Generator stator winding thermal fault early warning method and device based on koci causal network

Abstract

The application discloses a generator stator winding thermal fault early warning method and device based on a KOCMI causal network, and relates to the field of data processing.The method comprises the following steps: S1, collecting continuous sampling temperature data of the stator winding and preprocessing; S2, taking the continuous health samples of the preprocessed temperature data as reference samples of the causal network, and constructing the causal network by using a KOCMI method; S3, using a sliding window method to update the sampling temperature data in real time, and calculating the local entropy and the standard deviation of each node in the causal network under the current time window; S4, calculating the difference value of the local entropy and the difference value of the standard deviation of each node between the current time window and the previous time window; S5, calculating the global network entropy score under the current time window based on the difference value; and S6, performing fault early warning according to the abnormal dynamic change of the global network entropy score of the continuous time window.The application does not need to establish a model, does not need to train a large number of historical fault data, has strong generalization ability, and is helpful to early warning of the generator.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method and device for early warning of thermal faults in generator stator windings based on KOCMI causal networks. Background Technology

[0002] Steam turbine generators, as crucial equipment in power generation, are characterized by their complex structure and unique operating environment. With the continuous improvement of steam turbine generator unit capacity and efficiency, more complex engineering and operation and maintenance challenges have also emerged. Traditional preventative maintenance methods relying on fixed times or cycles are no longer sufficient to meet the high-efficiency operation requirements of modern units. The industry urgently needs to introduce more advanced operation and maintenance concepts and technologies to improve overall operating efficiency and ensure a continuous and stable power supply. Therefore, constructing a fault early warning system based on real-time data has significant engineering application value and practical significance for ensuring the safe operation of units, reducing maintenance costs, and improving the utilization rate and economy of unit operation.

[0003] Data-driven AI-based fault prediction technology has been widely applied in engineering fields due to the current development of artificial intelligence. It uses real-time data acquired by sensors as training samples to learn and achieve accurate and effective fault diagnosis.

[0004] Some scholars have proposed a stator winding fault monitoring method based on intelligent stator current data. This method uses short-time Fourier transform to extract fault features from the symmetrical components of the stator phase current signal, and then employs machine learning algorithms such as support vector machine, Naive Bayes classifier, and multilayer perceptron to achieve fault diagnosis. Other scholars have combined sparse autoencoders and long short-term memory networks to construct a temperature prediction model, and further combined this model with the sliding window method to create a thermal fault diagnosis model for turbine generator stator windings. This improves the accuracy of generator stator winding temperature prediction and enables early warning of thermal faults in turbine generator stator windings. However, these data-driven artificial intelligence methods have certain limitations:

[0005] 1. The demand for data samples, especially fault sample data, is large, but the relative scarcity of unit fault data limits the training of artificial intelligence models to identify abnormal states of steam turbine generator units.

[0006] 2. Training models consumes a lot of computing resources, and the high demand for computing power leads to high costs for supporting computing equipment.

[0007] 3. Traditional artificial intelligence models and methods struggle to capture the nonlinear coupling relationships between multiple variables, exhibiting poor adaptability and weak generalization ability when faced with complex operating conditions of steam turbine generator sets.

[0008] Studies of numerous complex systems have consistently revealed that systems, under the influence of long-term external disturbances or other factors, gradually deviate from their original stable operating state. When these influences persist, complex systems often transition from one stable state to another. A critical point exists during this state transition, near which the system exhibits critical slowdown. To capture this state transition, scholars have proposed Dynamic Network Markers (DNMs) to describe the critical dynamic characteristics of multivariable complex systems. Chinese invention patent application CN117992762A discloses a method for early warning of thermal faults in steam turbine generator sets based on DNM theory. However, this method relies on Pearson correlation coefficients for score calculation, and still suffers from limited model accuracy and high computational cost when dealing with complex nonlinear systems. Summary of the Invention

[0009] To address the above problems, this invention proposes a generator stator winding thermal fault early warning method and device based on KOCMI causal network. Based on collected stator winding monitoring data, a dynamic causal network of measurement point variables is established using the KOCMI method. By calculating the local network entropy and standard deviation of the causal network nodes, a global network entropy for stator winding fault early warning is constructed to identify whether a thermal fault has occurred in the stator winding. This method does not require model building, extensive training with historical fault data, or calculation of the three statistical conditions of the traditional DNM method. It exhibits strong generalization ability and facilitates early warning of generator faults.

[0010] On the one hand, the generator stator winding thermal fault early warning method based on KOCMI causal network has the following specific steps:

[0011] S1, collect and preprocess continuous sampling temperature data of the stator winding of the steam turbine generator;

[0012] S2, using the continuous healthy samples of preprocessed temperature data as reference samples for the causal network, and constructing the causal network using the KOCMI method;

[0013] S3, the preprocessed sampled temperature data is updated in real time using the sliding window method, and the local entropy and standard deviation of each measuring point node in the causal network under the current time window are calculated;

[0014] S4, calculate the difference in local entropy and standard deviation of each measuring point node between the current time window and the previous time window;

[0015] S5, calculate the global network entropy score under the current time window based on the difference between local entropy and the difference between standard deviation;

[0016] S6 provides fault warnings based on the abnormal dynamic changes in the global network entropy score over continuous time windows.

[0017] Preferably, the construction of the causal network using the KOCMI method is as follows:

[0018] Construct a data matrix of temperature data for the initial input time window. ;

[0019] For any two variables in the data matrix and ,by As the cause variable, As the outcome variable, and the remaining variables as condition variables, calculate... and The conditional mutual information value is used to obtain the original conditional mutual information value;

[0020] Data matrix Remove Variable column obtains subspace ;

[0021] Generate variables knockoff variable Use the knockoff variable Replace subspace variables in and calculate and The conditional mutual information value is obtained by replacing the original conditional mutual information value; the difference between the replaced conditional mutual information value and the original conditional mutual information value is calculated.

[0022] After performing the previous step several times to obtain several differences, a sign substitution test is performed to obtain the substitution test value. Two variables with a value less than the preset substitution test threshold are considered to have a causal relationship, and the causal strength of the two variables is calculated; otherwise, the two variables have no causal relationship.

[0023] A causal network is constructed by using each variable as a node and the causal strength between variables as edges connecting the nodes.

[0024] Preferably, the generated variable knockoff variable , is represented as:

[0025] ;

[0026] in, Represented as variables The average of all sampled values ​​within the current time window; Represents a centralized subspace, which is composed of subspaces The average value of the corresponding variable in each column is obtained by subtracting the average value of each column from the data in each column. Represents the identity matrix; The inverse matrix of the covariance matrix; Represents the difference factor matrix; This represents random noise.

[0027] Preferably, the causal strength is expressed as:

[0028] ;

[0029] in, Indicates the strength of the causal relationship between variables; Indicates the difference The absolute value of the average of all elements in the mixture; Indicates the difference Standard deviation; This represents a constant to prevent the denominator from being zero.

[0030] Preferably, the calculation of the local entropy and standard deviation of each node in the causal network under the current time window is as follows:

[0031] Calculate the causal proportion of each node's neighborhood to each node in a causal network; specifically, select a central node and calculate the causal proportion of each neighboring node to the central node, expressed as:

[0032] ;

[0033] in, Representing neighboring nodes The causal proportion of the central node; Representing neighboring nodes The absolute value of the causal strength between the central node and the central node. To express summation;

[0034] Calculate the standard deviation of each node in the current time window;

[0035] Taking each node as the center, calculate the local network entropy of each node in the current time window, expressed as:

[0036] ;

[0037] in, Represents a node Local network entropy within the current time window. Represents a node Neighboring nodes For nodes The causal proportion, where L represents the node. The number of its neighbors.

[0038] Preferably, the global network entropy score for the current time window is calculated based on the difference between local entropy and the difference between standard deviation, and is expressed as:

[0039] ;

[0040] in, This represents the global network entropy score for the (w+1)th time window; This indicates the time interval between the w-th time window and the (w+1)-th time window. The difference in local network entropy of each variable; This indicates the time interval between the w-th time window and the (w+1)-th time window. The difference in the standard deviations of the variables; Indicates the number of variables.

[0041] Preferably, the step of providing fault warning based on the abnormal dynamic changes in the global network entropy score over consecutive time windows specifically involves issuing a signal to provide fault warning when s or more instances of exceeding a preset global network entropy score threshold occur within five consecutive time windows.

[0042] Preferably, the preprocessing includes normalizing the temperature data.

[0043] Preferably, the temperature data is acquired through a DCS system.

[0044] On the other hand, the generator stator winding thermal fault early warning device based on KOCMI causal network includes the following:

[0045] The temperature data acquisition module is used to collect and preprocess continuous sampling temperature data of the stator winding of the steam turbine generator;

[0046] The KOCMI causal network construction module is used to construct causal networks by using continuous healthy samples of preprocessed temperature data as reference samples for causal networks and employing the KOCMI method.

[0047] The local entropy and standard deviation calculation module is used to update the sampled temperature data in real time after preprocessing using the sliding window method, and to calculate the local entropy and standard deviation of each measuring point node in the causal network under the current time window.

[0048] The difference calculation module is used to calculate the difference in local entropy and standard deviation of each measuring point node between the current time window and the previous time window.

[0049] The global network entropy score calculation module is used to calculate the global network entropy score under the current time window based on the difference between local entropy and the difference between standard deviation.

[0050] The fault warning module is used to provide fault warnings based on the abnormal dynamic changes in the global network entropy score over a continuous time window.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] This invention utilizes the KOCMI method to establish a dynamic causal network of measurement point variables based on the stator winding monitoring data collected by the DCS system of the target steam turbine generator set. By calculating the LNE (Local Network Entropy) value and standard deviation of the network nodes, a fault early warning index GNE (Global Network Entropy) for the stator winding is constructed to identify whether a thermal fault has occurred in the stator winding. This method does not require model building, training with a large amount of historical fault data, or calculation of the three statistical conditions of the traditional DNM method. It has strong generalization ability and helps to provide early warning for generators. Attached Figure Description

[0053] The present invention will now be described in further detail with reference to the accompanying drawings;

[0054] Figure 1 This is a flowchart illustrating the steps of the generator stator winding thermal fault early warning method based on KOCMI causal network according to an embodiment of the present invention.

[0055] Figure 2 This is a flowchart of a generator stator winding thermal fault early warning method based on KOCMI causal network according to an embodiment of the present invention;

[0056] Figure 3 This is a causal network diagram showing the relationships between input variables under normal operating conditions of the generator stator winding thermal fault early warning method based on KOCMI causal network according to an embodiment of the present invention.

[0057] Figure 4 This is a causal network diagram of the various input variables during a fault in the generator stator winding thermal fault early warning method based on KOCMI causal network according to an embodiment of the present invention.

[0058] Figure 5 This is a graph showing the trend of GNE value changing with time window during the evolution of thermal fault in the target steam turbine generator stator winding of the generator stator based on the KOCMI causal network-based early warning method according to an embodiment of the present invention.

[0059] Figure 6 This is a structural block diagram of a generator stator winding thermal fault early warning device based on a KOCMI causal network according to an embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described below through specific embodiments.

[0061] like Figure 1 and Figure 2 As shown, the generator stator winding thermal fault early warning method based on KOCMI causal network has the following specific steps:

[0062] S1 collects and preprocesses continuous sampling temperature data of the stator winding of the steam turbine generator.

[0063] The data collected by the DCS system from the stator winding sensor monitoring points needs to be preprocessed. The preprocessing method is to normalize the data.

[0064] ;

[0065] in, For variable values Normalized temperature data; For the collected temperature variable values, This represents the average value of each of the collected variables; The standard deviation of each of the collected variables.

[0066] S2 uses continuous healthy samples of preprocessed temperature data as reference samples for the causal network and constructs the causal network using the KOCMI method.

[0067] Based on m initial sample data sets, the KOCMI method is used to calculate the causal relationships between input variables within this time window, constructing an initial causal network. The local entropy and standard deviation of each node in the initial causal network are calculated, with the following specific steps:

[0068] S21, select the preprocessed monitoring data from m consecutive sampling times, n variables, and m sampling times to construct an input data matrix. , is represented as:

[0069] ;

[0070] in, Indicates the first The variable in the current time window The variable values ​​at each sampling time.

[0071] S22, use the KOCMI method to calculate the causal relationships between the input variables within this time window and construct an initial causal network. Calculate the original conditional mutual information values ​​between the variables within this time window. Suppose we calculate the original CMI value of variable X → variable Y, the calculation process is as follows:

[0072] ;

[0073] in, As the cause variable, For the outcome variable, For the remaining variables, for CMI value;

[0074] The Digamma function is defined as follows: ,in, For the generalization of factorial, it is represented as:

[0075] ;

[0076] Let be the number of neighbors in the K-nearest neighbors algorithm. This is the sampling time number (i.e., which group of sampling times) within the current time window.

[0077] Get nearest neighbor distance : Calculate the first The Chebyshev distances between the group of sampled data and other sampled data are calculated, and the results are arranged in ascending order. This represents the distance to the Kth non-zero value. The Chebyshev distance calculation formula is as follows:

[0078] ;

[0079] in, Indicates the first The data from the sampling time group are used as the center to calculate its relationship with the first group. Chebyshev distance between data points at different sampling times Indicates the first Group sampling time and the first The maximum absolute value of the difference between each variable at the sampling time of the group.

[0080] Indicates the first Centered on the group sampling time, in the data matrix Remove After the data in the variable column, the subspace In the nearest neighbor distance The number of samples within. The calculation formula is as follows:

[0081] ;

[0082] = ;

[0083] ;

[0084] in, Indicates the first Group sampling time The values ​​of each variable, Including, besides n-2 variables other than the variable, Indicates the first Group sampling time and the first Group sampling time subspace The maximum absolute value of the differences between the variables. Indicates the first The nearest neighbor distance centered on the data at the group sampling time; The function is an indicator function; it is 1 if the condition is met, and 0 otherwise. Indicates the first Centered on the group sampling time, in the data matrix Remove After the data in the variable column, the subspace In the nearest neighbor distance Number of samples within The calculation formula is as follows:

[0085] ;

[0086] ;

[0087] in, Indicates the first Group sampling time The values ​​of each variable, Indicates the first The nearest neighbor distance centered on the data at the group sampling time;

[0088] Indicates the first Centered on the group sampling time, in the data matrix Remove and After the data in the variable column, the subspace In the nearest neighbor distance Number of samples within The calculation formula is as follows:

[0089] ;

[0090] ;

[0091] in, Indicates the first Group sampling time The values ​​of each variable, Indicates the first Nearest neighbor distance centered on the group sampling time data

[0092] S23, based on the input matrix Remove Subspace of variable column Generate G variables knockoff variable :

[0093] ;

[0094] in, Represented as variables The average of all sampled values ​​within the current time window;

[0095] Represents a centralized subspace, which is composed of subspaces The average of the corresponding variable values ​​in each column is subtracted from the data in each column. For example, in the column corresponding to variable X, each data point should be subtracted from the average of that variable. It's important to note that Z represents not a single variable, but all variables other than X and Y; therefore, the average of the corresponding variable in each column of Z needs to be subtracted from the data in that column separately.

[0096] Represents the identity matrix. Represents the covariance matrix The inverse matrix and covariance matrix The calculation process is as follows:

[0097] ;

[0098] in, Represent the covariance matrix; This represents the operation of covariance between variable X and itself; Represents a centralized subspace; Indicates transpose; express Each sampling time;

[0099] The difference factor matrix is ​​represented as follows:

[0100] = ;

[0101] Among them, the elements on the diagonal represent the spatiality between the original variable and the ko variable. The selection should satisfy the following conditions: In this case, maximum.( The conditions guarantee ( The covariance matrix of this joint distribution is positive definite, meaning the variance is non-negative. The maximum is to ensure that the generated (Make the differences between the original variables as large as possible).

[0102] Random noise is used to increase the generation. The authenticity is determined through the following calculation process:

[0103] ;

[0104] ;

[0105] ;

[0106] in, A noise matrix whose elements are randomly generated and conform to a standard normal distribution. The noise covariance matrix The triangular matrix obtained through triangular decomposition.

[0107] S24, using the knockoff variable Replace the original variable Calculate the p-value to determine the variable. To determine whether a causal relationship exists between them, the calculation steps are as follows:

[0108] Generate G knockoff variables Replace the variables in the original input data matrix I Given a data column of G times, calculate the variable under the current time window. Conditional mutual information values ​​between , g represents the variable at the g-th time. Permutation. The formula for the P-value is expressed as:

[0109] ;

[0110] Where F represents the number of times the sign is randomly permuted, and f represents the f-th sign permutation; Represents the variable of the gth order. CMI difference before and after replacement: , , Represents the variable of the Gth order. The difference in CMI before and after the replacement; This represents the conditional mutual information value between variables X and Y before the substitution. This indicates that the knockoff variable has been replaced. With variables Conditional mutual information values ​​between them; This indicates that before the sign substitution is performed, The average value of each element in the mixture; Represents absolute value; This indicates the difference given during the f-th sign permutation. Randomly assigned symbols; This indicates an indicator function; if the condition within the parentheses is true, it takes the value 1; otherwise, it takes the value 0.

[0111] The permutation test measures the difference in random permutations ( ) Determine the signs of the elements in the expression and observe the results before and after the substitution. The reliability of the data is measured by the change in the mean of the elements. P represents the permutation test value. The closer it is to 0, the higher the reliability of the causal relationship. Statistically, 0.05 is generally used as the threshold.

[0112] In the f-th symbolic permutation, it means The sign (+1 or -1) is randomly assigned to the g-th element.

[0113] Indicates the sign before the sign is replaced. The average value of each element in the mixture; This is an indicator function; it is denoted as 1 if the condition is met, and 0 otherwise.

[0114] The causal strength (CS) between variables is calculated using the following formula:

[0115] ;

[0116] in, Indicates the strength of the causal relationship between variables; express The absolute value of the average of all elements in the mixture; Standard deviation; To prevent the constant from having a denominator of 0, it is generally taken as .

[0117] Based on the above, the causal relationships between the variables are calculated. A causal network for the current time window is then constructed based on these relationships, with each variable acting as a node and the causal strength between variables as the edges connecting the nodes.

[0118] S25, the calculation process of the local entropy and standard deviation of each node in the causal network in the initial time window is as follows:

[0119] Calculate the causal relationship between each node and its neighborhood in a causal network. Select a center node and calculate the causal relationship between each node's neighborhood and the center node. :

[0120] ;

[0121] in, Representing neighboring nodes The causal proportion of the central node; Representing neighboring nodes The absolute value of the causal strength between the central node and the central node. It represents the sum of the absolute values ​​of the causal strength between the nodes in the neighborhood and the central node.

[0122] Calculate the standard deviation of each node in the initial time window. .

[0123] Calculate the local network entropy of each node in the initial time window. , is represented as:

[0124] ;

[0125] in, Represents a node Local network entropy within the initial time window, Represents a node Neighboring nodes For nodes The causal proportion, where L represents the node. The number of its neighbors.

[0126] S3 uses the sliding window method to update the preprocessed sampled temperature data in real time, and calculates the local entropy and standard deviation of each measuring point node in the causal network under the current time window.

[0127] A sliding window method is used, with a window size of m and a step size of b. The initial window number is denoted as w, where w=0. The window slides forward, incorporating data from the next sampling time step and updating the data samples for the current time window. The local entropy of each node in the causal network is calculated for the current time window (w=w+1). with standard deviation According to S25, the local network entropy of each node in the current time window is expressed as: .

[0128] S4 calculates the difference in local entropy and standard deviation between each measuring point node in the current time window and the previous time window.

[0129] Calculate the difference in local entropy and standard deviation of each node between the current time window and the previous time window. The difference in network entropy characterizes the disturbance to the local network structure caused by the addition of new data samples. Under the influence of long-term external disturbances or other factors, the system will gradually deviate from its original stable operating state. When these influences persist, complex systems often transition from one stable state to another. There are critical points in the transition process of system states, and the system exhibits critical slowdown near these points. To capture this transition of system states, the Dynamic Network Marker (DNM) theory was proposed to describe the critical dynamic characteristics of multivariable complex systems. When the critical state is approached, at least one of all variables will fluctuate drastically; this variable is called the dominant group, and the dominant groups are highly correlated. Specifically, when the system approaches a critical state, it can be demonstrated that the dominant variable appears and satisfies the following three conditions: (1) the average standard deviation among key nodes in the network increases significantly; (2) the average Pearson correlation coefficient among key nodes in the network increases; and (3) the average Pearson correlation coefficient between key and non-key nodes in the network decreases. The difference in network entropy reflects the three statistical conditions in DNM theory. Therefore, the global network entropy value can be calculated based on the local network entropy value and standard deviation of each node to measure the overall fluctuation of the system and identify early signals of failure.

[0130] The difference in network entropy between nodes in two adjacent time windows is expressed as:

[0131] ;

[0132] in, This indicates the time interval between the (w+1)th time window and the wth time window. The local network entropy difference of each variable. Indicates the w-th time window. The local network entropy value of each variable. This indicates the (w+1)th time window. The local network entropy value of each variable.

[0133] The difference in standard deviation between nodes in two adjacent time windows is expressed as follows:

[0134] ;

[0135] in, This indicates the time interval between the (w+1)th time window and the wth time window. The standard deviation of each variable Indicates the w-th time window. The standard deviation of each variable This indicates the (w+1)th time window. The standard deviation of each variable.

[0136] S5 calculates the global network entropy score for the current time window based on the difference between the local entropy and the standard deviation.

[0137] Based on the difference between the local network entropy values ​​and standard deviations of each variable between the current time window (w+1) and the previous time window (w), the global network entropy score of the current time window (w+1) is calculated as a comprehensive state index of the system.

[0138] ;

[0139] Calculate the global network entropy score for all time windows, and project the global network entropy score onto a two-dimensional coordinate axis according to the time window. The resulting graph has the horizontal axis representing the time window number and the vertical axis representing the global network entropy score.

[0140] S6 provides fault warnings based on the abnormal dynamic changes in the global network entropy score over continuous time windows.

[0141] Fault warnings are issued based on the dynamic changes in the global network entropy score of each time window. If the global network entropy score exceeds the threshold in a or more time windows within c consecutive time windows, the system is judged to have entered a critical state and a warning signal is issued.

[0142] Verification experiments were conducted on embodiments of the present invention, using a large steam turbine generator unit in a power plant as the object. Data acquisition was completed by the unit's DCS system, with a sampling period of 1 minute. This experiment included 300 consecutive sampling time points, including the fault time. The first 50 consecutive sampling data points (m=50) were used as the initial time window to construct an initial causal network using the KOCMI method. A sliding window method was adopted, with the window size set to 50 sampling time points and the window sliding step size set to 1 sampling time point. Figure 3 As shown in the figure, this diagram illustrates the causal network between the input variables under normal operating conditions, constructed using the KOCMI method. Figure 4 The diagram shows the causal network between the input variables during a fault state. Each node in the network represents an input variable (different temperature measurement points of the stator winding sensor), and the lines connecting the nodes represent the causal relationships between these input variables. The thickness of the lines indicates the strength of the causal relationship. A comparison of the two diagrams clearly shows that as the sliding window incorporates new samples, the causal relationships between the input variables become stronger, the coupling becomes tighter, and the system enters a critical state.

[0143] The Global Network Entropy Score (GNE) is calculated for all time windows using the GNE method, and early warnings are issued based on the dynamic fluctuations of the GNE score. To reduce false alarms caused by sampling interference, continuous early warning monitoring is introduced. A fault warning is issued when the GNE value exceeds a set threshold in three or more consecutive time windows out of five. Figure 5 As shown in the figure, this graph illustrates the change in global network entropy score over time after calculating the network entropy score for all time windows using the GNE method. The horizontal axis represents the time window number, the vertical axis represents the GEN value, the red dashed line represents the GNE threshold, and the red crosses indicate points exceeding the threshold. Figure 5 As can be seen from the first two hundred normal operating time windows, the GNE score, although fluctuating slightly, tends to be stable overall. The first warning window is numbered 208, indicating that the system has entered a critical state. After the first warning window, i.e., in the critical state, the GNE score begins to fluctuate significantly, with the curve frequently exceeding the threshold. The time window when the GNE score reaches its peak is numbered 226, which corresponds to the actual failure time.

[0144] like Figure 6 As shown, the present invention also discloses a generator stator winding thermal fault early warning device based on a KOCMI causal network, comprising:

[0145] Temperature data acquisition module 601 is used to acquire and preprocess continuous sampled temperature data of the stator winding of steam turbine generator;

[0146] KOCMI Causal Network Construction Module 602 is used to construct a causal network by using continuous healthy samples of preprocessed temperature data as reference samples for the causal network and employing the KOCMI method.

[0147] The local entropy and standard deviation calculation module 603 is used to update the sampled temperature data in real time after preprocessing using the sliding window method, and to calculate the local entropy and standard deviation of each measuring point node in the causal network under the current time window.

[0148] The difference calculation module 604 is used to calculate the difference in local entropy and standard deviation of each measuring point node between the current time window and the previous time window.

[0149] The global network entropy score calculation module 605 is used to calculate the global network entropy score under the current time window based on the difference between the local entropy and the difference between the standard deviation.

[0150] The fault warning module 606 is used to provide fault warnings based on the abnormal dynamic changes in the global network entropy score over a continuous time window.

[0151] The specific implementation of the generator stator winding thermal fault early warning device based on KOCMI causal network is the same as the generator stator winding thermal fault early warning method based on KOCMI causal network, and will not be described again in this embodiment.

[0152] 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 generator stator winding thermal fault early warning method based on KOCMI causal network, characterized in that, Includes the following steps: S1, collect and preprocess continuous sampling temperature data of the stator winding of the steam turbine generator; S2, using the continuous healthy samples of preprocessed temperature data as reference samples for the causal network, and constructing the causal network using the KOCMI method; S3, the preprocessed sampled temperature data is updated in real time using the sliding window method, and the local entropy and standard deviation of each measuring point node in the causal network under the current time window are calculated; S4, calculate the difference in local entropy and standard deviation of each measuring point node between the current time window and the previous time window; S5, calculate the global network entropy score under the current time window based on the difference between local entropy and the difference between standard deviation; S6, based on the abnormal dynamic changes in the global network entropy score over a continuous time window, provides fault warnings; The construction of the causal network using the KOCMI method is as follows: Construct a data matrix of temperature data for the initial input time window. ; For any two variables in the data matrix and ,by As the cause variable, As the outcome variable, and the remaining variables as condition variables, calculate... and The conditional mutual information value is used to obtain the original conditional mutual information value; Data matrix Remove Variable column obtains subspace ; Generate variables knockoff variable Use the knockoff variable Replace subspace variables in and calculate and The conditional mutual information value is used to obtain the replaced conditional mutual information value; Calculate the difference between the conditional mutual information value after replacement and the original conditional mutual information value; After performing the previous step several times to obtain several differences, a sign substitution test is performed to obtain a substitution test value. Two variables with a value less than the preset substitution test threshold are identified as having a causal relationship, and the causal strength of the two variables is calculated. Otherwise, the two variables have no causal relationship; A causal network is constructed by using each variable as a node and the causal strength between variables as edges connecting the nodes.

2. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The generated variables knockoff variable , is represented as: ； in, Represented as variables The average of all sampled values ​​within the current time window; Represents a centralized subspace, which is composed of subspaces The average value of the corresponding variable in each column is obtained by subtracting the average value of each column from the data in each column. Represents the identity matrix; The inverse matrix of the covariance matrix; Represents the difference factor matrix; This represents random noise.

3. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The causal strength is expressed as: ； in, Indicates the strength of the causal relationship between variables; Indicates the difference The absolute value of the average of all elements in the mixture; Indicates the difference Standard deviation; This represents a constant to prevent the denominator from being zero.

4. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The calculation of the local entropy and standard deviation of each node in the causal network under the current time window is as follows: Calculate the causal proportion of each node's neighborhood to each node in a causal network; specifically, select a central node and calculate the causal proportion of each neighboring node to the central node, expressed as: ； in, Representing neighboring nodes The causal proportion of the central node; Representing neighboring nodes The absolute value of the causal strength between the central node and the central node. To express summation; Calculate the standard deviation of each node in the current time window; Taking each node as the center, calculate the local network entropy of each node in the current time window, expressed as: ； in, Represents a node Local network entropy within the current time window. L represents a node. The number of its neighbors.

5. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The global network entropy score for the current time window is calculated based on the difference between local entropy and standard deviation, and is expressed as follows: ； in, This represents the global network entropy score for the (w+1)th time window; This indicates the time interval between the w-th time window and the (w+1)-th time window. The difference in local network entropy of each variable; This indicates the time interval between the w-th time window and the (w+1)-th time window. The difference in the standard deviations of the variables; Indicates the number of variables.

6. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The method of providing fault warning based on abnormal dynamic changes in the global network entropy score over consecutive time windows is as follows: when a or more times the global network entropy score exceeds a preset threshold within c consecutive time windows, a signal is issued to provide fault warning; where c is greater than or equal to a.

7. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The preprocessing includes normalizing the temperature data.

8. The generator stator winding thermal fault early warning method based on KOCMI causal network according to claim 1, characterized in that, The temperature data is collected through a DCS system.

9. A generator stator winding thermal fault early warning device based on KOCMI causal network, characterized in that, Including the following: The temperature data acquisition module is used to collect and preprocess continuous sampling temperature data of the stator winding of the steam turbine generator; The KOCMI causal network construction module is used to construct causal networks by using continuous healthy samples of preprocessed temperature data as reference samples for causal networks and employing the KOCMI method. The local entropy and standard deviation calculation module is used to update the sampled temperature data in real time after preprocessing using the sliding window method, and to calculate the local entropy and standard deviation of each measuring point node in the causal network under the current time window. The difference calculation module is used to calculate the difference in local entropy and standard deviation of each measuring point node between the current time window and the previous time window. The global network entropy score calculation module is used to calculate the global network entropy score under the current time window based on the difference between local entropy and the difference between standard deviation. The fault warning module is used to provide fault warnings based on the abnormal dynamic changes in the global network entropy score over a continuous time window. The construction of the causal network using the KOCMI method is as follows: Construct a data matrix of temperature data for the initial input time window. ; For any two variables in the data matrix and ,by As the cause variable, As the outcome variable, and the remaining variables as condition variables, calculate... and The conditional mutual information value is used to obtain the original conditional mutual information value; Data matrix Remove Variable column obtains subspace ; Generate variables knockoff variable Use the knockoff variable Replace subspace variables in and calculate and The conditional mutual information value is used to obtain the replaced conditional mutual information value; Calculate the difference between the conditional mutual information value after replacement and the original conditional mutual information value; After performing the previous step several times to obtain several differences, a sign substitution test is performed to obtain a substitution test value. Two variables with a value less than the preset substitution test threshold are identified as having a causal relationship, and the causal strength of the two variables is calculated. Otherwise, the two variables have no causal relationship; A causal network is constructed by using each variable as a node and the causal strength between variables as edges connecting the nodes.

Images