi-source fusion graph under normal operating conditions is then input into a graph neural network for training, resulting in a health benchmark model that learns the spatial correlation characteristics of the normal state. Real-time operating data is acquired, a real-time multi-source fusion graph is constructed, and input into the health benchmark model. The reconstruction error between the predicted output and the real-time data is calculated to determine whether the system is abnormal. When an anomaly is determined, a real-time fine-grained graph representing the current abnormal state is acquired. This graph undergoes multi-level coarsening to generate a sequence of coarse graphs at different resolutions. The anomaly distribution is analyzed layer by layer to locate the fault initiation node from coarse to fine. The fault type is then determined by combining the correlation change characteristics between this node and its neighboring nodes. This invention achieves accurate detection of multi-parameter correlation anomalies and rapid location of fault sources in thermal systems.

[0011] Another objective of this invention is to provide a device for detecting and locating multi-parameter correlation anomalies in a power plant thermal system.

[0012] To achieve the above objectives, this invention proposes a method for detecting and locating multi-parameter correlation anomalies in a power plant thermal system, comprising: Historical normal operation data is obtained, a process topology map is constructed based on the physical connection relationship of the thermal system, and a correlation feature map is constructed based on the statistical correlation of historical normal operation data. The process topology map and the correlation feature map are fused to obtain a multi-source fusion map. Acquire historical operating data and corresponding multi-source fusion graphs under normal operating conditions, input them into a graph neural network model for training, so that the graph neural network model can learn the spatial correlation features between monitoring nodes under normal conditions and obtain a health benchmark model. Acquire real-time operational data, construct a real-time multi-source fusion graph, input the real-time multi-source fusion graph into the health benchmark model, obtain the prediction output, calculate the reconstruction error between the prediction output and the real-time operational data, and output the judgment result of whether the system has an anomaly. When an anomaly is detected in the system, a real-time fine-grained map representing the current abnormal state is obtained. The real-time fine-grained map is then subjected to multi-level graph coarsening to generate a sequence of coarsened maps with different resolutions. The anomaly distribution of each level of coarsened map is analyzed layer by layer to locate the fault initiation node from coarse to fine. The fault type is obtained by combining the correlation and change characteristics between the fault initiation node and its corresponding neighboring nodes.

[0013] In one embodiment of the present invention, the steps of acquiring historical normal operation data, constructing a process topology map based on the physical connection relationship of the thermal system, constructing a correlation feature map based on the statistical correlation of the historical normal operation data, and fusing the process topology map and the correlation feature map to obtain a multi-source fusion map include: Obtain the design drawings and working fluid flow patterns of the thermal system, abstract key equipment as node sets, abstract the connecting pipes between equipment as edge sets, construct the process topology graph, and output the corresponding adjacency matrix. ; Obtain the time series of any two nodes i and j from the historical normal operation data, and calculate the similarity using the Pearson correlation coefficient. The absolute value of the similarity is compared with a set threshold. If it is greater than the threshold, the corresponding position is set to 1; otherwise, it is set to 0. A binary correlation adjacency matrix is ​​constructed as the correlation feature map, and the corresponding adjacency matrix is ​​obtained. ; Set the balance coefficient α, and Multiply by α Multiply by (1) After summing α), we obtain the merged adjacency matrix:

[0014] based on Construct and output a multi-source fusion graph.

[0015] In one embodiment of the present invention, the time series of any two nodes i and j in historical normal operation data are obtained, and the similarity is calculated using the Pearson correlation coefficient. The absolute value of the similarity is compared with a set threshold, including: Get the kth sample value of node i and the corresponding mean The k-th sample value of node j and the corresponding mean Based on the total number of samples N, the relevance feature matrix elements are calculated using the input formula. :

[0016] The calculated The absolute value is compared with a set threshold to obtain a binarized correlation adjacency matrix.

[0017] In one embodiment of the present invention, a balance coefficient is set, a fused adjacency matrix is ​​calculated, and based on... Output a multi-source fusion graph, including: optimizing the balance coefficient α to 0.5, and inputting... and ,calculate Output the merged adjacency matrix .

[0018] In one embodiment of the present invention, the step of acquiring historical operating data and corresponding multi-source fusion graphs under normal operating conditions, and inputting them into a graph neural network model for training, so that the graph neural network model learns the spatial correlation features between monitoring nodes under normal conditions to obtain a health benchmark model, includes: The multi-source fusion graph and corresponding node feature matrix under historical normal operating conditions are obtained as input data, and a graph convolutional network model is constructed as a health baseline model. Multi-layer feature extraction based on propagation rules:

[0019] in, This is the node feature matrix of layer 1. To add a self-loop adjacency matrix, For degree matrix, The trainable weight matrix of the first layer, For activation functions; With the goal of minimizing the reconstruction error between the model output and the real input, the mean squared error loss function is used to iteratively update the model parameters, resulting in a healthy benchmark model that can accurately reconstruct the node features under normal conditions.

[0020] In one embodiment of the present invention, the steps of acquiring real-time operational data, constructing a real-time multi-source fusion graph and inputting the real-time multi-source fusion graph into a health benchmark model to obtain a prediction output, calculating the reconstruction error between the prediction output and the real-time operational data, and outputting a determination result on whether the system has experienced an anomaly include: The system collects real-time operating data, constructs a process topology map based on the physical connection relationship of the thermal system, constructs a correlation feature map based on the statistical correlation of historical normal operation data, and weights and fuses the two to obtain a multi-source fusion map. The resulting real-time multi-source fusion map is then input into a trained health benchmark model to obtain a prediction output. Calculate the root mean square error between the predicted output and the actual real-time running data:

[0021] in, The total number of nodes. To ensure real-time operation and accurate data values, The model predicts the output; The calculated root mean square error is compared with a threshold determined by the statistical distribution of errors of normal samples in the source domain. When the root mean square error exceeds the threshold, a judgment result indicating that the system has malfunctioned is output.

[0022] In one embodiment of the present invention, multi-level graph coarsening and fault location include: A real-time fine-grained map representing the current abnormal state is obtained, and a multi-level compression is performed using a spectral clustering algorithm. After L-level coarsening, a series of coarsened map sequences with resolution from high to low are generated. Construct the adjacency matrix of the coarsened graph step by step according to the formula:

[0023] in, and Representing the first The nodes in the level graph are at the 1st level. Two clusters formed in the hierarchical graph and These represent the number of nodes within the cluster; Starting with the coarsest diagram, the analysis proceeds layer by layer upwards. Based on the node anomaly score, the large functional area is first located, and then the scope is gradually narrowed down to the subsystem. Finally, the starting node that caused the anomaly is obtained on the original fine diagram.

[0024] In one embodiment of the present invention, distinguishing fault types includes: Obtain the fault initiation node and all its corresponding neighbor nodes, calculate the correlation change between the fault initiation node and all its neighbor nodes before and after the anomaly, and obtain the reconstruction error of the corresponding node and its correlation coefficient with its strongly correlated neighbors. If the reconstruction error of the fault initiation node is higher than that of other nodes and its own characteristics change abruptly, and its correlation coefficient with all strongly correlated neighbors decreases to close to 0, then the instrument fault is output. If the anomaly of the fault initiation node causes a chain reaction and directional deviation in the characteristics of multiple downstream neighboring nodes, and the correlation coefficient between the neighboring nodes remains stable or increases, then the performance of the output device is truly degraded.

[0025] In one embodiment of the present invention, it further includes: Acquire fault cases and corresponding location results confirmed by operation and maintenance personnel during continuous system operation as new fault samples; The new fault samples are combined with the accumulated normal operation data to form an updated training set, and the health benchmark model is incrementally trained. Incremental training allows the model to adapt to long-term drift caused by system aging and environmental changes, outputting an updated health benchmark model.

[0026] This invention also proposes a multi-parameter correlation anomaly detection and location device for power plant thermal systems, comprising: The graph construction module is used to acquire historical normal operation data, construct a process topology graph based on the physical connection relationship of the thermal system, and construct a correlation feature graph based on the statistical correlation of historical normal operation data. The process topology graph and the correlation feature graph are then fused to obtain a multi-source fusion graph. The model training module is used to acquire historical operating data and corresponding multi-source fusion graphs under normal operating conditions, and input them into the graph neural network model for training, so that the graph neural network model can learn the spatial correlation features between each monitoring node under normal conditions and obtain a health benchmark model. The detection module is used to acquire real-time operating data, construct a real-time multi-source fusion map, input the real-time multi-source fusion map into the health benchmark model, obtain the prediction output, calculate the reconstruction error between the prediction output and the real-time operating data, and output the judgment result of whether the system has an anomaly. The localization module is used to acquire a real-time fine map representing the current abnormal state when the system is determined to be abnormal. The real-time fine map is subjected to multi-level graph coarsening processing to generate a sequence of coarsening maps with different resolutions. The abnormal distribution of each level of coarsening map is analyzed layer by layer to locate the fault starting node from coarse to fine. The fault type is obtained by combining the correlation and change characteristics between the fault starting node and the corresponding neighboring nodes.

[0027] This invention discloses a method and apparatus for detecting and locating multi-parameter correlation anomalies in a power plant thermal system. It constructs a multi-source fusion graph by integrating a physical topology graph and a statistical correlation graph, establishes a health baseline model using a graph neural network, and employs multi-level graph coarsening to achieve coarse-to-fine fault initiation node location. This effectively solves the problems of inaccurate detection and difficult location caused by neglecting system topology correlation and fault propagation characteristics in existing technologies. It achieves an integrated solution across the entire process from data acquisition, graph modeling, anomaly detection, and fault location and type differentiation, significantly improving the accuracy and reliability of multi-parameter correlation anomaly detection in thermal systems and enhancing the method's adaptability and engineering practical value under complex system conditions.

[0028] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a general flowchart of a method for detecting and locating multi-parameter correlation anomalies in a power plant thermal system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the process topology and fault propagation of a power plant feedwater system according to an embodiment of the present invention; Figure 3 This is a training curve diagram of the GCN health benchmark model according to the method of the embodiment of the present invention; Figure 4 This is a comparison chart of the anomaly detection performance of different methods according to embodiments of the present invention; Figure 5 This is a graph showing the fault location performance and computational efficiency of the method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a multi-parameter correlation anomaly detection and location device for a power plant thermal system according to an embodiment of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0032] The following description, with reference to the accompanying drawings, describes a method and apparatus for detecting and locating multi-parameter correlation anomalies in a power plant thermal system according to an embodiment of the present invention.

[0033] Figure 1 This is a flowchart of a method for detecting and locating multi-parameter correlation anomalies in a power plant thermal system according to an embodiment of the present invention.

[0034] A method for detecting and locating multi-parameter correlation anomalies in a power plant thermal system includes the following steps: S1. Obtain historical normal operation data, construct a process topology map based on the physical connection relationship of the thermal system, and construct a correlation feature map based on the statistical correlation of historical normal operation data. Merge the process topology map and the correlation feature map to obtain a multi-source fusion map. S2, acquire historical operating data and corresponding multi-source fusion graphs under normal operating conditions, input them into the graph neural network model for training, so that the graph neural network model learns the spatial correlation features between each monitoring node under normal conditions, and obtains the health benchmark model; S3: Acquire real-time operating data, construct a real-time multi-source fusion graph, input the real-time multi-source fusion graph into the health benchmark model, obtain the prediction output, calculate the reconstruction error between the prediction output and the real-time operating data, and output the judgment result of whether the system has an anomaly. S4. When the system is determined to be abnormal, a real-time fine map representing the current abnormal state is obtained. The real-time fine map is subjected to multi-level graph coarsening processing to generate a sequence of coarsening maps with different resolutions. The abnormal distribution of each level of coarsening map is analyzed layer by layer to locate the fault starting node from coarse to fine. The fault type is obtained by combining the correlation change characteristics between the fault starting node and the corresponding neighboring nodes.

[0035] Specifically, this invention addresses the core shortcomings of current anomaly detection methods for thermal systems, such as neglecting system topology, difficulty in fault location, insufficient consideration of fault propagation characteristics, and low computational efficiency. It proposes a multi-parameter correlation anomaly detection and location method for power plant thermal systems. This invention constructs an integrated architecture encompassing system mechanism topology graph construction, multi-graph fusion and spatial correlation feature learning, GCN health benchmark modeling, and multi-scale coarsening graph anomaly location. Based on graph neural networks, multi-graph fusion theory, and graph coarsening theory, it achieves accurate location of system-level anomaly detection and component-level fault sources, providing a core algorithm engine for intelligent operation and maintenance of thermal systems.

[0036] Furthermore, the present invention mainly involves the following steps: First, to address the issue of traditional methods neglecting the physical connections of the system, a topology graph construction method based on system mechanisms is introduced. Based on the thermal system design drawings and the flow patterns of the working fluid, key equipment is abstracted as nodes in the graph, and the connecting pipes between equipment are abstracted as edges. The resulting process topology graph explicitly encodes the connections of the physical system into computer-processable graph structure data, laying a clear data foundation for subsequent graph neural network analysis.

[0037] Secondly, considering the strong coupling of multiple parameters in the thermal system, multi-graph fusion and GCN health benchmark modeling are implemented. Statistical correlations between data are introduced on top of the basic topology graph to construct a correlation feature map, which captures variables that are physically distant but closely related in their operational patterns. The basic topology graph and the correlation feature map are then fused to construct a more informative multi-source fused graph. Subsequently, a Graph Convolutional Network (GCN) model is constructed, using historical operating data of the unit under normal operating conditions as input. The GCN model automatically learns the spatial correlation characteristics of each node in the graph structure under normal operating conditions. In this way, a benchmark model capable of representing the system's health status is trained.

[0038] Furthermore, to address the challenge of accurately locating the source of a fault after an anomaly occurs, a fault location mechanism based on multi-scale coarsening maps is introduced. When the output of the GCN model of real-time operating data deviates significantly from the health baseline, indicating a system anomaly, the location process is initiated. First, a spectral clustering map coarsening algorithm is applied to the real-time fine map representing the current system state, performing multi-level compression to generate a series of coarsening maps with decreasing resolution. Then, each level of coarsening map is analyzed layer by layer. The coarsest map can quickly determine which major functional area the anomaly occurs in; as the analysis deepens, the scope is gradually narrowed until the starting node causing the anomaly is located on the original fine map. Furthermore, by analyzing the anomaly pattern of the located node and its correlation with neighboring nodes, the fault type can be further distinguished. If a node undergoes a sudden change while neighboring nodes remain normal, it can be identified as an instrument failure; if a node anomaly triggers a chain reaction in downstream nodes along the working fluid flow direction, it can be identified as a genuine degradation of equipment performance.

[0039] This invention constructs an end-to-end intelligent diagnostic framework integrating detection and localization through the synergy of the aforementioned steps. Verification has shown that, compared to traditional methods using only single-point thresholds, PCA / MLP models that do not consider graph structures, and GCN models that do not incorporate multi-scale localization mechanisms, the method of this invention significantly improves core indicators such as anomaly detection accuracy, fault location precision, computational efficiency, and false alarm rate, providing a reliable technical tool for the lean operation and maintenance of thermal systems.

[0040] like Figure 1 As shown, the specific implementation process of this invention revolves around the construction of a system mechanism topology diagram, multi-graph fusion and GCN health benchmark modeling, and multi-scale coarsening graph fault location. Combined with the stringent requirements for accuracy and practicality in power plant thermal systems, a technical solution that can be implemented in engineering is formed. The specific steps are as follows: Step 1: Based on the design drawings and operating mechanism of the thermodynamic system, construct a process topology diagram representing the physical structure of the system. Among them, the node set Represents key equipment or key measurement points in the system; edge set This represents the flow paths of the working fluid or the energy exchange relationships between devices. The constructed process topology diagram explicitly encodes the connections of the physical system into computer-processable graph structure data, laying a physically meaningful data foundation for subsequent graph neural network analysis.

[0041] Step 2: Building upon the basic topology graph constructed in Step 1, further introduce statistical correlations between data to construct a correlation feature graph. This is used to capture variables that are strongly correlated even when physically far apart.

[0042] Step 2.1: Calculate the similarity between the time series of any two nodes (measuring points) in the historical normal operation data using the Pearson correlation coefficient (PCC), as a new edge relationship between the nodes. Use this similarity as a component of the correlation feature matrix to construct a correlation feature graph. elements The calculation is as follows: (1) in For nodes The Each sample value Its mean, The total number of samples. This formula measures the degree of linear correlation between two groups of data by calculating the ratio of their covariance to standard deviation, with values ​​ranging from... Between. When When the absolute value is greater than a set threshold, it is considered that there is a strong correlation between the two nodes, and the corresponding position is set to 1; otherwise, it is set to 0, thus obtaining a binary correlation adjacency matrix.

[0043] Step 2.2: Convert the adjacency matrix of the basic topology graph Adjacency matrix of correlation feature maps The adjacency matrix is ​​obtained by merging the adjacency matrix. : (2) (3) in The total number of graph structures, Feature weights This is the normalization function. In practice, it is simplified to a weighted summation form. Balance coefficient The result was determined after optimization. Through this step, a multi-source fusion graph that simultaneously contains physical connections and statistical correlations was constructed. .

[0044] Step 2.3: Construct a Graph Convolutional Network (GCN) model as a health baseline model. This involves fusing multi-source graphs from historical normal operating conditions. and its node feature matrix The GCN model is trained using this as input. The propagation rules for the GCN layers are as follows: (4) in This is the node feature matrix of layer 1. To add a self-loop adjacency matrix, For degree matrix, The trainable weight matrix of the first layer, This is the activation function.

[0045] Through training, the GCN model learns the spatial correlation features between nodes under normal operating conditions, resulting in a baseline model that can accurately reconstruct or predict node features under normal conditions. The training objective is to minimize the model output. Compared to real input The reconstruction error between the data and the operating state can be calculated using the mean squared error as the loss function. This model establishes a mapping relationship between system data and operating state.

[0046] Step 3: Anomaly detection based on real-time data graphs. When the system is running in real-time, the collected real-time data is constructed into a real-time multi-source fusion graph using the methods described in Steps 1 and 2. Input it into the baseline model trained in step S2. In the middle, the calculation model output Compared with the true value The reconstruction error between the two. The reconstruction error can be expressed as the root mean square error (RMSE): (5) in This represents the total number of nodes. This error measure indicates the degree to which the current operating state deviates from the normal association pattern. When the error... Exceeding the set threshold When an anomaly occurs, the system is identified and a real-time detailed graph is recorded. Used for subsequent localization. Threshold It can be determined by the error statistical distribution of normal samples in the source domain, taking the 95th percentile.

[0047] Step 4: Accurate fault source localization based on multi-scale coarsening map. Once an anomaly is detected, the fault localization mechanism based on multi-scale coarsening map is activated.

[0048] Step 4.1: Create a detailed real-time graph representing the current abnormal state of the system. Multi-level compression is performed using a spectral clustering algorithm. Spectral clustering groups similar nodes into one class by analyzing the characteristic structure of the Laplacian matrix of the graph. Level coarsening generates a series of coarsened maps with resolutions ranging from high to low. ,in This is the original detailed image. The coarsened graph with the fewest nodes. Adjacency matrix of the coarsened graph. Build it as follows: (6) in and Representing the first The nodes in the level graph are at the 1st level. Two clusters formed in the hierarchical graph and These represent the number of nodes within each cluster. In this formula, the numerator is the sum of all edge weights between the two clusters, and the denominator is the product of the number of nodes in the two clusters. Dividing the two yields the average connection strength between a single node. Diagonal elements This represents the average connection strength between unit nodes within a cluster.

[0049] Step 4.2: Analyze the coarsening plots layer by layer. First, analyze the coarsest plot. Based on the anomaly score of its nodes, it is determined which major functional area the anomaly occurred in. Then, it is zoomed in to the next level of more refined coarseness plot. Further analysis within this area is conducted to pinpoint specific subsystems or devices. Finally, the original detailed map is used to refine the analysis. The algorithm focuses on the local subgraph related to the device, analyzes the abnormal patterns of its internal nodes, and finally locates the starting node that caused the abnormality.

[0050] Step 4.3: Based on the anomaly patterns of the located starting node and its neighboring nodes, and combined with the fault propagation characteristics, further determine the fault type. Calculate the correlation changes between this node and all its neighboring nodes before and after the anomaly. If its reconstruction error is significantly higher than other nodes, indicating a sudden change in the node's own characteristics, and its correlation coefficient with all strongly correlated neighbors is high... A sharp drop approaching 0 indicates a preliminary diagnosis of an instrument failure at that node. If the anomaly of this node not only alters its own characteristics but also causes a chain reaction of directional deviations in the characteristics of multiple downstream neighboring nodes, and the correlation coefficients between them remain stable or increase, then it can be determined that the equipment performance corresponding to this node has truly deteriorated.

[0051] Step 5: As the system continues to run, the accumulated fault cases and normal operation data can be used for iterative optimization of the model. Newly generated fault samples and their location results, confirmed by operations and maintenance personnel, are periodically added to the training set to improve the GCN health baseline model from Step 2. Incremental training enables the model to adapt to long-term drift caused by system aging and environmental changes, and to maintain high-precision detection and localization capabilities.

[0052] This invention presents a complete implementation plan integrating detection and localization, encompassing system mechanism topology construction, multi-graph fusion and GCN health benchmark modeling, and multi-scale coarsening graph fault location. By introducing system topology, it addresses the problem of traditional methods severing system correlations. Through multi-graph fusion and GCN deep mining of spatiotemporal correlation features, it achieves sensitive detection of early anomalies. Multi-scale graph coarsening analysis enables precise and efficient localization from the system level to the measurement point level, and it can effectively distinguish between instrument failures and equipment performance degradation by combining fault propagation characteristics. The final structured early warning report can be seamlessly integrated into existing asset health management platforms, providing quantitative, interpretable, and traceable technical support for predictive maintenance of thermal systems and the construction of smart power plants. This completes the technical chain from multi-source data input of the thermal system to anomaly detection and precise fault localization.

[0053] Furthermore, to fully verify the effectiveness and engineering applicability of the technical solution described in this invention, a multi-source data fusion test environment for the grid-side thermal system was constructed and systematically verified, using the feedwater system of a 1000MW ultra-supercritical coal-fired power generating unit as the object. Test data was collected from two years of historical operating data and six months of fault injection experimental data of the unit, covering 30 monitoring variables including parameters such as temperature, pressure, and flow rate from the deaerator, feedwater pump, three high-pressure heaters, regulating valves, and key measuring points, with a sampling frequency of once per minute. The data includes normal operating conditions and three typical faults: high-pressure heater tube bundle leakage, feedwater pump cavitation, and flowmeter drift. Each type of fault includes different degrees of degradation processes and sudden fault samples. A total of 50,000 sets of normal data and 3,000 sets of fault data (1,000 sets for each type of fault) were collected. The first 18 months of normal data were used as the training set to construct a health baseline model, and the last 6 months of data were used as the test set, which included 500 sets of normal data and all fault data (mixed in chronological order). All fault samples were labeled by domain experts with fault type, occurrence time, and faulty equipment / measuring point location for positioning accuracy verification. The experiment verified the technical effectiveness based on core indicators such as anomaly detection accuracy, false alarm rate, fault location accuracy, positioning time, and computational efficiency, which are detailed below with reference to the attached figures.

[0054] like Figure 1 The diagram shown is a flowchart of the multi-parameter correlation anomaly detection and localization method for thermal systems based on graph neural networks, according to the present invention. The diagram fully illustrates the closed-loop process from data input to anomaly detection, fault localization, and model updating, and presents the technical architecture and data flow of the present invention.

[0055] like Figure 2 The diagram shows the process topology and fault propagation of the power plant's feedwater system. Figure 2(a) is a process topology diagram constructed based on the system mechanism. In this diagram, core equipment such as the deaerator, feedwater pump, high-pressure heater, regulating valve, and economizer are abstracted as nodes. The feedwater main pipeline, extraction steam / drainage pipeline, and measuring point signal connections connecting these equipment are abstracted as edges, distinguished by different colors: solid blue lines represent the main feedwater path, solid orange lines represent the extraction steam / drainage branch, and dashed gray lines represent measuring point signal connections. Gray circular nodes in the diagram represent key measuring points. This topology diagram explicitly encodes the working fluid flow relationships of the physical system into graph-structured data, laying a clear physical data foundation for subsequent GCN modeling. Figure 2 (b) is a fault propagation diagram of the water supply system, which presents the fault propagation path in a three-layer structure of typical fault sources, intermediate abnormal states, and system-level consequences. The gray arrows indicate the propagation path of the fault along the direction of the working fluid flow, reflecting the physical law that the actual deterioration of equipment performance will trigger a chain reaction in downstream nodes, and implicitly indicating that if a node is abnormal but neighboring nodes are not affected, it can be identified as an instrument failure.

[0056] like Figure 3 The figure shows the training curves of the GCN health baseline model. The horizontal axis represents the number of training iterations (Epochs), and the vertical axis represents the loss function value. The solid blue line represents the training set loss decrease curve, and the dashed red line represents the validation set loss curve. It can be seen that the model converges after approximately 200 iterations, with the training set loss decreasing to 0.012 and the validation set loss stabilizing at 0.015. This indicates that the model successfully learned the spatial correlation features between nodes under normal operating conditions and did not exhibit significant overfitting.

[0057] like Figure 4 The figure shows a comparison of the anomaly detection performance of the method of this invention with four baseline models. Experiments were conducted on a test set containing 500 normal samples and 3,000 faulty samples. The accuracies of each method were as follows: single-point thresholding method 78.3%, PCA 82.5%, MLP 85.1%, ordinary GCN 89.7%, and the method of this invention 96.2%. The method of this invention improves performance by 6.5 percentage points compared to ordinary GCN and by 17.9 percentage points compared to the traditional single-point thresholding method. Furthermore, the false positive rate of the method of this invention is only 2.1%, and the false negative rate is 3.8%, significantly better than other methods. This comparison fully demonstrates the ability of the multi-graph fusion mechanism to deeply mine system correlation features and the sensitivity of the health baseline model to early, subtle anomalies.

[0058] like Figure 5 As shown, this paper demonstrates the performance of the method of the present invention in terms of fault location and computational efficiency. Figure 5(a) The location accuracy for three types of faults: high-pressure heater tube bundle leakage 94.3%, feedwater pump cavitation 91.8%, and flowmeter drift 97.2%, with a comprehensive location accuracy of 94.5%. In contrast, ordinary GCN can only output anomaly scores but cannot locate the fault, requiring manual investigation with an average time of about 45 minutes; the method of this invention, through a multi-scale coarsening location mechanism, takes only 3.2 seconds on average from detecting an anomaly to outputting the fault node, thus improving the location efficiency. Figure 5 (b) Comparison of average inference time per sample for each method. This figure verifies that the present invention, while ensuring high positioning accuracy, also possesses the real-time performance required for engineering deployment.

[0059] In summary, the method of this invention, through multi-graph fusion GCN health benchmark modeling and multi-scale coarsening localization technology, achieves an accuracy of 96.2% and a false alarm rate of only 2.1% in anomaly detection of water supply systems; the fault location accuracy reaches 94.5%, and the localization time is shortened to the second level, greatly improving diagnostic efficiency; the single-sample inference time is 2.43ms, meeting the real-time requirements of engineering. Experimental data fully verify the significant advantages of this invention in multi-parameter correlation anomaly detection and accurate localization in thermal systems, providing accurate, efficient, and interpretable technical support for intelligent operation and maintenance of power grid generation-side equipment.

[0060] To implement the method of the above embodiments, such as Figure 6 As shown, this invention proposes a multi-parameter correlation anomaly detection and location device 10 for power plant thermal systems, comprising: The graph construction module 100 is used to acquire historical normal operation data, construct a process topology graph based on the physical connection relationship of the thermal system, construct a correlation feature graph based on the statistical correlation of historical normal operation data, and fuse the process topology graph and the correlation feature graph to obtain a multi-source fusion graph.

[0061] The model training module 200 is used to acquire historical operating data and corresponding multi-source fusion graphs under normal operating conditions, and input them into the graph neural network model for training, so that the graph neural network model learns the spatial correlation features between monitoring nodes under normal conditions and obtains a health benchmark model.

[0062] The detection module 300 is used to acquire real-time operating data, construct a real-time multi-source fusion map, input the real-time multi-source fusion map into the health benchmark model, obtain the prediction output, calculate the reconstruction error between the prediction output and the real-time operating data, and output the judgment result of whether the system has an anomaly.

[0063] The positioning module 400 is used to acquire a real-time fine map representing the current abnormal state when the system is determined to be abnormal, perform multi-level graph coarsening processing on the real-time fine map to generate a sequence of coarsening maps with different resolutions, analyze the abnormal distribution of each level of coarsening map to locate the fault starting node from coarse to fine, and obtain the fault type by combining the correlation change characteristics between the fault starting node and the corresponding neighboring nodes.

[0064] The apparatus of this invention effectively solves the problems of neglecting system topology, difficulty in fault location, and ambiguity in type identification in the prior art. It realizes an integrated solution for the entire process from data acquisition, health modeling, anomaly detection to accurate fault source location, significantly improving the accuracy of anomaly detection and fault location in thermal systems, and enhancing the adaptability and engineering practical value of the method under complex coupled conditions.

[0065] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0066] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for detecting and locating multi-parameter correlation anomalies in a power plant thermal system, characterized in that, include: Historical normal operation data is obtained, a process topology map is constructed based on the physical connection relationship of the thermal system, and a correlation feature map is constructed based on the statistical correlation of historical normal operation data. The process topology map and the correlation feature map are fused to obtain a multi-source fusion map. Acquire historical operating data and corresponding multi-source fusion graphs under normal operating conditions, input them into a graph neural network model for training, so that the graph neural network model can learn the spatial correlation features between monitoring nodes under normal conditions and obtain a health benchmark model. Acquire real-time operational data, construct a real-time multi-source fusion graph, input the real-time multi-source fusion graph into the health benchmark model, obtain the prediction output, calculate the reconstruction error between the prediction output and the real-time operational data, and output the judgment result of whether the system has an anomaly. When an anomaly is detected in the system, a real-time fine-grained map representing the current abnormal state is obtained. The real-time fine-grained map is then subjected to multi-level graph coarsening to generate a sequence of coarsened maps with different resolutions. The anomaly distribution of each level of coarsened map is analyzed layer by layer to locate the fault initiation node from coarse to fine. The fault type is obtained by combining the correlation and change characteristics between the fault initiation node and its corresponding neighboring nodes.

2. The method according to claim 1, characterized in that, The process involves acquiring historical normal operation data, constructing a process topology map based on the physical connections of the thermal system, and constructing a correlation feature map based on the statistical correlations of the historical normal operation data. The process topology map and the correlation feature map are then fused to obtain a multi-source fusion map, including: Obtain the design drawings and working fluid flow patterns of the thermal system, abstract key equipment as node sets, abstract the connecting pipes between equipment as edge sets, construct the process topology graph, and output the corresponding adjacency matrix. ; Obtain the time series of any two nodes i and j from the historical normal operation data, and calculate the similarity using the Pearson correlation coefficient. The absolute value of the similarity is compared with a set threshold. If it is greater than the threshold, the corresponding position is set to 1; otherwise, it is set to 0. A binary correlation adjacency matrix is ​​constructed as the correlation feature map, and the corresponding adjacency matrix is ​​obtained. ; Set the balance coefficient α, and Multiply by α Multiply by (1) After summing α), we obtain the merged adjacency matrix: based on Construct and output a multi-source fusion graph.

3. The method according to claim 2, characterized in that, Obtain the time series of any two nodes i and j from the historical normal operation data, and calculate the similarity using the Pearson correlation coefficient. The absolute value of the similarity is compared with a set threshold, including: Get the kth sample value of node i and the corresponding mean The k-th sample value of node j and the corresponding mean Based on the total number of samples N, the relevance feature matrix elements are calculated using the input formula. : The calculated The absolute value is compared with a set threshold to obtain a binarized correlation adjacency matrix.

4. The method according to claim 2, characterized in that, Set the balance coefficient, calculate the fused adjacency matrix, and based on... Output a multi-source fusion graph, including: optimizing the balance coefficient α to 0.5, and inputting... and ,calculate Output the merged adjacency matrix .

5. The method according to claim 1, characterized in that, The process of acquiring historical operating data and corresponding multi-source fusion graphs under normal operating conditions, inputting them into a graph neural network model for training, and enabling the graph neural network model to learn the spatial correlation features between monitoring nodes under normal conditions to obtain a health benchmark model includes: The multi-source fusion graph and corresponding node feature matrix under historical normal operating conditions are obtained as input data, and a graph convolutional network model is constructed as a health baseline model. Multi-layer feature extraction based on propagation rules: in, This is the node feature matrix of layer 1. To add a self-loop adjacency matrix, For degree matrix, The trainable weight matrix of the first layer, For activation functions; With the goal of minimizing the reconstruction error between the model output and the real input, the mean squared error loss function is used to iteratively update the model parameters, resulting in a healthy benchmark model that can accurately reconstruct the node features under normal conditions.

6. The method according to claim 1, characterized in that, The process of acquiring real-time operational data, constructing a real-time multi-source fusion graph, inputting the real-time multi-source fusion graph into a health benchmark model to obtain a prediction output, calculating the reconstruction error between the prediction output and the real-time operational data, and outputting a determination result on whether the system has experienced an anomaly includes: The system collects real-time operating data, constructs a process topology map based on the physical connection relationship of the thermal system, constructs a correlation feature map based on the statistical correlation of historical normal operation data, and weights and fuses the two to obtain a multi-source fusion map. The resulting real-time multi-source fusion map is then input into a trained health benchmark model to obtain a prediction output. Calculate the root mean square error between the predicted output and the actual real-time running data: in, The total number of nodes. To ensure real-time operation and accurate data values, The model predicts the output; The calculated root mean square error is compared with a threshold determined by the statistical distribution of errors of normal samples in the source domain. When the root mean square error exceeds the threshold, a judgment result indicating that the system has malfunctioned is output.

7. The method according to claim 1, characterized in that, Multi-level graph coarsening and fault location, including: A real-time fine-grained map representing the current abnormal state is obtained, and a multi-level compression is performed using a spectral clustering algorithm. After L-level coarsening, a series of coarsened map sequences with resolution from high to low are generated. Construct the adjacency matrix of the coarsened graph step by step according to the formula: in, and Representing the first The nodes in the level graph are at the 1st level. Two clusters formed in the hierarchical graph and These represent the number of nodes within the cluster; Starting with the coarsest diagram, the analysis proceeds layer by layer upwards. Based on the node anomaly score, the large functional area is first located, and then the scope is gradually narrowed down to the subsystem. Finally, the starting node that caused the anomaly is obtained on the original fine diagram.

8. The method according to claim 1, characterized in that, Distinguish between fault types, including: Obtain the fault initiation node and all its corresponding neighbor nodes, calculate the correlation change between the fault initiation node and all its neighbor nodes before and after the anomaly, and obtain the reconstruction error of the corresponding node and its correlation coefficient with its strongly correlated neighbors. If the reconstruction error of the fault initiation node is higher than that of other nodes and its own characteristics change abruptly, and its correlation coefficient with all strongly correlated neighbors decreases to close to 0, then the instrument fault is output. If the anomaly of the fault initiation node causes a chain reaction and directional deviation in the characteristics of multiple downstream neighboring nodes, and the correlation coefficient between the neighboring nodes remains stable or increases, then the performance of the output device is truly degraded.

9. The method according to claim 1, characterized in that, Also includes: Acquire fault cases and corresponding location results confirmed by operation and maintenance personnel during continuous system operation as new fault samples; The new fault samples are combined with the accumulated normal operation data to form an updated training set, and the health benchmark model is incrementally trained. Incremental training allows the model to adapt to long-term drift caused by system aging and environmental changes, outputting an updated health benchmark model.

10. A device for detecting and locating multi-parameter correlation anomalies in a power plant thermal system, characterized in that, include: The graph construction module is used to acquire historical normal operation data, construct a process topology graph based on the physical connection relationship of the thermal system, and construct a correlation feature graph based on the statistical correlation of historical normal operation data. The process topology graph and the correlation feature graph are then fused to obtain a multi-source fusion graph. The model training module is used to acquire historical operating data and corresponding multi-source fusion graphs under normal operating conditions, and input them into the graph neural network model for training, so that the graph neural network model can learn the spatial correlation features between each monitoring node under normal conditions and obtain a health benchmark model. The detection module is used to acquire real-time operating data, construct a real-time multi-source fusion map, input the real-time multi-source fusion map into the health benchmark model, obtain the prediction output, calculate the reconstruction error between the prediction output and the real-time operating data, and output the judgment result of whether the system has an anomaly. The localization module is used to acquire a real-time fine map representing the current abnormal state when the system is determined to be abnormal. The real-time fine map is subjected to multi-level graph coarsening processing to generate a sequence of coarsening maps with different resolutions. The abnormal distribution of each level of coarsening map is analyzed layer by layer to locate the fault starting node from coarse to fine. The fault type is obtained by combining the correlation and change characteristics between the fault starting node and the corresponding neighboring nodes.

Images