Hydro-generator stator thermal fault diagnosis method based on time sequence and spatial distribution

Abstract

The present application relates to a kind of water turbine generator stator thermal fault diagnosis method based on timing and spatial distribution.The method includes: division area, establishment graph structure, construct graph model edge relationship matrix M, construct each node feature vector x, construct graph model feature matrix X T , construct timing graph feature matrix X, data acquisition, graph neural network training, nine steps of fault diagnosis.The present application adopts graph neural network and can learn to the temperature distribution characteristics of fault state in stator structure, temperature variation characteristics, can distinguish the influence difference of different types of measuring point temperature under different working conditions in normal state, the temperature difference of measuring point caused by spatial structure, the temperature difference of measuring point caused by temperature change lagging behind load change.The present application method is used for stator thermal fault diagnosis, can improve the sensitivity of analysis and diagnosis stator thermal fault, easy to popularization and application.

Description

Technical Field

[0001] This invention belongs to the field of condition analysis and diagnosis technology for large hydro-generators in smart power plants, and relates to a method for diagnosing stator thermal faults of hydro-generators based on time sequence and spatial distribution. Background Technology

[0002] With the transformation of the energy structure and the integration of hydropower, wind power, and solar power in power operation, the maintenance methods of hydropower plants are gradually transitioning from planned and preventive maintenance to condition-based maintenance. The effectiveness of condition-based maintenance depends on the reliability of the condition analysis results of the equipment under maintenance. Temperature is a key characteristic of stator condition. Poor contact of winding electrical joints, short circuits between strands of conductors, short circuits between turns of the core, aging of core insulation, and abnormal cooling and ventilation will all manifest in the form of heat. The temperature distribution will also change with the cooling efficiency and the spatial location of the fault point, and the temperature distribution of the load curve will also be different at different time series.

[0003] With the continuous increase in generator capacity, taking a 650MW large hydro-generator as an example, its stator diameter exceeds 10 meters, its height exceeds 5 meters, the number of conductor bars exceeds 1000, the core is axially segmented into more than 60 segments, and there are 14 coolers. If the grid is divided according to the number of slots and core segments, it can be divided into 37,950 regions. Radially, it can be further divided into windings, core, frame, and coolers, with a radial width greater than 1 meter. Radially divided according to function and material, the core, windings, and coolers exceed 100,000 smaller regions. Currently, stator temperature measurement is achieved by locally embedded thermal resistors. Taking the above generator as an example, there are 144 measurement points for the windings, 64 for the core, 32 for the pressure plate, and 28 for the coolers. The temperature measurement coverage of the windings (144 / 37,950) is 0.38%, and the temperature measurement coverage of the core (64 / 37,950) is 0.17%. Currently, stator temperature is detected through point-based inspections and threshold-based early warning systems. However, these methods do not consider the spatial distribution and temporal characteristics of temperature. Reflecting the overall stator temperature status through a limited number of measurement points results in low sensitivity for identifying stator overheating faults, leading to unreliable stator condition analysis. Therefore, overcoming the shortcomings of existing technologies is a pressing issue that needs to be addressed in this field. Summary of the Invention

[0004] The purpose of this invention is to address the problems in current stator temperature analysis and diagnosis of large hydro-generators, which suffer from low sensitivity and low reliability due to the limited measurement points used in analysis. These problems include the failure to consider the differences in stator temperature distribution caused by variations in stator spatial structure and measurement point location, the failure to consider the different effects of different operating conditions on the temperature of different types of measurement points, and the failure to consider the temporal characteristics between load P and Q changes and stator temperature rise changes. The invention provides a method for diagnosing stator thermal faults in hydro-generators based on temporal and spatial distribution.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for diagnosing stator thermal faults in hydro-generators based on temporal and spatial distribution includes the following steps:

[0007] Step (1), Divide the area: Divide the stator into areas based on the cooler, and indicate the temperature measurement points of each component in the area, including the winding, core, pressure finger, and cooler.

[0008] Step (2), establish graph structure: take the temperature measurement points obtained in step (1) as nodes, establish graph structure according to the relationship between stator loss heat generation, cooling heat dissipation and cooling air circulation, and set directed connections between nodes according to the generation and transfer of heat when establishing graph structure.

[0009] Step (3), construct the graph model edge relationship matrix M: the connection relationship between each node is represented by a two-dimensional matrix M, and the elements in the matrix M ij This indicates whether there is a connection between nodes i and j, i.e., a directed edge from node i to j. If it exists, then M... ij =1, if not, then M ij =0, n represents the number of nodes in the graph structure, i = 0, 1, 2, 3....n-1, j = 0, 1, 2, 3....n-1, and M is the element in the matrix. ij The value of 0 or 1 is determined based on the interaction relationship between pairs of nodes in the specific physical model;

[0010] Step (4), construct the node feature vector using a one-dimensional array x i express:

[0011] x i =[P0, P1, Q0, Q1, T0, T1]

[0012] Wherein, P0 is the generator active power at the previous moment of the node, P1 is the generator active power at the current moment of the node; Q0 is the generator reactive power at the previous moment of the node, Q1 is the generator reactive power at the current moment of the node; T0 is the temperature at the previous moment of the node, T1 is the temperature at the current moment of the node.

[0013] Step (5): Construct a two-dimensional array X of the graph model feature matrix. Tt :

[0014] X Tt =[x0,x1,x2,...x n-3 ,x n-2 ,x n-1 ]

[0015] n is the number of nodes, 1, 2, 3...n represents the node index, x0, x1, x2,...x n-3 ,x n-2 ,x n-1X represents the n feature vectors of nodes 1 to n in sequence; Tt Let be the feature matrix of the graph model of n nodes at a certain time t;

[0016] Step (6), construct the three-dimensional array X of the time series graph model feature tensors:

[0017] X = [X] T0 ,X T1 ,X T2 ,.....X Tt-3 ,X Tt-2 ,X Tt-1 ]

[0018] The X-series graphical model feature tensor consists of t graphical model feature matrices X. Tt Composition, X T0 X is the feature matrix of the graphical model at time 0. T1 For example, X is the feature matrix of the time-1 graphical model, and so on. Tt-1 The feature matrix of the graphical model at time t-1;

[0019] Step (7), Data Acquisition: Collect historical data of the normal state and fault state of each node from historical data, and mark them using fault tags; the historical data includes generator active power, generator reactive power, temperature and corresponding time;

[0020] The faults mentioned include short circuits between core laminations, short circuits between winding strands, reduced cooler efficiency, abnormal measuring points, and other temperature abnormalities.

[0021] Step (8), training of graph neural network: with two-dimensional matrix M and time series graph feature matrix X corresponding to each state historical data as input, and corresponding fault label Y as output, the graph neural network is trained and verified to obtain the hydro-generator stator thermal fault diagnosis model.

[0022] Step (9), Fault diagnosis: Real-time collection of generator active power, generator reactive power, temperature and corresponding time data of each node, and acquisition of the corresponding time sequence feature matrix X, input into the hydro-generator stator thermal fault diagnosis model obtained in step (8), and the final fault diagnosis result is obtained based on the output Y of the hydro-generator stator thermal fault diagnosis model.

[0023] Furthermore, preferably, in step (2), the connection between nodes is established by single-phase arrows; nodes with constant quantities or measuring points that have no effect on temperature analysis are not introduced, and no connection is established between nodes that have no direct effect.

[0024] Furthermore, preferably, in step (4), when selecting each feature in the feature vector, the time interval between the current time and the previous time is set to 1 minute, and the time interval can be adjusted for different sizes of units.

[0025] Furthermore, preferably, the fault label is 0 for normal conditions, 1 for short circuit between core laminations, 2 for short circuit between winding strands, 3 for reduced cooler efficiency, 4 for abnormal measuring points, and 5 for other abnormal temperatures.

[0026] Other temperature anomalies mentioned in this invention refer to other atypical stator thermal faults, such as blocked core ventilation channels leading to poor ventilation circulation and temperature rise, increased overall core loss and temperature rise, and high temperature of the pressure finger plate. After obtaining the corresponding fault data, add labels sequentially based on the original label number, update X and Y, and retrain the model according to step (8) to increase the model's fault diagnosis capability.

[0027] Compared with the prior art, the beneficial effects of this invention are as follows:

[0028] Currently, large generator stators monitor operating temperatures through distributed temperature measurement points, including winding temperature measurement, core temperature measurement, pressure finger plate temperature measurement, and cooling air temperature. Anomalies are detected through measurement point inspections and threshold values, and anomaly identification is based on range, average temperature rise, and maximum temperature. However, as generator capacity increases and size grows, the spatial structure causes temperature deviations at different measurement points exceeding 3°C. Early localized faults such as poor contact at winding electrical joints, short circuits between strands of winding bars, and short circuits between turns of the core can lead to even lower localized temperature rises at the detection points due to solid insulation and core conduction, potentially lower than the temperature difference caused by the stator structure itself. Therefore, existing methods struggle to detect these early faults.

[0029] This invention constructs a graph model based on a thorough consideration of the large generator structure, cooling method, temperature measurement point layout, and relationships between measurement points. It then constructs node feature vectors within the graph model structure, including operating condition information, operating condition change information, node temperature information, and node temperature change information. A graph feature matrix is ​​constructed using these feature vectors, containing temperature distribution information between nodes in space. Finally, a graph feature matrix dataset is constructed using equally spaced samples, containing the temporal characteristics of temperature changes with load. By constructing the graph model using this invention's graph modeling method and employing a graph neural network algorithm, the model can fully learn the temperature change characteristics under fault conditions in the stator structure. It can distinguish the differences in the impact of different operating conditions on the temperature of different types of measurement points under normal conditions, the temperature differences at measurement points caused by spatial structure, and the temperature differences at measurement points caused by temperature changes lagging behind load changes. Using this invention's modeling method and training the model with a graph neural network for stator fault diagnosis improves the sensitivity of analyzing and diagnosing stator thermal faults. Attached Figure Description

[0030] Figure 1 The diagram shows the arrangement of temperature measurement on the stator of the cooler unit; where a) is a circumferential cross-sectional view of the stator, b) is a cross-sectional view of K1, and c) is a cross-sectional view of K2.

[0031] Figure 2 The structure diagram of the stator temperature measurement analysis diagram for the cooler unit;

[0032] Figure 3 This is a diagram of the structure of the application example;

[0033] Figure 4 This is a graph showing the operating conditions in an application example.

[0034] Figure 5 The graph shows the loss result of the graph neural network in an application example. Detailed Implementation

[0035] The present invention will now be described in further detail with reference to the embodiments.

[0036] Those skilled in the art will understand that the following embodiments are for illustrative purposes only and should not be construed as limiting the scope of the invention. Where specific techniques or conditions are not specified in the embodiments, they are performed in accordance with the techniques or conditions described in the literature in the field or according to the product instructions. Materials or equipment whose manufacturers are not specified are all conventional products that can be obtained by purchase.

[0037] 1. Divide the stator into zones based on the cooler, and mark the temperature measurement points for each component within each zone, including the windings, core, pressure fingers, and cooler. For example... Figure 1 Stator temperature measurement layout diagram for the cooler unit area. Figure 1 The measurements from ① to ⑩ are, in order: upper stator winding temperature measurement, middle stator winding temperature measurement, lower stator winding temperature measurement, upper core tooth pressure plate pressure finger temperature measurement, lower core tooth pressure plate pressure finger temperature measurement, lower section of the core back temperature measurement, middle section of the core back temperature measurement, upper section of the core back temperature measurement, cooler hot air temperature measurement, and cooler cold air temperature measurement. The arrows in the diagram indicate the cooling air circulation path of the closed-loop self-circulating air cooling system. The cold air cooled by the cooler enters the air gap through the upper and lower ends of the stator, simultaneously cooling the stator end pressure plates pressure fingers and end windings. The cooling air then enters the stator core ventilation groove through the air gap, carrying away the temperature of the windings and core. After exiting the ventilation groove, it re-enters the cooler for heat dissipation and cooling, becoming cold air again to begin the next cycle.

[0038] 2. Using the cooler as a unit, perform graphical modeling of the measuring points within the area.

[0039] 1) Based on Figure 1 A graphical model is established to show the relationship between the unit structure measurement point layout and stator loss heat generation, cooling heat dissipation, and cooling air circulation, as follows: Figure 2 .

[0040] Figure 2There are 10 nodes in total. Nodes 1, 2, and 3 are the upper, middle, and lower stator winding temperature measurement nodes, respectively. Nodes 4 and 5 are the lower and upper stator core pressure plate pressure finger temperature measurement nodes, respectively. Nodes 6, 7, and 8 are the lower, middle, and upper stator core temperature measurement nodes, respectively. Nodes 9 and 10 are the hot and cold air temperature measurement nodes of the cooler, respectively.

[0041] Connections between nodes are established using single-phase arrows. The number of nodes and connection edges can be increased based on the actual number of measuring points on the equipment. Nodes with constant values ​​or measuring points that have no impact on temperature analysis are not introduced, and connections are not established between nodes that have no direct impact. Node 1 outputs to nodes 8 and 9 via a one-way arrow; Node 2 outputs to nodes 7 and 9 via a one-way arrow; Node 3 outputs to nodes 6 and 9 via a one-way arrow, indicating that the winding temperature directly affects the temperature of the nearby core and the hot air temperature of the cooler; Node 4 outputs to node 9 via a one-way arrow; Node 5 outputs to node 9 via a one-way arrow, indicating that the temperature of the core pressure plate finger directly affects the hot air temperature of the cooler; Node 6 outputs to node 9 via a one-way arrow; Node 7 outputs to node 9 via a one-way arrow; Node 8 outputs to node 9 via a one-way arrow, indicating that the winding temperature directly affects the hot air temperature of the auxiliary cooler; Node 9 outputs to node 10 via a one-way arrow, indicating that the hot air temperature of the cooler directly affects the cold air temperature; Node 10 outputs to nodes 1, 2, 3, 4, 5, 6, 7, and 8 via a one-way arrow, indicating that the cold air temperature of the cooler directly affects the temperature of the stator winding, core, pressure plate finger, and hot air temperature of the cooler.

[0042] 2) Construct the connection relationships between nodes from the graph structure graph, represented by a two-dimensional matrix M, where M is the element of the matrix. ij This indicates whether there is a connection between nodes i and j, i.e., a directed edge from node i to j. If it exists, then M... ij =1, if not, then M ij = 0, n represents the number of nodes in the graph structure; i = 0, 1, 2, 3....n-1, j = 0, 1, 2, 3....n-1, M is the element in the matrix. ij The value of 0 or 1 is determined based on the interaction relationships between pairs of nodes in the specific physical model. Matrix M is represented as follows:

[0043]

[0044] Figure 2 The connection matrix M of the Chinese diagram structure is as follows:

[0045]

[0046] 3) Construct node feature vectors using a one-dimensional array x i This indicates that the feature matrix of a graph model composed of multiple nodes is represented by a two-dimensional array X. Tt The time series graph feature tensor, composed of feature matrices from multiple sample graphs over a continuous period, is represented by a three-dimensional array X; xi X Tt X represents the characteristic information of each temperature measuring point at different locations and different functional components, the temperature distribution characteristics of each measuring point, and the temporal characteristics of the temperature distribution as the operating conditions change, respectively.

[0047] x i =[P0, P1, Q0, Q1, T0, T1]

[0048] x i Let x be the feature vector of the i-th node, where i = 0, 1, 2, 3, ..., n-1, and n represents the number of nodes in the graph structure. i In the diagram, P represents the active power of generator operating condition parameters, P0 represents the active power of generator at the previous moment, and P1 represents the active power of generator at the current moment; Q represents the reactive power of generator operating condition parameters, Q0 represents the reactive power of generator at the previous moment, and Q1 represents the reactive power of generator at the current moment; T represents the node temperature, T0 represents the temperature of node i at the previous moment, and T1 represents the temperature of node i at the current moment; the time interval between the current moment and the previous moment is set to 1 minute, and the time interval can be adjusted for different sizes of generator units.

[0049] X Tt The feature matrix of the graph model consists of n node feature vectors x0, x1, x2, ..., xn. n-3 ,x n-2 ,x n-1 constitute.

[0050] X Tt =[x0,x1,x2,...x n-3 ,x n-2 ,x n-1 ]

[0051] n is the number of nodes, 1, 2, 3...n represents the node index, x0, x1, x2,...x n-3 ,x n-2 ,x n-1 X represents the n feature vectors from node 1 to node n, in that order. Tt The feature matrix of a graph model of n nodes at a certain time t

[0052] The X-series graphical model feature tensor consists of t graphical model feature matrices X. Tt The sequence is composed of X, where t is the time length of the sequence t = 0, 1, 2, ..., t-1. T0 The feature matrix of the graphical model at time 0, X T1 The feature matrix of the time-1 graphical model is X. Tt-1 Let X be the feature matrix of the time series graphical model at time t-1. Then, the feature tensor X of the time series graphical model at time length t is as follows:

[0053] X = [X] T0 ,XT1 ,X T2 ,.....X Tt-3 ,X Tt-2 ,X Tt-1 ]

[0054] 3. Following the above graphical model requirements, using historical data from generators of the same model or with identical stator structure measurement point layout and cooling methods, construct time-series graphical model feature tensor three-dimensional array X and fault label one-dimensional array Y for different faults, i.e., the fault sample dataset. Based on historical data and fault information, refer to Table 1 to construct one or more fault samples.

[0055] Table 1

[0056]

[0057]

[0058] Note: The number of elements in Y0 to Y5 is the same as the number of samples.

[0059] 4. Based on the above graph model and fault dataset, a graph neural network is used to learn fault features (i.e., the graph neural network is trained and validated using a two-dimensional matrix M and the time series graph feature matrix X corresponding to the historical data of each state as input and the corresponding fault label Y as output). The trained graph neural network model (i.e., the hydro-generator stator thermal fault diagnosis model) and parameters are used to diagnose and determine the fault in the generator sample data with unknown faults.

[0060] Application Examples

[0061] Taking generator No. 8 of a factory with a capacity of 650MW and 14 coolers as an example, this paper conducts graphical modeling and analysis diagnosis. The stator is divided into 14 stator units, centered on the coolers. The example will use cooler unit No. 14 as an example for graphical modeling, analysis, and diagnosis. The modeling, analysis, and diagnosis methods for the remaining cooler units are the same as those for cooler unit No. 14 in the example, and will not be described again in the example.

[0062] 1. Based on the stator temperature measurement layout diagram of the generator cooler unit, count the temperature measurement points of each component in the 14th cooler unit area, including the winding, core, pressure finger, and cooler, and label the measurement point serial number starting from zero; count the slot number and segment number of each measurement point (from left to right and from top to bottom), and determine the spatial position of the measurement point in the stator, i.e., upper, middle, and lower, based on the slot number and segment number. List the category and spatial position of the measurement points according to the measurement point name and spatial position, as shown in Table 2, to establish the diagram structure.

[0063] Table 2

[0064]

[0065]

[0066] 2. Perform graphical modeling of stator cooler unit 14.

[0067] 1) Based on the relationship between stator loss heat generation, cooling heat dissipation, and cooling air circulation, establish the graph structure as shown in the figure below. Based on the stator structure, heat generation and cooling mechanism, and temperature measurement point arrangement, set the graph node n. In the example, as shown in Table 2, this unit has 18 measurement points, so n = 18. Based on the generation and transfer of heat, set the directed connections (edges) between each node. In the example, there are 44 edges.

[0068] 2) Construct the graph model edge relation matrix. In the example graph structure, the model nodes n=18. Based on the stator thermal failure mechanism and ventilation and heat dissipation principle, 44 directed edges are determined. The edge relation matrix M obtained according to the example graph structure is as follows:

[0069]

[0070] M ij This indicates whether there is a connection between nodes i and j, i.e., a directed edge from node i to j. If it exists, then M... ij =1, if not, then M ij =0, number of elements 18*18, i = 0, 1, 2, 3....17, j = 0, 1, 2, 3....17, element M in the matrix ij The value of 0 or 1 is determined based on the interaction between nodes in the specific physical model. The physical meaning is that the coil temperature will directly affect the hot air temperature of the cooler. Similarly, there are a total of 44 directed edges, and each edge has its own physical meaning.

[0071] 4) Construct node feature vectors and graph feature matrices.

[0072] Based on the information of cooler unit node 14, two sample data points were obtained for the generator at 15:22 and 15:23 on May 22, 2022, as shown in Table 3 below:

[0073] Table 3

[0074]

[0075] The one-dimensional array x of node feature vectors for constructing the graph model of cooler unit No. 14 i A 1×6 array of feature matrices is constructed from the given data. Tt Items of size 18 x 6:

[0076] At 15:23, the graph structure contains 18 nodes and 18 feature vectors in a one-dimensional array x0 to x... 17 In order:

[0077] x0=[568.6,567.9084,58.45078,57.77768,64.3,64.3]

[0078] x1=[568.6,567.9084,58.45078,57.77768,62.7,62.7]

[0079] x2=[568.6,567.9084,58.45078,57.77768,61.7,61.7]

[0080] x3=[568.6,567.9084,58.45078,57.77768,49.3,49.3]

[0081] x4=[568.6,567.9084,58.45078,57.77768,54.2,54.2]

[0082] x5=[568.6,567.9084,58.45078,57.77768,66.1,66.1]

[0083] x6=[568.6,567.9084,58.45078,57.77768,64.7,64.7]

[0084] x7=[568.6,567.9084,58.45078,57.77768,64.0,64.0]

[0085] x8=[568.6,567.9084,58.45078,57.77768,27.2,27.2]

[0086] x9=[568.6,567.9084,58.45078,57.77768,52.5,52.5]

[0087] x 10 =[568.6,567.9084,58.45078,57.77768,61.9,61.9]

[0088] x 11 =[568.6,567.9084,58.45078,57.77768,62.7,62.7]

[0089] x 12 =[568.6,567.9084,58.45078,57.77768,60.6,60.6]

[0090] x 13=[568.6,567.9084,58.45078,57.77768,61.5,61.5]

[0091] x 14 =[568.6,567.9084,58.45078,57.77768,61.9,61.9]

[0092] x 15 =[568.6,567.9084,58.45078,57.77768,60.9,60.9]

[0093] x 16 =[568.6,567.9084,58.45078,57.77768,63.3,63.3]

[0094] x 17 =[568.6,567.9084,58.45078,57.77768,60.0,60.0]

[0095] The graph feature matrix is ​​a two-dimensional array X. T0 :

[0096] X T0=[[568.6,567.9084,58.45078,57.77768,64.3,64.3],[568.6,567.9084,58.45078,57.77768,62.7,62.7],[568.6,567.9084,58.45078,57.77768,61.7,61.7],[568.6,567.9084,58.45078,57.77768,49.3,49.3],[568.6,567.9084,58.45 078,57.77768,54.2,54.2],[568.6,567.9084,58.45078,57.77768,66.1,66.1],[568.6,567.9084,58.45078,57.77768,64.7,64.7],[568.6,567.9084,58.45078,57.77768,64.0,64.0],[568.6,567.9084,58.45078,57.77768,27.2,27.2] ,[568.6,567.9084,58.45078,57.77768,52.5,52.5],[568.6,567.9084,58.45078,57.77768,61.9,61.9],[568.6,567.9084,58.45078,57.77768,62.7,62.7],[568.6,567.9084,58.45078,57.77768,60.6,60.6],[568.6,567.9084,58.450 78,57.77768,61.5,61.5],[568.6,567.9084,58.45078,57.77768,61.9,61.9],[568.6,567.9084,58.45078,57.77768,60.9,60.9],[568.6,567.9084,58.45078,57.77768,63.3,63.3],[568.6,567.9084,58.45078,57.77768,60.0,60.0]]

[0097] exist Figure 3 The generator in the structure was selected from 15:23 to 16:22 on May 22, 2022. One set of samples was obtained every minute for one hour, and a feature matrix X consisting of 60 consecutive sequence graph models was constructed. T0 -X T59 The three-dimensional array X of the characteristic tensor of the timing diagram model for cooler unit 14, with a size of 60×18×6, is as follows:

[0098] X＝[[[547.2239,547.8749,52.41906,52.59641,64.2,64.2],[547.2239,547.8749,52.41906,52.59641,62.8,62.8],[547.2239,547.8749,52.41906,52.59641,61.8,61.8],[547.2239,547.8749,52.41906,52.59641,49.2,49.2],[547.2239,547.8749,52.41906,52.59641,54.4,54.4],[547.2239,547.8749,52.41906,52.59641,66.4,66.3],[547.2239,547.8749,52.41906,52.59641,64.5,64.5],[547.2239,547.8749,52.41906,52.59641,64.1,64.1],[547.2239,547.8749,52.41906,52.59641,27.6,27.6],[547.2239,547.8749,52.41906,52.59641,52.7,52.7],[547.2239,547.8749,52.41906,52.59641,62.1,62.1],[547.2239,547.8749,52.41906,52.59641,63.3,63.3],[547.2239,547.8749,52.41906,52.59641,60.8,60.8],[547.2239,547.8749,52.41906,52.59641,61.5,61.5],[547.2239,547.8749,52.41906,52.59641,61.9,61.9],[547.2239,547.8749,52.41906,52.59641,61.0,61.0],[547.2239,547.8749,52.41906,52.59641,63.5,63.5],[547.2239,547.8749,52.41906,52.59641,60.3,60.3]],

[0099]

[0100] [[443.0389,444.6879,36.3532,34.67438,59.41,59.32],[443.0389,444.6879,36.3532,34.67438,57.99,57.89],[443.0389,444.6879,36.3532,34.67438,57.32,57.24],[443.0389,444.6879,36.3532,34.67438,47.46,47.04],[443.0389,444.6879,36.35 32,34.67438,52.00,52.00],[443.0389,444.6879,36.3532,34.67438,60.71,60.60],[443.0389,444.6879,36.3532,34.67438,59.62,59.52],[443.0389,444.6879,36.3532,34.67438,59.12,59.03],[443.0389,444.6879,36.3532,34.67438,27.00,27.00], [443.0389,444.6879,36.3532,34.67438,50.90,50.90],[443.0389,444.6879,36.3532,34.67438,60.30,60.30],[443.0389,444.6879,36.3532,34.67438,61.50,61.50],[443.0389,444.6879,36.3532,34.67438,59.60,59.60],[443.0389,444.6879,36.353 2,34.67438,57.09,57.02],[443.0389,444.6879,36.3532,34.67438,57.33,57.25],[443.0389,444.6879,36.3532,34.67438,56.53,56.44],[443.0389,444.6879,36.3532,34.67438,58.61,58.53],[443.0389,444.6879,36.3532,34.67438,56.02,55.92]]]

[0101] 3. Construction of the fault dataset.

[0102] Different fault sample sequences are obtained at equal time intervals. The length includes all operating conditions under normal operation, including load increase, load decrease, stable load, and different combinations of active and reactive power. The fault label for normal state is 0, the fault label for short circuit between core laminations is 1, the fault label for short circuit between winding strands is 2, the fault label for reduced cooler efficiency is 3, the fault label for abnormal measuring point is 4, and the fault label for other abnormal temperature is 5.

[0103] Select working conditions such as Figure 4 As shown. Figure 4 In the table, the horizontal axis represents the sample number, and the vertical axis represents the amplitudes of P and Q. The sample includes 4727 samples from 3 start-stop cycles. A partial sample table of core faults is shown in Table 3. A partial sample table of normal samples is shown in Table 4.

[0104] Table 3. Sample Table of Core Faults (Partial)

[0105]

[0106]

[0107] Table 4. Partial List of Normal Samples

[0108]

[0109]

[0110] 4. Graph models are used for stator temperature fault diagnosis in the No. 14 cooler area. Time-series feature datasets of different faults are constructed using different fault samples to train a graph neural network model. The trained graph neural network model is then used to analyze and diagnose the generator.

[0111] 1) Structure and parameters of the Graph Neural Network (GCN) algorithm

[0112] Convolutional layer 1, input feature dimension 12, output feature dimension 128; (conv1):GCNConv(12,128);

[0113] Graph downsampling, input feature dimension 128, retain the number of nodes 80% of the input node count, output feature dimension remains unchanged; (pool1):TopKPooling(128, ratio = 0.8, multiplier = 1.0);

[0114] Graph convolutional layer 2, input feature dimension 128, output feature dimension 128; (conv2):GCNConv(128,128);

[0115] Graph downsampling, input feature dimension 128, retaining 80% of the input node count, output feature dimension unchanged; (pool2):TopKPooling(128, ratio = 0.8, multiplier = 1.0);

[0116] The graph shows convolutional layer 3 with an input feature dimension of 128 and an output feature dimension of 128; (conv3):GCNConv(128,128);

[0117] Graph downsampling, input feature dimension 128, retaining 80% of the input node count, output feature dimension unchanged; (pool3):TopKPooling(128, ratio = 0.8, multiplier = 1.0);

[0118] Fully connected layer 1, input feature dimension 128, output feature dimension 128; (lin1):Linear(in_features=128,out_features=128,bias=True);

[0119] Fully connected layer 2, input feature dimension 128, output feature dimension 64; (lin2):Linear(in_features = 128, out_features = 64, bias = True);

[0120] Fully connected layer 3, input feature dimension 64, output feature dimension 1; (lin3):Linear(in_features=64,out_features=1,bias=True);

[0121] Batch normalization layer 1, performs normalization operation on 128-dimensional features; (bn1):BatchNorm1d(128,eps=1e-05,momentum=0.1,affine=True,track_running_stats=True);

[0122] Batch normalization layer 2 performs normalization operation on 64-dimensional features; (bn2):BatchNorm1d(64,eps=1e-05,momentum=0.1,affine=True,track_running_stats=True);

[0123] Activation function 1: ReLU; (act1): ReLU();

[0124] Activation function 2: ReLU; (act2): ReLU().

[0125] 2) Model Training

[0126] The optimizer, torch.optim.Adam, optimizes the model's parameters.

[0127] Learning rate: lr is set to 0.001, which is the step size the optimizer uses each time it updates the parameters;

[0128] Loss functions: torch.nn.BCELoss binary cross-entropy loss function, torch.nn.CrossEntropyLoss loss function for multi-class classification;

[0129] Training steps: forward propagation, loss calculation, backpropagation, and parameter update;

[0130] Training parameters: Epoch = 20, batch_size = 600;

[0131] The training results are shown in the figure below. After 20 iterations, the loss decreased to 0.3. Figure 5 ;

[0132] Training result evaluation score: roc_auc_score: 0.9227018940062575.

[0133] 3) Diagnostic analysis application. Samples 1, 2, and 3 in Table 5 were normal, while samples 4, 5, and 6 had localized thermal faults in their cores (stator core temperature 53 was 2°C higher than normal, corresponding to a 1°C higher hot air temperature and a 0.5°C higher cold air temperature in the cooler). The data was input into the model for classification and diagnosis, and the accuracy of the output results was 0.9227.

[0134] Note: The fault label for normal conditions is 0, the fault label for short circuit between core laminations is 1, the fault label for short circuit between winding strands is 2, the fault label for reduced cooler efficiency is 3, the fault label for abnormal measuring points is 4, and the fault label for other abnormal temperatures is 5.

[0135] The samples to be diagnosed are shown in Table 5 below.

[0136] Table 5

[0137]

[0138]

[0139] Diagnostic output results:

[0140] Six samples were input into the trained diagnostic model for fault classification. The output classification prediction results showed fault label values ​​of 0.05099515, 0.06590305, 0.17380856, 0.9895475, 0.9888508, and 0.98820704, respectively. The true label values ​​were 0, 0, 0, 1, 1, and 1, respectively, with an accuracy of 0.9227.

[0141] Results Explanation:

[0142] Sample 1 is a normal sample with a label of 0 and a predicted label of 0.05099515, indicating a 94.9% probability that it is a normal sample.

[0143] Sample 2 is a normal sample with a label of 0 and a predicted label of 0.05099515, indicating a 93.4% probability that it is a normal sample.

[0144] Sample 3 is a normal sample with a label of 0 and a predicted label of 0.05099515, indicating an 82.6% probability that it is a normal sample.

[0145] Sample 4, with a core failure, has a label of 1 and a predicted label of 0.9895475, indicating a 98.9% probability that it is a core failure sample.

[0146] Sample 5, with a core failure, has a label of 1 and a predicted label of 0.9888508, indicating a 98.8% probability that it is a core failure sample.

[0147] Sample 6, with a core fault, has a label of 1 and a predicted label of 0.98820704, indicating a 98.8% probability that it is a core fault sample.

[0148] In summary, the time-series and spatial distribution-based stator thermal fault diagnosis method for hydro-generators provided in this embodiment of the invention has the following technical effects or advantages compared with existing technologies such as measurement point range variation analysis, fixed threshold, dynamic threshold, and multivariate evaluation:

[0149] 1. This invention constructs a graph model based on the structure, cooling method, temperature measurement point layout, and relationships between measurement points of a large generator; it links the physical model between generator stator losses, heat generation, ventilation, and cooling through the relationships between nodes and edges, resulting in good interpretability of the diagnostic analysis results;

[0150] 2. This invention combines a graph model structure to construct node feature vectors, which include operating condition information, operating condition change information, measuring point temperature information, and measuring point temperature change information; thus avoiding the false alarms and lack of alarms caused by the inability to distinguish operating condition differences in traditional threshold methods. For example, when the load drops rapidly to a light load, the low load and high temperature will cause false alarms, and the abnormality of high load and low temperature during the rapid load increase process after startup cannot be detected and no alarm is triggered.

[0151] 3. This invention constructs a graph feature matrix using feature vectors, containing temperature distribution information between nodes in space. Existing methods suffer from low detection sensitivity due to temperature distribution differences caused by the stator structure, resulting in excessively high thresholds to avoid false alarms. For example, if the temperature measurements at the upper and lower cores differ by 1-5℃, the threshold should be greater than 5℃ using the range method. However, this approach fails to detect temperature rise faults below 5K. This invention, through model learning of multi-point temperature distribution characteristics, can detect abnormal states based on normal distribution, exhibiting high sensitivity. In the example, it diagnosed a localized thermal fault in the core (stator core No. 53 temperature was 2℃ higher, corresponding to a 1℃ higher hot air temperature and a 0.5℃ higher cold air temperature in the cooler).

[0152] 4. This invention constructs a graphical feature matrix dataset using equally spaced samples, containing the temporal characteristics of temperature variation with load. In addition to the advantages mentioned above, this method of constructing feature data can reflect the distribution characteristics of each measuring point over time and with load, resulting in high reliability for diagnostic analysis.

[0153] 5. This invention constructs a graph model using a graph modeling method and employs a graph neural network algorithm to fully learn the temperature change characteristics of the stator structure under fault conditions. It can distinguish the differences in the impact of different operating conditions on the temperature of different types of measuring points under normal conditions, the temperature differences at measuring points caused by spatial structure, and the temperature differences at measuring points caused by temperature changes lagging behind load changes. Most existing algorithms suffer from poor interpretability due to low correlation between the data model and the physical model of generator stator ventilation and heat dissipation; simulation calculation methods have good physical model conformity, but the boundary conditions set differ from those of actual operating generators, leading to deviations in calculation results. The method of this invention is not only correlated with the stator heat loss and ventilation and heat dissipation physical model, but also learns actual generator parameters through operating generator data. This improves the sensitivity and reliability of stator thermal fault diagnosis.

[0154] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing stator thermal faults in hydro-generators based on temporal and spatial distribution, characterized in that, The steps include the following: Step (1), Divide the area: Divide the stator into areas based on the cooler, and indicate the temperature measurement points of each component in the area, including the winding, core, pressure finger, and cooler. Step (2), establish graph structure: take the temperature measurement points obtained in step (1) as nodes, establish graph structure according to the relationship between stator loss heat generation, cooling heat dissipation and cooling air circulation, and set directed connections between nodes according to the generation and transfer of heat when establishing graph structure. Step (3), construct the graph model edge relationship matrix M: the connection relationship between each node is represented by a two-dimensional matrix M, and the elements in the matrix M ij This indicates whether there is a connection between nodes i and j, i.e., a directed edge from node i to j. If it exists, then M... ij =1, if not, then M ij =0, n represents the number of nodes in the graph structure, i = 0, 1, 2, 3....n-1, j = 0, 1, 2, 3....n-1, and M is the element in the matrix. ij The value of 0 or 1 is determined based on the interaction relationship between pairs of nodes in the specific physical model; Step (4), construct the node feature vector using a one-dimensional array x i express: x i =[P0,P1,Q0,Q1,T0,T1] Wherein, P0 is the generator active power at the previous moment of the node, P1 is the generator active power at the current moment of the node; Q0 is the generator reactive power at the previous moment of the node, Q1 is the generator reactive power at the current moment of the node; T0 is the temperature at the previous moment of the node, T1 is the temperature at the current moment of the node. Step (5): Construct a two-dimensional array X of the graph model feature matrix. Tt : X Tt =[x0,x1,x2,.....x n-3 , x n-2 ,x n-1 ] n is the number of nodes, 1, 2, 3...n represents the node index, x0, x1, x2,...x n-3 ,x n-2 ,x n-1 X represents the n feature vectors of nodes 1 to n in sequence; Tt Let be the feature matrix of the graph model of n nodes at a certain time t; Step (6), construct the three-dimensional array X of the time series graph model feature tensors: X=[X T0 ,X T1 ,X T2 ,.....X Tt-3 ,X Tt-2 ,X Tt-1 ] The X-series graphical model feature tensor consists of t graphical model feature matrices X. Tt Composition, X T0 X is the feature matrix of the graphical model at time 0. T1 For example, X is the feature matrix of the time-1 graphical model, and so on. Tt-1 The feature matrix of the graphical model at time t-1; Step (7), Data Acquisition: Collect historical data of the normal state and fault state of each node from historical data, and mark them using fault tags; the historical data includes generator active power, generator reactive power, temperature and corresponding time; The faults mentioned include short circuits between core laminations, short circuits between winding strands, reduced cooler efficiency, abnormal measuring points, and other temperature abnormalities. Step (8), training of graph neural network: with two-dimensional matrix M and time series graph feature matrix X corresponding to each state historical data as input, and corresponding fault label Y as output, the graph neural network is trained and verified to obtain the hydro-generator stator thermal fault diagnosis model. Step (9), Fault diagnosis: Real-time collection of generator active power, generator reactive power, temperature and corresponding time data of each node, and acquisition of the corresponding time sequence feature matrix X, input into the hydro-generator stator thermal fault diagnosis model obtained in step (8), and the final fault diagnosis result is obtained based on the output Y of the hydro-generator stator thermal fault diagnosis model.

2. The method for diagnosing stator thermal faults of a hydro-generator based on time-series and spatial distribution as described in claim 1, characterized in that, In step (2), connections between nodes are established using single-phase arrows; nodes with constant quantities or measuring points that have no effect on temperature analysis are not introduced, and connections are not established between nodes that have no direct impact.

3. The method for diagnosing stator thermal faults of a hydro-generator based on time-series and spatial distribution according to claim 1, characterized in that, In step (4), the time interval between the current time and the previous time is set to 1 minute.

4. The method for diagnosing stator thermal faults of a hydro-generator based on time-series and spatial distribution as described in claim 1, characterized in that, The fault label for normal conditions is 0, the fault label for short circuit between core laminations is 1, the fault label for short circuit between winding strands is 2, the fault label for reduced cooler efficiency is 3, the fault label for abnormal measuring points is 4, and the fault label for other abnormal temperatures is 5.

Images