Method and device for early warning of thermal fault of stator winding of steam turbine generator

Abstract

The application discloses a steam turbine generator stator winding thermal fault early warning method and device, and relates to the field of data processing.The method comprises the following steps: S1, taking the continuously sampled steam turbine generator stator winding temperature history health data as a reference data set; S2, taking the variables of the reference data set as nodes, mining the causal dependence relationship between the variables to construct a directed edge, and obtaining a directed reference network; S3, constructing a dynamic deviation network according to the causal deviation of the current window of the to-be-diagnosed temperature variable data and the directed reference network; S4, constructing a local module of each node and calculating a local score; S5, selecting a plurality of nodes with the highest local score in the current time window and calculating the average score thereof; and S6, taking the average score difference between the current time window and the previous time window as a comprehensive state score DNDS, and using the DNDS to determine the thermal fault.The application does not need to rely on a large number of historical fault labeled samples, does not need to establish a complex thermodynamic mechanism model, has low calculation cost, and has strong working condition adaptability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and device for early warning of thermal faults in the stator windings of steam turbine generators. Background Technology

[0002] As a core component of a generator, the stator winding endures the combined effects of high voltage, high current, and mechanical vibration during operation, making it highly susceptible to localized overheating faults. If not detected and addressed promptly, these faults can rapidly escalate into serious accidents such as insulation breakdown and grounding short circuits. Therefore, constructing an efficient and reliable early warning technology system for stator winding thermal faults is of significant engineering application value and practical importance for improving equipment operational safety and reducing maintenance costs. Currently, Distributed Control Systems (DCS) are widely used in power plants, providing a reliable data foundation for identifying abnormal states and diagnosing faults in steam turbine generator sets. Existing DCS-based data-driven methods for diagnosing thermal faults in steam turbine generator stator windings mainly fall into three categories: expert system diagnostic methods, traditional machine learning diagnostic methods, and deep learning diagnostic methods. Some scholars have attempted to combine expert systems with neural networks, using neural networks to construct a fault diagnosis knowledge base and verifying the feasibility of a hybrid expert system based on fuzzy neural networks in generator fault identification tasks. Other scholars have established real-time status evaluation indicators for generator water-cooled systems using similar clouds and verified the effectiveness of this method through practical cases. Still others have proposed an early warning method based on SAE-LSTM and a sliding window approach for overheating defects in the stator windings of water-cooled turbine generators. Results show that, compared with traditional threshold-based early warning methods, the proposed method can provide early warning of defects up to 85 hours in advance. Existing DCS-based data-driven methods for thermal faults in turbine generator stator windings have made some progress in early warning and fault diagnosis, but still suffer from limitations such as high dependence on fault samples, lagging early warning capabilities, and high computational resources and costs.

[0003] In recent years, the theory of "critical phase transition" in nonlinear system dynamics has gradually attracted attention. When the state of a complex multivariable system evolves to a critical point, a set of strongly correlated states and highly fluctuating dominant variables emerge, generating three possible warning signals: a significant increase in the standard deviation of each dominant variable, enhanced correlation among the dominant variables, and a decrease in the correlation between the dominant variables and external non-dominant variables. To achieve early warning of critical slowdown phenomena in complex systems, Chen Luonan et al. from the Chinese Academy of Sciences proposed the Dynamic Network Marker (DNM) theory, applying it to early warning of severe biological diseases. Based on this theory, in the field of thermal fault diagnosis of turbine generator stator windings, Jin Liang et al. pioneered the application of DNM to early warning of stator winding overheating defects, achieving good results. Summary of the Invention

[0004] To address the above problems, this invention proposes a method and device for early warning of thermal faults in the stator windings of steam turbine generators. Based on the monitored temperature variables of the target steam turbine generator stator windings, these variables are mapped to network nodes based on causal relationships. Local scores are calculated using three characteristic quantitative indicators based on dynamic network markers. The nodes with the highest local scores in the dynamic deviation network are extracted as dynamic network marker subnetworks, thereby constructing a Density-Related State Score (DNDS) for the stator windings to identify when the unit transitions from steady-state operation to a critical state. This method does not rely on a large number of historical fault-labeled samples for supervised training using deep learning methods; it does not require establishing complex thermodynamic mechanism models, has low computational cost, and strong adaptability to operating conditions, thus facilitating early warning of overheating defects in the steam turbine generator stator windings.

[0005] On the one hand, the method for early warning of thermal faults in the stator windings of steam turbine generators includes the following specific steps:

[0006] S1, acquire the continuous temperature variable data of the turbine generator stator winding; use the continuously sampled historical temperature health data as a reference dataset;

[0007] S2 uses the variables of the reference dataset as nodes to mine the causal dependencies between variables to construct directed edges, thus obtaining a directed reference network;

[0008] S3, use a sliding window to update the temperature variable data to be diagnosed, and construct a dynamic bias network based on the causal deviation between the current window of the temperature variable data to be diagnosed and the directed reference network;

[0009] S4, take each node of the dynamic bias network and all its first-order downstream nodes as local modules of that node, and calculate the local scores of the comprehensive local module internal bias, internal synergy and external decorrelation of each node.

[0010] S5: Select the nodes with the highest local scores within the current time window and calculate their average score;

[0011] S6, the difference between the average score of the current time window and the previous time window is used as the comprehensive status score DNDS of the turbine generator stator winding; the turbine generator stator winding with abnormal dynamic changes in DNDS value is identified as a thermal fault and an early warning is issued.

[0012] Preferably, the process involves using variables from the reference dataset as nodes to mine causal dependencies between variables to construct directed edges, thus obtaining a directed reference network; specifically as follows:

[0013] Based on the reference dataset, the time-delay causal relationship and instantaneous coupling relationship between variables are learned through statistical tests, resulting in directed time-delay causal edges and undirected undirected causal networks;

[0014] Using time-delayed causal edges and undirected causal networks as constraints, a hill-climbing search is performed on the undirected edges to determine the optimal direction of the instantaneous edges, resulting in a complete causal network with timestamps.

[0015] The complete causal network with timestamps is mapped to a directed reference network containing only physical variables. The Pearson correlation coefficient of each edge in the directed reference network is calculated using the reference dataset and used as the association strength weight of that edge.

[0016] Preferably, the hill-climbing search adopts the minimum description length scoring criterion.

[0017] Preferably, based on the reference dataset, the time-delay causal relationship and instantaneous coupling relationship between variables are learned through statistical tests, resulting in directed time-delay causal edges and undirected undirected causal networks, as detailed below:

[0018] Perform a rank transformation on the reference dataset to obtain the rank-transformed data;

[0019] A candidate set of lagged parent nodes for the target variable is constructed based on the data after rank transformation; the candidate set of lagged parent nodes includes the values ​​of all other variables over a past time window;

[0020] Using the partial rank correlation coefficient, the net correlation between each candidate lagged variable and the target variable in the candidate set of lagged parent nodes is calculated sequentially, and the conditional independence test is performed.

[0021] For two variables that do not meet the conditional independence test, construct a time-delayed causal edge from the candidate lagged variable to the target variable.

[0022] Preferably, the undirected causal network is obtained through the following steps:

[0023] A forward heuristic search is performed, specifically: for each target variable in the rank-transformed data, the Max-Min heuristic strategy is used to evaluate and filter all other variables at the same time to obtain the instantaneous candidate parent-child set of the target variable;

[0024] Perform backward conditional independence tests and screening on the instantaneous candidate parent-child sets to obtain the instantaneous parent-child sets; specifically: if a conditional set exists... If P > θ, then remove E from the instantaneous candidate parent-child set; otherwise, keep E; where, Represents an instantaneous candidate parent-child set; Represents the set of lagging parent nodes; Represents the union; This indicates the removal of candidate nodes from the instantaneous candidate parent-child set. ;

[0025] Using all variables as nodes, undirected edges are used to connect the target variable and the variables in its instantaneous parent and child sets to obtain an undirected causal network.

[0026] Preferably, the construction of a dynamic bias network based on the causal deviation between the current window of the temperature variable data to be diagnosed and the directed reference network is as follows:

[0027] Calculate the Pearson correlation coefficients for any two variables within the current time window of the directed reference network and the temperature variable data to be diagnosed;

[0028] The statistical significance of the two Pearson correlation coefficients was tested using a two-sample Fisher Z test to obtain a statistical significance score. The edge between two variables with a statistical significance score greater than a preset significance threshold was determined to have undergone a significant perturbation at the current time; otherwise, the edge was not considered to have undergone a significant perturbation.

[0029] The dynamic bias network is constructed by using all variables as nodes, edges that have undergone significant perturbations as edges, and the absolute value of the difference in the Pearson correlation coefficient between two variables as the weight of the edge.

[0030] Preferably, the Pearson correlation coefficient is expressed as:

[0031] ;

[0032] in, Representing variables and variables The Pearson correlation coefficient at time t; Indicates the length of the sliding window; Indicates timestamp as Time variable Observed values; Indicates the first in the sliding window One data point; Indicates timestamp as Time variable Observed values; Representing variables The arithmetic mean of all observations within the current sliding window; Representing variables The arithmetic mean of all observations within the current sliding window.

[0033] Preferably, step S6 further includes: mapping the scores of nodes corresponding to all variables within all time windows to a three-dimensional coordinate graph to draw a three-dimensional visualization of the entire process.

[0034] Preferably, the local score is expressed as:

[0035] ;

[0036] in, express Nodes of the time window Local scores; express Nodes of the time window The internal deviation of a local module reflects the internal deviation of the local module; express Nodes of the time window The strength of abnormal correlations outside the local module reflects the decorrelation outside the local module; express Nodes of the time window The abnormal driving strength within a local module reflects the internal coordination of the local module; This indicates that positive numbers with a denominator of 0 should be prevented.

[0037] On the other hand, the steam turbine generator stator winding thermal fault early warning device includes the following:

[0038] The data acquisition module is used to acquire continuous temperature variable data of the stator winding of the steam turbine generator; and uses continuously sampled historical temperature health data as a reference dataset.

[0039] The directed reference network building module is used to construct directed edges by mining the causal dependencies between variables using the variables of the reference dataset as nodes, thus obtaining a directed reference network.

[0040] The dynamic bias network construction module is used to update the temperature variable data to be diagnosed using a sliding window, and to construct a dynamic bias network based on the causal bias between the current window of the temperature variable data to be diagnosed and the directed reference network.

[0041] The local score acquisition module is used to treat each node of the dynamic bias network and all its first-order downstream nodes as local modules of that node, and to calculate the local score of each node's comprehensive local module internal bias, internal synergy and external decorrelation.

[0042] The time window average score acquisition module is used to select several nodes with the highest local scores within the current time window and calculate their average score.

[0043] The thermal fault determination module is used to take the average score difference between the current time window and the previous time window as the comprehensive status score DNDS of the turbine generator stator winding; and to identify the turbine generator stator winding with abnormal dynamic changes in DNDS value as a thermal fault and issue an early warning.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention maps the monitored variables covered by the DCS system of the target steam turbine generator into causal network nodes. By calculating three characteristic quantitative indicators and local scores of dynamic network markers, it extracts the dynamic network marker subnetwork from the dynamic deviation network, and then constructs the Unit Stator Winding Comprehensive State Score (DNDS). A dynamic threshold is set to identify when the unit transitions from steady-state operation to a critical state. This method does not require supervised training using deep learning methods based on a large number of historical fault-labeled samples; it does not require establishing complex thermodynamic mechanism models, has low computational cost, and strong adaptability to operating conditions, which helps to provide early warning of overheating defects in the steam turbine generator stator winding. Attached Figure Description

[0046] The present invention will now be described in further detail with reference to the accompanying drawings;

[0047] Figure 1 This is a flowchart illustrating the steps of the steam turbine generator stator winding thermal fault early warning method according to an embodiment of the present invention.

[0048] Figure 2 This is a flowchart illustrating the early warning method for thermal faults in the stator winding of a steam turbine generator according to an embodiment of the present invention.

[0049] Figure 3 This is a directed reference network diagram constructed based on historical health data for the early warning method of thermal faults in the stator winding of a steam turbine generator according to an embodiment of the present invention.

[0050] Figure 4 This is the complete process of each variable in 671 time windows of the early warning method for thermal faults in the stator winding of a steam turbine generator according to an embodiment of the present invention. Value 3D visualization;

[0051] Figure 5This is a graph showing the trend of DNDS values ​​changing at each stage of the evolution of a thermal fault in the stator winding of a steam turbine generator, as described in an embodiment of the present invention.

[0052] Figure 6 This is a dynamic deviation network diagram of the first abnormality in the steam turbine generator stator winding thermal fault early warning method according to an embodiment of the present invention.

[0053] Figure 7 This is a structural block diagram of a steam turbine generator stator winding thermal fault early warning device according to an embodiment of the present invention. Detailed Implementation

[0054] The present invention will be further described below through specific embodiments.

[0055] like Figure 1 and Figure 2 As shown, the method for early warning of thermal faults in the stator winding of a steam turbine generator includes the following specific steps:

[0056] S1, acquire the continuous temperature variable data of the turbine generator stator winding; use the continuously sampled historical temperature health data as a reference dataset.

[0057] Collect continuous monitoring variable data covered by the DCS system of the steam turbine generator set, and select... Historical health data from consecutive sampling points are used as a reference dataset. Data from consecutive sampling times were used as samples to be diagnosed and preprocessed.

[0058] Temperature changes in the stator windings of a steam turbine generator are a complex thermo-mechanical coupling process. Based on physical mechanism analysis, redundant nodes that are unrelated to the evolution of thermal faults are eliminated.

[0059] Assuming the filtered reference dataset is The sample to be diagnosed is .

[0060] Standardize the data:

[0061] ;

[0062] in, For the collected variable values, For variable values Standardized data; This is the average of the collected variable values; This represents the standard deviation of the collected variable values.

[0063] S2 uses the variables of the reference dataset as nodes to mine the causal dependencies between variables to construct directed edges, thus obtaining a directed reference network.

[0064] Set the length and step size parameters of the sliding window, and construct a time window data matrix using the preprocessed samples to be diagnosed.

[0065] Let the length of the sliding window be 'a' and the step size be 'a'. To obtain a time window data matrix, we acquire m variables from the sample to be diagnosed, continuously collect the variable values ​​of m variables at a time points, and obtain the values ​​of m variables at a time points a. , represented as:

[0066] ;

[0067] in, This represents the j-th variable in the current time window. One variable value;

[0068] Repeat the above steps Next time, obtain A data matrix with a time window.

[0069] A directed reference network for a steam turbine generator set based on historical health data is established using a time-series hybrid constraint hill-climbing algorithm.

[0070] S201, Perform rank transformation on the reference dataset:

[0071] Select a preprocessed reference dataset (n sets in total), and obtain the original time series vector P for each monitoring point, denoted as... ,in This represents the original physical observation value of the i-th sampling point, where n is the total number of sampling points; all elements in the original time series vector P are sorted monotonically in ascending order according to their numerical values ​​to generate an ordered sequence. , recorded as = ,satisfy ≤ ≤ Based on the original observations In an ordered sequence The position index in the matrix determines its corresponding rank. If the original observed values In an ordered sequence If a position is unique, then its rank value is... Equal to its sorted position index If the original observed values In an ordered sequence If there are c identical values ​​in a given set, and these values ​​occupy consecutive positions from the b-th position to the (b+c-1)-th position, then its rank is... The arithmetic mean of the above position indices is calculated; its mathematical expression is:

[0072] ;

[0073] Using the calculated rank value Replace the original observations Generate a rank sequence for subsequent parameter calculation and causal analysis. = .

[0074] S202, based on the data after rank transformation, uses the partial rank correlation coefficient to identify the time-delay causal relationship between each monitoring node:

[0075] Build target node The set of lagging parent nodes:

[0076] ;

[0077] in, For target sensor The node at time t; For the i-th sensor in the past The numerical value at any given moment; For the time lag, ; The maximum lag time is determined based on the sampling interval.

[0078] Iterative calculation of partial rank correlation coefficient:

[0079] ;

[0080] Where Z is the current set of condition variables, which is used during the iteration process. The subset of nodes other than the node being tested; Under the influence of the condition variable set Z, the variable and The partial rank correlation coefficient; The Pearson correlation coefficient between rank data A and B; Variables , The rank sequence of the condition set Z.

[0081] S203. To verify the significance of the correlation, a conditional independence test is performed:

[0082] Establish the statistical test T:

[0083] ;

[0084] in, For sensors and The partial rank correlation coefficient; is the sample size of the reference dataset; m is the total number of nodes in the network.

[0085] Then calculate the corresponding P-value:

[0086] ;

[0087] in, The cumulative distribution function of the standard normal distribution. Given a significance level θ, p-value > θ, and a sensor... and If the edges are unrelated, delete them; if p-value < θ, the sensor... and The edges are related and should be preserved.

[0088] S204, based on the data in the reference dataset after rank transformation, uses a Max-Min heuristic strategy to identify the coupling relationships between nodes at the same time:

[0089] Constructing the instantaneous neighbor set: for the target node Q= Initialize the instantaneous candidate parent-child set CPC(Q) = .

[0090] Calculate the minimum correlation coefficient MinAssoc between candidate node E and target node Q:

[0091] ;

[0092] Where S is a subset of the current candidate parent-child set CPC(Q); Let E be the partial rank correlation coefficient between E and Q given a subset S.

[0093] Select the node with the largest MinAssoc value Calculate the partial rank correlation coefficient and verify the significance of the correlation. If the p-value < θ, then... Add to CPC(Q); if p-value > θ, it means the variables are considered conditionally independent, and the node will not be added to CPC(Q). Then, lag the parent node set. Merge with CPC(Q) and perform a backward elimination check on each node E in CPC(Q). If a condition set exists... If p-value > θ, it means that the association between node E and target node Q is caused by the propagation of other variables, so E is removed from CPC(Q); if p-value < θ, it means that even if all other factors are considered, node E still has a direct influence on target node Q; repeat the above process until all instantaneous candidate nodes have been traversed, and finally, output the undirected network skeleton.

[0094] S205, based on the Minimum Description Length (MDL) scoring criterion, determines the orientation of instantaneous edges in the undirected network skeleton:

[0095] ;

[0096] in, Let be the predicted value of the i-th parameter at time t; At time t, pointing to All parent nodes (including lagging parent nodes) Instantaneous parent node with other parameters A numerical vector consisting of (components); This is the corresponding least squares regression coefficient vector; .

[0097] ;

[0098] in, Let be the true value of the i-th parameter at time t; This represents the sum of squared residuals of the local network structure.

[0099] ;

[0100] in, The score represents the total minimum description length of the network structure G; m is the total number of parameters. For pointing The number of elements in the set of all parent nodes. Then, an iterative search using a constraint hill-climbing search algorithm is performed to find the number of elements in the set of all parent nodes. The minimum optimal network structure: For instantaneous edges, perform operations such as adding, deleting, or reversing the direction, choosing the one that minimizes... The operation with the largest descent updates the network structure until the score converges, ultimately outputting a causal network graph containing explicit causal direction and time lag information. .

[0101] S206, Define a causal network diagram. = , where the node set Includes variable nodes with lag time. ; directional reference network = , This represents m network parameter nodes. All nodes Unified mapping to The only physical node in For causal network diagrams any → or → If it exists, then it exists in the directed reference network. China → The unique directed physical connection ultimately constructs a directed reference network containing a clear causal direction. .

[0102] S207, using the preprocessed, rank-less transformed reference dataset to calculate the Pearson correlation coefficient between variables and quantify the edge weights of the directed reference network:

[0103] ;

[0104] in, Representing variables and variables Pearson correlation coefficient for the reference sample.

[0105] S3, use a sliding window to update the temperature variable data to be diagnosed, and construct a dynamic bias network based on the causal deviation between the current window of the temperature variable data to be diagnosed and the directed reference network.

[0106] S301, Let the current time be t, the sliding window length be a, and the step size be h. Obtain the time window matrix. For each directed edge in the directed reference network, the target vector is... = The source vector is = Calculate parameters Pearson correlation coefficient of the sliding window at time t :

[0107] ;

[0108] in, Representing variables and variables The Pearson correlation coefficient at time t; Indicates the length of the sliding window; Indicates timestamp as Time variable The observed values, Indicates the first in the sliding window One data point; Indicates timestamp as Time variable Observed values; Representing variables The arithmetic mean of all observations within the current sliding window; Representing variables The arithmetic mean of all observations within the current sliding window;

[0109] S302, for the same pair of variables, calculate the directed reference network and The specific Pearson correlation coefficient of the time window at time t is expressed as:

[0110] ;

[0111] in, Representing variables and variables In the The specific Pearson correlation coefficient within the time window of time.

[0112] S303. To ensure that the analyzed data reflects true dynamic changes rather than the influence of random noise, a two-sample Fisher Z-test is performed on the disturbance:

[0113] ;

[0114] Where z is the z-statistic after the correlation coefficient is converted into a normal distribution by Fisher's Z-transform.

[0115] ;

[0116] in, This is the standard error of the difference, used to correct for the variance effect caused by different sample sizes. The length of the sliding window; This is the sample size for the reference dataset.

[0117] ;

[0118] Where Z is the statistical significance score; The standardized statistic is the real-time Pearson correlation coefficient calculated based on the current sliding window data, after Fisher Z-transformation. The standardized statistic is the historical baseline Wilson correlation coefficient calculated based on the reference dataset, after Fisher's Z-transformation.

[0119] Set a significance threshold If |Z|> If |Z|≤ If the change does not reach a statistically significant level, it is considered random noise fluctuation and the edge is ignored. Then, for... The dynamic deviation network edge weights are assigned values ​​at each time step. For activated edges, their weights are set to... Finally, the following was constructed: Dynamic bias network with causal perturbation weights at different times .

[0120] S4 treats each node of the dynamic bias network and all its first-order downstream nodes as local modules of that node, and calculates the local scores of the comprehensive local module internal bias, internal synergy, and external decorrelation for each node.

[0121] For any node in the network Define its local modules This module consists of nodes. It consists of itself and all its first-order downstream nodes in the dynamic deviation network.

[0122] S401, Calculate internal deviation:

[0123] ;

[0124] in, The internal deviation of a local module of node x at time window time is used to measure the degree of deviation of all parameters within the local module relative to the directed reference network at the current time window. For nodes The number of first-order downstream nodes; For nodes exist The mean within the sliding window at any given time; and Nodes representing the reference dataset The mean and standard deviation.

[0125] S402, Calculate internal synergy:

[0126] ;

[0127] in, for Nodes of the time window The local module's abnormal drive strength is used to measure the node's... The correlation between it and all its downstream nodes deviates from the average magnitude of the reference state. For nodes The set of first-order downstream nodes; For nodes The first-order downstream node.

[0128] S403, Calculate external decorrelation:

[0129] ;

[0130] in, for Nodes of the time window The strength of anomaly associations outside local modules, measured by the nodes The degree to which the correlation between the first-order downstream module driven and the rest of the system deviates from the reference state; For nodes The number of all nodes outside the local module; For nodes The first-order downstream node; This refers to nodes outside the local module.

[0131] S404, calculate the overall score (or local score) of each node in the dynamic deviation network at the current moment:

[0132] ;

[0133] in, for Nodes of the time window DND score; For very small positive numbers, to prevent A value of 0 cannot be calculated. .

[0134] S5: Select the nodes with the highest local scores within the current time window and calculate their average score.

[0135] Construct a three-dimensional coordinate system with the X-axis representing variables, the Y-axis representing time, and the Z-axis representing node scores. Place the calculated scores of each node corresponding to each variable within each time window into the three-dimensional coordinate axis to obtain a three-dimensional visualization of the entire process.

[0136] S6, the difference between the average score of the current time window and the previous time window is used as the comprehensive status score DNDS of the turbine generator stator winding; the turbine generator stator winding with abnormal dynamic changes in DNDS value is identified as a thermal fault and an early warning is issued.

[0137] The Comprehensive State Fraction (DNDS) of the stator winding of a steam turbine generator is expressed as:

[0138] ;

[0139] in, Let be the local score of the j-th node at time t; Let be the local score of the j-th node at time t-1.

[0140] Verification experiments were conducted on embodiments of the present invention, using a large steam turbine generator from a power plant as the object. Data acquisition was performed by the unit's DCS system, with a sampling period of 1 minute. This experiment selected 1440 consecutive sampling times (i.e.,... A directed reference network is constructed based on health data (corresponding to 24 hours), such as... Figure 3 As shown in the figure, this diagram illustrates the directed reference network topology constructed from time-series monitoring data of the turbine generator stator winding using the TS-MCHC method. Since the original result generated by TS-MCHC is a causal graph containing both time lag and instantaneous information, a directed edge is established in the graph to visually demonstrate the spatial coupling relationships between the physical measurement points, provided a causal connection exists between any two physical nodes at any time step. In the diagram, nodes represent temperature measurement points of the stator winding, and directed edges represent causal dependencies. This network can serve as a baseline directed reference network.

[0141] In the event of a thermal fault in the generator stator winding, the actual moment of the fault is used as the time reference point, and data from 671 sampling moments before and after that moment are extracted as the samples to be diagnosed (i.e., ...). =671), select 120 consecutive sampling times (i.e. The data is constructed within a time window, with a window length of 2 hours. A sliding window mechanism is used, with the window sliding step set to one sampling time (i.e., ...). Set parameter M=3, which is the local score. The top three variables are defined as the difference between the average scores of nodes before and after a window, which is used as the Comprehensive State Score (DNDS) of the turbine generator stator winding for that window, to quantify the dynamic correlation and change characteristics between variables. To reduce the false alarm rate, a multi-window criterion mechanism is introduced: if the Comprehensive State Score (DNDS) changes abruptly at least three times within five consecutive time windows, the warning trigger condition is met.

[0142] Local scores for each monitored variable are calculated using a sliding window. Mapped to a 3D graph, such as Figure 4 As shown, X represents the variable number, the Y-axis represents the sliding time window, and the Z-axis represents the DND local score. Normal operation period (green zone): Within windows 0 to 350, the local DND scores of all variables remain at low levels with minimal fluctuations. This indicates that the system is in a stable equilibrium state, and the correlation between variables is dominated by normal physical coupling. Critical transition period (orange zone): Starting from window 413, the system enters the critical slowdown phase. The local DND scores of some variables begin to deviate from the baseline, showing a significant upward trend. This phenomenon occurs approximately 225 time windows earlier than the actual fault occurrence (window 638). Fault outbreak period (red zone): After window 638, the fault characteristics fully manifest, and the DND scores of specific variables explode exponentially, reaching a peak, indicating a fundamental topological change in the network structure.

[0143] Local scores for each monitored variable are calculated using a sliding window. Choose the one with the highest score. Each node is defined as the difference between the average scores of nodes before and after a window, which is used as the comprehensive state score (DNDS) of the turbine generator stator winding. Early thermal fault warnings for the turbine generator stator winding are then implemented based on the dynamic changes in the comprehensive state score. For example... Figure 5 As shown, the solid green line represents the overall state score, the red "X" marks detected abnormal abrupt changes, the orange dashed line represents the threshold, and the red vertical dashed line indicates the moment of the first anomaly (window 413), indicating a warning. Normal operating state: In windows 0 to 400, the green curve fluctuates slightly around 0 and remains below the orange threshold line, indicating that under normal operating conditions, the overall state score changes smoothly without abrupt changes. Critical state: In windows 413 to 638, the green curve momentarily forms a sharp peak, indicating an increase in anomalies, signifying the moment the system reaches a critical point. The curve then quickly falls back, slightly higher than the normal period, entering a new critical stable state. Fault state: Window 638 and subsequent windows.

[0144] Figure 6 The dynamic deviation network structure is shown at the instant the system triggers the warning signal (window 413). Figure 3 Compared to the reference network, the network topology at this moment exhibits significant structural sparsity and local clustering characteristics. The original dense structure disappears, and only a very few strongly activated edges remain in the network. This intuitively reflects that the system has entered a critical state, and the cooperative relationships within the system have undergone drastic changes and the decoupling between the internal and external systems has occurred.

[0145] like Figure 7 As shown, the present invention also discloses a steam turbine generator stator winding thermal fault early warning device, comprising:

[0146] The data acquisition module 701 is used to acquire continuous temperature variable data of the stator winding of the steam turbine generator; and to use the continuously sampled historical temperature health data as a reference dataset.

[0147] The directed reference network building module 702 is used to construct directed edges by mining the causal dependencies between variables using the variables of the reference dataset as nodes, thus obtaining a directed reference network.

[0148] The dynamic bias network construction module 703 is used to update the temperature variable data to be diagnosed using a sliding window, and to construct a dynamic bias network based on the current window of the temperature variable data to be diagnosed and the causal bias towards the reference network.

[0149] The local score acquisition module 704 is used to treat each node of the dynamic bias network and all its first-order downstream nodes as local modules of that node, and to calculate the local score of each node's comprehensive local module internal bias, internal synergy and external decorrelation.

[0150] The time window average score acquisition module 705 is used to select several nodes with the highest local scores within the current time window and calculate their average score.

[0151] The thermal fault determination module 706 is used to take the average score difference between the current time window and the previous time window as the comprehensive status score DNDS of the turbine generator stator winding; and to identify the turbine generator stator winding with abnormal dynamic changes in DNDS value as a thermal fault and issue an early warning.

[0152] The specific implementation of the steam turbine generator stator winding thermal fault early warning device is the same as the steam turbine generator stator winding thermal fault early warning method, and will not be repeated in this embodiment.

[0153] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for early warning of thermal faults in the stator winding of a steam turbine generator, characterized in that, Includes the following steps: S1, acquire the continuous temperature variable data of the turbine generator stator winding; use the continuously sampled historical temperature health data as a reference dataset; S2 uses the variables of the reference dataset as nodes to mine the causal dependencies between variables to construct directed edges, thus obtaining a directed reference network; S3, use a sliding window to update the temperature variable data to be diagnosed, and construct a dynamic bias network based on the causal deviation between the current window of the temperature variable data to be diagnosed and the directed reference network; S4, take each node of the dynamic bias network and all its first-order downstream nodes as local modules of that node, and calculate the local scores of the comprehensive local module internal bias, internal synergy and external decorrelation of each node. S5: Select the nodes with the highest local scores within the current time window and calculate their average score; S6, the difference between the average score of the current time window and the previous time window is used as the comprehensive status score DNDS of the turbine generator stator winding; the turbine generator stator winding with abnormal dynamic changes in DNDS value is identified as a thermal fault and an early warning is issued.

2. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 1, characterized in that, The process involves using variables from a reference dataset as nodes to mine causal dependencies between variables to construct directed edges, resulting in a directed reference network; specifically as follows: Based on the reference dataset, the time-delay causal relationship and instantaneous coupling relationship between variables are learned through statistical tests, resulting in directed time-delay causal edges and undirected undirected causal networks; Using time-delayed causal edges and undirected causal networks as constraints, a hill-climbing search is performed on the undirected edges to determine the optimal direction of the instantaneous edges, resulting in a complete causal network with timestamps. The complete causal network with timestamps is mapped to a directed reference network containing only physical variables. The Pearson correlation coefficient of each edge in the directed reference network is calculated using the reference dataset and used as the association strength weight of that edge.

3. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 2, characterized in that, The hill-climbing search uses the minimum description length scoring criterion.

4. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 2, characterized in that, Based on the reference dataset, the time-delay causal relationships and instantaneous coupling relationships between variables are learned through statistical tests, resulting in directed time-delay causal edges and undirected undirected causal networks, as detailed below: Perform a rank transformation on the reference dataset to obtain the rank-transformed data; A candidate set of lagged parent nodes for the target variable is constructed based on the data after rank transformation; the candidate set of lagged parent nodes includes the values ​​of all other variables over a past time window; Using the partial rank correlation coefficient, the net correlation between each candidate lagged variable and the target variable in the candidate set of lagged parent nodes is calculated sequentially, and the conditional independence test is performed. For two variables that do not meet the conditional independence test, construct a time-delayed causal edge from the candidate lagged variable to the target variable.

5. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 4, characterized in that, The undirected causal network is obtained through the following steps: A forward heuristic search is performed, specifically: for each target variable in the rank-transformed data, the Max-Min heuristic strategy is used to evaluate and filter all other variables at the same time to obtain the instantaneous candidate parent-child set of the target variable; Perform backward conditional independence tests and screening on the instantaneous candidate parent-child sets to obtain the instantaneous parent-child sets; specifically: if a conditional set exists... If P > θ, then remove E from the instantaneous candidate parent-child set; otherwise, keep E; where, Represents an instantaneous candidate parent-child set; Represents the set of lagging parent nodes; Represents the union; This indicates the removal of candidate nodes from the instantaneous candidate parent-child set. ; Using all variables as nodes, undirected edges are used to connect the target variable and the variables in its instantaneous parent and child sets to obtain an undirected causal network.

6. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 1, characterized in that, The dynamic bias network is constructed based on the causal deviation between the current window of the temperature variable data to be diagnosed and the directed reference network, as detailed below: Calculate the Pearson correlation coefficients for any two variables within the current time window of the directed reference network and the temperature variable data to be diagnosed; The statistical significance of the two Pearson correlation coefficients was tested using a two-sample Fisher Z test, and the statistical significance score was obtained. An edge between two variables whose statistical significance scores are greater than a preset significance threshold is considered to have undergone a significant perturbation at the current time; otherwise, the edge is considered not to have undergone a significant perturbation. The dynamic bias network is constructed by using all variables as nodes, edges that have undergone significant perturbations as edges, and the absolute value of the difference in the Pearson correlation coefficient between two variables as the weight of the edge.

7. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 6, characterized in that, The Pearson correlation coefficient is expressed as: ； in, Representing variables and variables The Pearson correlation coefficient at time t; Indicates the length of the sliding window; Indicates timestamp as Time variable Observed values; Indicates the first in the sliding window One data point; Indicates timestamp as Time variable Observed values; Representing variables The arithmetic mean of all observations within the current sliding window; Representing variables The arithmetic mean of all observations within the current sliding window.

8. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 1, characterized in that, S6 further includes: mapping the scores of nodes corresponding to all variables within all time windows to a three-dimensional coordinate graph, and drawing a three-dimensional visualization of the entire process.

9. The method for early warning of thermal faults in the stator winding of a steam turbine generator according to claim 1, characterized in that, The local score is expressed as: ； in, express Nodes of the time window Local scores; express Nodes of the time window The internal deviation of a local module reflects the internal deviation of the local module; express Nodes of the time window The strength of abnormal correlations outside the local module reflects the decorrelation outside the local module; express Nodes of the time window The abnormal driving strength within a local module reflects the internal coordination of the local module; This indicates that positive numbers with a denominator of 0 should be prevented.

10. A thermal fault early warning device for the stator winding of a steam turbine generator, characterized in that, Including the following: The data acquisition module is used to acquire continuous temperature variable data of the stator winding of the steam turbine generator; and uses continuously sampled historical temperature health data as a reference dataset. The directed reference network building module is used to construct directed edges by mining the causal dependencies between variables using the variables of the reference dataset as nodes, thus obtaining a directed reference network. The dynamic bias network construction module is used to update the temperature variable data to be diagnosed using a sliding window, and to construct a dynamic bias network based on the causal bias between the current window of the temperature variable data to be diagnosed and the directed reference network. The local score acquisition module is used to treat each node of the dynamic bias network and all its first-order downstream nodes as local modules of that node, and to calculate the local score of each node's comprehensive local module internal bias, internal synergy and external decorrelation. The time window average score acquisition module is used to select several nodes with the highest local scores within the current time window and calculate their average score. The thermal fault determination module is used to take the average score difference between the current time window and the previous time window as the comprehensive status score DNDS of the turbine generator stator winding. The stator winding of the steam turbine generator exhibiting abnormal dynamic changes in DNDS value is identified as a thermal fault and an early warning is issued.

Images