Wind turbine generator set fault early warning method and device based on dynamic network markers

Abstract

The present invention discloses a wind turbine generator set fault early warning method and device based on dynamic network markers, which relate to the field of data processing technology. The method comprises the following steps: collecting data of continuous monitoring variables covered in the wind turbine generator set's own SCADA system, selecting data at K consecutive sampling moments as experimental samples and performing preprocessing; constructing a time window data matrix using the preprocessed experimental samples; calculating the Pearson correlation of variable pairs, thereby constructing a differential network of the current time window; evaluating the score of each node in the differential network of the current time window; mapping the scores of the nodes corresponding to all variables in all time windows to a three-dimensional coordinate graph to realize a three-dimensional visualization graph; selecting N nodes with the highest scores, defining the accumulated value of the node scores as a comprehensive state index (ICI) of the wind turbine generator set, and realizing early fault early warning of the wind turbine generator set according to the dynamic change of the ICI value.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method and device for early warning of wind turbine generator set faults based on dynamic network markers. Background Art

[0002] Wind farms are often located in mountainous or offshore areas with abundant wind resources but harsh natural environments. These complex operating conditions result in significantly higher wind turbine failure rates than other types of power generation equipment. Furthermore, key wind turbine components are often installed at high altitudes, and equipment operation and maintenance is limited by operating windows and meteorological conditions, making operation and maintenance difficult and costly. With the continued growth of my country's wind power installed capacity and the gradual approaching of the design lifespan of some early-operating units, the demand for operation and maintenance in the wind power sector and the associated market size are also expanding. Therefore, establishing an efficient and reliable wind turbine early fault warning technology system has significant engineering application value and practical significance for improving equipment operational safety, reducing maintenance costs, and ensuring the economic operation of wind power. Currently, the vast majority of wind farms have deployed Supervisory Control and Data Acquisition (SCADA) systems, providing rich data support for unit operational status monitoring and fault analysis. Wind turbine early warning and fault diagnosis methods based on SCADA data have attracted widespread attention and in-depth research from scholars both domestically and internationally, and have achieved promising results in engineering practice. Existing wind turbine early warning and fault diagnosis methods based on SCADA data can be primarily categorized into two types: supervised learning-based classification models and Normal Behavior Model (NBM)-based methods. Classification models build an automatic classifier by learning and modeling labeled normal and abnormal data samples, thereby classifying typical faults. For example, some researchers have used the XGboost classifier to classify fault states. To address the scarcity of wind turbine fault samples and imbalanced classification, research has introduced improved generative adversarial networks (GANs) that generate high-quality fault samples through data augmentation techniques, alleviating the sample shortage issue and providing support for the effective construction of classification models. Compared to classification models that rely on a large number of fault samples, NBM methods are more effective in addressing data imbalance. This method typically utilizes a large amount of historical SCADA data from wind turbines in normal operation to construct a health status model. By comparing the residuals between real-time monitoring data and the model's predicted values, it determines whether the current state deviates, thereby enabling anomaly detection and fault warning. Since it does not rely on a large number of fault samples, the NBM method shows strong adaptability and robustness in early fault warning and hidden danger identification, and has gradually become an important research direction in the field of wind power condition monitoring. Although the existing wind turbine SCADA data-driven methods have made some progress in early warning and fault diagnosis, they still have the following limitations: (1) They rely on a large amount of historical data.Most of the current mainstream early fault warning methods rely on a large amount of normal historical data, combined with machine learning models for training and prediction. Their warning effects are not only limited by the accuracy of the model itself, but are also easily affected by the deviation of the prediction residual threshold setting, making it difficult to ensure robustness and accuracy under complex working conditions. (2) Insufficient modeling of individual heterogeneity. Wind turbines have significant differences in design, manufacturing and operating conditions, resulting in different degradation paths for each unit. However, existing methods mostly rely on group data modeling and tend to extract universal state evolution patterns. Although they can identify common fault types, they are difficult to reflect the characteristic evolution laws of a single unit. (3) High computing resources and costs. Machine learning algorithms require a lot of computing resources and time in the model learning and training stages. Especially when facing high-dimensional, multi-variable SCADA data, the training process is complex and costly, which restricts the widespread application of such methods in resource-constrained scenarios.

[0003] In recent years, the theory of "critical phase transitions" in nonlinear system dynamics has gained increasing attention. When a complex system evolves to a critical state, even small changes in control parameters or system stress can trigger a critical phase transition, potentially leading to catastrophic accidents. Theoretical studies of critical phase transitions have found that before a critical phase transition occurs, the system exhibits critical slowing behavior near the critical point, generating three potential early warning signals: slow perturbation recovery, reduced autocorrelation, and increased variance. To describe the dynamic characteristics of critical phase transitions in multivariable complex systems, researchers have proposed the concept of dynamic network markers (DNMs) and demonstrated their effectiveness in biological, ecological, and financial systems. Chinese invention patent application publication number CN107291991A discloses a method for early warning of wind turbine defects based on dynamic network markers. However, this method relies on a predictive model to model and predict each monitored variable individually. The results are largely limited by the accuracy and generalization ability of the selected model, which in turn affects the reliability of the overall diagnostic results and introduces a certain degree of uncertainty. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems in the prior art.

[0005] The technical solution adopted by the present invention to solve the technical problem is to provide a wind turbine generator set fault early warning method and device based on dynamic network markers, including the following steps:

[0006] Collect the data of continuous monitoring variables covered by the wind turbine's own SCADA system, select The data of consecutive sampling moments are used as experimental samples and preprocessed;

[0007] Set the length and step size parameters of the sliding window and use the preprocessed experimental samples to construct the time window data matrix;

[0008] Calculate the Pearson correlation of the variable pair in the current time window and the previous time window, and construct a differential network for the current time window based on the Pearson correlation of the variable pair; construct a differential network for all time windows;

[0009] Evaluate the variability and correlation characteristics of each node in the differential network of the current time window, and calculate the score of each node in the differential network of the current time window based on the variability and correlation characteristics; calculate the scores of the nodes in the differential network of all time windows;

[0010] Map the scores of the nodes corresponding to all variables in all time windows to a three-dimensional coordinate graph, and draw a three-dimensional visualization of the entire process;

[0011] Select the highest score nodes, define the The accumulated score of each node is used as the comprehensive status index ICI of the wind turbine, and the early fault warning of the wind turbine is realized according to the dynamic change of the ICI value.

[0012] Preferably, the data of continuous monitoring variables covered by the SCADA system of the wind turbine generator set itself is collected, and the data is selected The data of consecutive sampling moments are used as experimental samples and preprocessed. The preprocessing refers to normalizing the data, which is expressed as:

[0013] ;

[0014] in, is the collected variable value, For variable values Normalized data of is the average value of each variable collected; is the standard deviation of the collected variable values.

[0015] Preferably, the setting of the sliding window length and step size parameters and the construction of a time window data matrix using the preprocessed experimental samples include the following steps:

[0016] Set the length of the sliding window to , the step size is , get the number of variables as ;Continuous acquisition variables in The variable value at each moment is used to obtain the time window data matrix , expressed as:

[0017] ;

[0018] in, Indicates the The variable in the current time window variable values;

[0019] Repeat the above steps times, obtained A time window data matrix.

[0020] Preferably, the step of calculating the Pearson correlation of the variable pair in the current time window and the previous time window, and constructing the differential network of the current time window according to the Pearson correlation of the variable pair, comprises the following steps:

[0021] Calculate the Pearson correlation between any pair of variables in the previous and current time windows, expressed as:

[0022] ;

[0023] in, Indicates the Variables in a time window and variables Pearson correlation of ; Representing variables In the The time window variable values, Representing variables In the The time window variable values; Indicates the variables within a time window The mean of the variable values, Indicates the variables within a time window The mean of the variable values;

[0024] For the same pair of variables, the specific Pearson correlation coefficient is calculated based on the Pearson correlation of the previous time window and the current time window, which is expressed as:

[0025] ;

[0026] in, Representing variables and variables In the Specific Pearson correlation coefficient in the time window;

[0027] Taking the variables as nodes, we determine whether there is an edge connection between the two nodes corresponding to the pair of variables based on whether the value of the specific Pearson correlation coefficient reaches the threshold.

[0028] The above operations are performed on all variable pairs to construct the differential network of the current time window.

[0029] Preferably, the calculation process of the node score includes the following steps:

[0030] The differential expression deviation CV of the calculation node is defined as:

[0031] ;

[0032] in, Representation node The first-order field is composed of nodes It is composed of nodes connected by edges. express The nodes in Represents a first-order domain The number of midpoints, Representation node the standard deviation of the values ​​in the current time window, Representation node The average of the values ​​in the current time window, Represents a first-order domain The standard deviation of the node value in the current time window, In the first-order domain The average value of the node in the current time window;

[0033] Calculate the nodes of a node and its first-order neighbors The mean value is defined as:

[0034] ;

[0035] in, Representation node With first-order fields The average value of the specific correlation coefficients of all nodes in the current time window;

[0036] Calculate the relationship between the nodes in the first-order neighborhood of the node and the nodes in the second-order neighborhood of the node average value , expressed as:

[0037] ;

[0038] in, Representation node The second-order neighborhood of The nodes of are composed of nodes connected by edges. Represents the second-order neighborhood The number of nodes, Represents the second-order neighborhood Nodes in

[0039] Based on the above calculation results, the score of each node in the differential network under the current time window is calculated, which is expressed as:

[0040] ;

[0041] in, Representation node score.

[0042] Preferably, mapping the scores of the nodes corresponding to all variables in all time windows into a three-dimensional coordinate graph and drawing a three-dimensional visualization graph of the entire process includes the following steps:

[0043] Construct a three-dimensional coordinate system with the X-axis representing the variable, the Y-axis representing the time window, and the Z-axis representing the node score;

[0044] The calculated scores of the nodes corresponding to each variable in each time window are placed in the three-dimensional coordinate axis to obtain a three-dimensional visualization diagram of the whole process.

[0045] Preferably, the wind turbine integrated status index ICI is expressed as:

[0046] ;

[0047] in, Representing variables The corresponding node score.

[0048] The present invention also provides a wind turbine generator set fault warning device based on dynamic network markers, comprising:

[0049] The sample collection module collects the data of continuous monitoring variables covered by the wind turbine's own SCADA system and selects The data of consecutive sampling moments are used as experimental samples and preprocessed;

[0050] The time window construction module sets the length and step size parameters of the sliding window and uses the preprocessed experimental samples to construct the time window data matrix;

[0051] A differential network construction module calculates the Pearson correlation of the variable pair in the current time window and the previous time window, and constructs the differential network of the current time window based on the Pearson correlation of the variable pair; and constructs the differential network for all time windows;

[0052] The node score calculation module evaluates the variability and correlation characteristics of each node in the differential network of the current time window, and calculates the score of each node in the differential network of the current time window based on the variability and correlation characteristics; the score is calculated for the nodes in the differential network of all time windows;

[0053] The node score visualization module maps the node scores corresponding to all variables in all time windows into a three-dimensional coordinate graph and draws a three-dimensional visualization of the entire process;

[0054] Fault prediction module, select the one with the highest score nodes, define the The accumulated score of each node is used as the comprehensive status index ICI of the wind turbine, and the early fault warning of the wind turbine is realized according to the dynamic change of the ICI value.

[0055] The present invention has the following beneficial effects: Based on the monitoring variables covered by the SCADA (Supervisory Control and Data Acquisition, SCADA) system of the target wind turbine, the present invention maps them into a variable node network. By calculating three characteristic quantitative indicators of dynamic network markers, the dynamic network marker subnetwork in the node network is extracted, and then a comprehensive state indicator of the wind turbine is constructed to identify when the wind turbine has transitioned from a normal state to a critical state. This method does not rely on a large amount of normal historical sample data for learning and training using machine learning methods; it also does not require model building, has strong generalization capabilities, and facilitates early warning of generator defects.

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A diagram showing the steps of a method according to an embodiment of the present invention;

[0058] Figure 2 A flow chart of a method for early warning of wind turbine faults based on three-dimensional visual dynamic network markers provided by an embodiment of the present invention;

[0059] Figure 3 The three-dimensional visualization of the LDS values ​​of each variable in 100 time windows under normal operation;

[0060] Figure 4 This is the ICI value change trend chart for 100 time windows under normal operation;

[0061] Figure 5 This is the LDS value diagram of each variable in two time windows under normal operation;

[0062] Figure 6 The three-dimensional visualization of the LDS values ​​of each variable in 131 time windows of the generator bearing fault;

[0063] Figure 7 This is the ICI value change trend diagram of each time window in the evolution process of the generator bearing fault;

[0064] Figure 8 This is the LDS value diagram of each variable in the time window of each stage during the evolution of the generator bearing fault;

[0065] Figure 9 The three-dimensional visualization of the LDS values ​​of each variable in 131 time windows of hydraulic system failure;

[0066] Figure 10 The ICI value change trend diagram of 131 time windows of hydraulic system failure;

[0067] Figure 11 It is the LDS value diagram of each variable in each time window during the evolution of hydraulic system failure;

[0068] Figure 12 2 is a structural diagram of a device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] See also Figure 1 , which is a method step diagram and a method flow chart of an embodiment of the present invention, includes the following steps:

[0070] S101, collect the data of continuous monitoring variables covered by the wind turbine generator set's own SCADA system, select The data of consecutive sampling moments are used as experimental samples and preprocessed;

[0071] S102, setting the length and step size parameters of the sliding window, and constructing a time window data matrix using the preprocessed experimental samples;

[0072] S103, calculating the Pearson correlation of the variable pair in the current time window and the previous time window, and constructing a differential network for the current time window based on the Pearson correlation of the variable pair; and constructing a differential network for all time windows;

[0073] S104, evaluating the variability and correlation characteristics of each node in the differential network of the current time window, and calculating the score of each node in the differential network of the current time window based on the variability and correlation characteristics; and calculating the scores of the nodes in the differential networks of all time windows;

[0074] S105, mapping the scores of the nodes corresponding to all variables in all time windows into a three-dimensional coordinate graph, and drawing a three-dimensional visualization graph of the entire process;

[0075] S106, select the one with the highest score nodes, define the The accumulated score of each node is used as the comprehensive status index ICI of the wind turbine, and the early fault warning of the wind turbine is realized according to the dynamic change of the ICI value.

[0076] Specifically, in S101, the operation monitoring data collected in real time by the SCADA system needs to be preprocessed. The preprocessing method is to normalize the data:

[0077] ;

[0078] in, is the normalized data of the variable at a certain sampling moment; It is the variable value collected at a certain sampling moment; is the average value of each variable at a certain sampling moment; is the standard deviation of each variable at a certain sampling moment.

[0079] Specifically, the window data matrix is ​​constructed in S102, which specifically includes: setting the length of the sliding window to , the step size is , get the number of variables as , to construct the window sample data matrix . Select The data of consecutive sampling moments are taken as experimental samples.

[0080] ;

[0081] Specifically, in S103, a differential network of wind turbine generators in the current time window is established, and the specific steps include:

[0082] S201, select the first To The SCADA data of each sampling moment is used as the data matrix of the time window (a total of group), where Represents the number of sampling moments contained in the time window (i.e. the above-mentioned window length). , the data of the corresponding sampling time range is used as the reference time window data, and any pair of samples in the time window and The correlation between them is expressed using the Pearson correlation coefficient (PCC) [symbol , variable and The definition of monitoring variables in SCADA is:

[0083] ;

[0084] Where: and are the monitoring variables in the window samples respectively and monitoring variables The value of and Monitor variables and monitoring variables the mean value in the reference sample group;

[0085] S202, as the sampling time increases, , time window , the time window is in steps of Slide forward, and the data samples contained in the window are updated accordingly. Specifically, the variable and After the data in the new time window The value of and The difference between the two is due to the change of the window data, which represents the correlation between the new window data matrix and the previous time window data matrix. From this, we can get the specific Pearson correlation coefficient (SPCC) of the sample and [with the symbol It is defined as:

[0086] ;

[0087] The above formula is a difference equation, which means that when the window data matrix changes, the relationship between the variables is: The disturbance of the value. follows a normal distribution, so Follows the difference normal distribution. Assume that each variable is mapped to a node in the network, The value of and nodes The basis for whether there is an edge between them: for each The correlation test is performed on the values. When the values ​​of a variable in adjacent time windows are If it is higher than 0.05, the variable and There are differential edges between them, that is, there are edge connections; otherwise, they are considered variables and There is no edge between them. So far, the A differential network (DN) with a time window.

[0088] Specifically, in S104, the variability and correlation characteristics of each node in the differential network of the current time window are evaluated, specifically: in order to obtain the dynamic network markers (DNM) sub-network in the DN within the time window, it is necessary to assign a local DNM score (LDS) to each node. The essence of DNM is a group of interacting nodes. When the network state develops to a critical state, they will show three characteristics: ① The average coefficient of variation of the nodes in the sub-network increases; ② The absolute value of the correlation of the nodes in the sub-network increases; ③ The absolute value of the correlation between the nodes in the sub-network and the external variable nodes decreases. Therefore, the node LDS that needs to be constructed is actually to quantify the three characteristics of the above-mentioned DNM.

[0089] Specifically, for the first feature of DNM, using nodes and all nodes in its first-order neighborhood ( Nodes in the network Variables connected by edges form nodes The coefficient of variation (CV) of the first-order neighborhood of It is represented by:

[0090] ;

[0091] Where: Representation node The first-order field is composed of nodes It is composed of nodes connected by edges. express The nodes in Represents a first-order domain The number of midpoints, Representation node the standard deviation of the values ​​in the current time window, Representation node The average of the values ​​in the current time window, Represents a first-order domain The standard deviation of the node value in the current time window, In the first-order domain The average value of the node in the current time window;

[0092] The second feature of DNM is node and the nodes in its first-order neighborhood Average quantized by value (in symbols denoted), defined as:

[0093] ;

[0094] The third feature of DNM is to consider nodes Nodes in the first-order neighborhood of and the nodes of the second-order neighborhood (DN network and node Variables in the first-order neighborhood of Variables connected by edges form nodes The second-order neighborhood of Average value (in symbol denoted), defined as:

[0095] ;

[0096] Where: Represents the specific difference between nodes in the network Nodes in the first-order neighborhood of Nodes consisting of nodes connected by edges The second-order neighborhood of Represents the number of nodes in the second-order neighborhood.

[0097] The calculation of the score of each node in the differential network at the current sampling moment in S104 specifically includes:

[0098] S401, using the above calculated related values ​​to calculate the nodes in the DN at the current sampling time The local DNM score (LDS) [denoted by It is defined as:

[0099] ;

[0100] Can be considered as a node The local score of the current time window is obtained by analogy, and the score of each node in the current time window is obtained by analogy. ;

[0101] S402, determine whether To determine whether the The corresponding nodes of the DN in the time window If not satisfied, return to S202 to obtain the data matrix of the next time window; if satisfied, proceed to S105.

[0102] Specifically, in S105, the scores of each node obtained at all sampling moments are used to draw a three-dimensional coordinate graph, and the specific steps include: Put it into the three-dimensional coordinate axis and get an axis as the variable number. The axis is the time window, Axis A three-dimensional plot of the values.

[0103] Specifically, in said S106, the comprehensive state index of the wind turbine in the current time window is constructed and calculated, specifically according to the calculated state index of each node in different time windows. , and use it as the basis for sorting key nodes. Top value ranking The variable nodes of the DNM sub-network under the current time window are used to characterize the critical state change characteristics of the system. On this basis, all nodes in the DNM sub-network are defined The accumulation of the values ​​is used as the integrated condition index (ICI) to evaluate the critical transition of the wind turbine state. ) to achieve quantitative characterization of the overall operating status of the system and the construction of fault warning criteria, Expressed as:

[0104] .

[0105] A verification experiment was conducted on the embodiment of the present invention, with the wind turbine unit T07 of a certain wind farm as the object. The unit has a horizontal axis three-blade structure, a rated power of 2MW, a rated wind speed of 12m / s, and a cut-in wind speed and a cut-out wind speed of 4m / s and 25m / s respectively. The diameter of the wind rotor is 90m, and its maximum speed can reach 14.9r / min. The gearbox adopts a three-stage planetary gear structure to drive an asynchronous generator with a maximum speed of 2016r / min, a rated voltage of 690V, and a grid access frequency of 50Hz. The unit tower is a steel pipe structure with a hub height of 80m. Data collection is completed by the SCADA system supporting the unit, and the sampling period is 10min. This experiment selects 36 consecutive sampling moments (i.e. ) data to build a reference time window, the window length corresponds to 6 hours, a sliding window mechanism is used, and the window sliding step is set to 1 sampling moment (i.e. The parameter N is set to 4, meaning the sum of the scores of the four highest-scoring variables serves as the ICI value for that window. The same experimental parameters are set for all three instances. To reduce the false alarm rate, a multi-window criterion mechanism is introduced: if the comprehensive status indicator (ICI) undergoes a sudden change at least three times within five consecutive time windows, the warning trigger condition is considered met.

[0106] Under normal operation, the target wind turbines are imported into the 136 (i.e. ) sampling time SCADA data set, based on the set window parameters, a reference window data matrix is ​​constructed and experiments are carried out. To intuitively present the time series evolution of variables, the obtained LDS values ​​are projected into a three-dimensional coordinate graph, where the X-axis represents the variable number, the Y-axis represents the time window number, and the Z-axis represents the LDS score of the corresponding variable. Figure 3 The LDS value distribution of each variable in 100 time windows of the wind turbine under normal operating conditions (i.e., visualization effect) is shown. Figure 4 It can be seen that when the target wind turbine is in normal operation, the ICI value corresponding to each time window fluctuates within the set range, and the overall trend remains stable without significant abnormal mutations. It is consistent with the unit operating status recorded in the operation log, verifying the accuracy of the state identification of this method under normal operating conditions. Figure 5 It further shows the LDS values ​​obtained for each variable within a certain time window.

[0107] In the case of generator bearing failure, the actual moment of wind turbine failure is taken as the time reference point, and the data of 136 sampling moments before and 30 sampling moments after the moment are respectively intercepted as analysis samples (i.e. The data from the first 36 sampling moments are used to construct a reference window. Subsequently, under the sliding time window mechanism, LDS values ​​are calculated for all monitored variables within each time window to quantify the dynamic correlation and change characteristics between variables.

[0108] Depend on Figure 6 The three-dimensional coordinate diagram can be divided into five operating states: the first (green) indicates the unit is in normal operation, during which the LDS values ​​of various variables fluctuate within a relatively stable range without experiencing abnormal mutations; the second (yellow) corresponds to the unit transitioning from normal to critical state, during which the LDS values ​​of some variables experience a continuous and significant increase for the first time, with the overall value much higher than the first category; the third (orange) indicates the unit is stable in critical state, with LDS values ​​falling back and remaining within an oscillating range, maintaining overall stability; the fourth (red) corresponds to the unit fault state, during which the LDS values ​​of some variables experience another sudden increase, forming a second localized spike; and the fifth (dark red) represents the steady-state phase after a fault occurs, during which the LDS values ​​return to a new stable range after experiencing a sudden change. This three-dimensional visualization clearly illustrates the dynamic evolution of the LDS values ​​of each monitored variable within each time window. By comparing and analyzing the LDS change trends of the second and fourth categories with those of other state types, the dynamic response characteristics of key variables in the process of the unit evolving from criticality to failure can be effectively identified, thereby providing key decision-making basis for early fault warning and identification of wind turbines.

[0109] Based on the LDS values ​​calculated above, the LDS values ​​of each variable in each time window are sorted, and the variables with the top four LDS values ​​are selected. Their LDS values ​​are accumulated as the comprehensive status index (ICI) value of the wind turbine. The dynamic evolution curve of the ICI values ​​of all time windows is obtained, as shown in the figure below: Figure 7 As shown in the figure, from the 1st to the 55th time windows, the ICI value remained stable below 0.2, with minimal fluctuations, indicating stable wind turbine operation. Starting from the 56th time window, the ICI curve showed its first significant abrupt change. It then fluctuated continuously from the 57th to the 59th time window, meeting the warning trigger conditions and triggering a warning, indicating that the turbine had transitioned from a stable state to a critical state. Subsequently, from the 60th to the 88th time window, the ICI value remained stable within the normal operating range, indicating that the wind turbine was stable in a critical state. Although the ICI value increased in the 89th, 94th, and 97th time windows, it did not reach the preset warning threshold, and therefore no warning signal was triggered. The 101st time window corresponds to the actual fault occurrence, indicating that the turbine had transitioned from a critical state to a fault dynamic process. Within the 30 time windows after the fault, the ICI value fluctuated before returning to the 0–0.2 range, maintaining a relatively stable state. This indicates that the system has entered a new stable state, namely, a wind turbine shutdown fault state. The results show that the early warning signal was issued in the 57th time window, 440 minutes earlier than the actual fault window, which verifies that this method can achieve effective early warning before the wind turbine bearing failure.

[0110] Figure 8 The specific distribution of local scores of each variable under typical working conditions is further explained: Figure 8 (a) corresponds to the scores of each variable in a certain time window when the wind turbine is in normal operating conditions; Figure 8 (b) corresponds to the scores of each variable at the moment when the unit enters the critical state; Figure 8 (c) in the equation reflects the characteristics of the unit being stable in a critical state; Figure 8 (d) and (e) in the figure represent fault alarm windows; Figure 8 (f) in the figure indicates that the wind turbine enters the fault steady state. Figure 7 Variables 4, 5, 6, and 8 have significantly higher scores than other variables, forming the key subnetwork for this time window. Upon investigation, these variables correspond to the average temperature of generator phase 1, the average temperature of generator phase 2, the average temperature of generator phase 3, and the average temperature of the generator bearings, respectively. Combined with the wind turbine's operation log, it can be determined that the outage was caused by a generator bearing failure. These abnormally scored variables align with the characteristics of the generator bearing failure-related outages recorded in the wind turbine operation log. Therefore, the identification of abnormal variables in the generator set can be achieved based on the key subnetwork variables.

[0111] In the case of hydraulic system failure, the same data selection strategy as that of the generator bearing failure is used to analyze the SCADA data before and after the typical hydraulic system failure of the target wind turbine. The three-dimensional coordinate diagram of the LDS value in 131 time windows is obtained by sliding window calculation, as shown in the figure below. Figure 9 shown.

[0112] Based on the above results, the ICI value of each time window is calculated according to the pre-set rules. Figure 10 The results show that from the 1st to the 22nd time windows, the ICI value remained stable below 0.2, with minimal fluctuations, indicating that the wind turbine was operating in a stable state during this period. In the 22nd time window, the ICI curve experienced its first significant abrupt change, followed by continuous fluctuations from the 23rd to the 27th time windows, meeting the warning criteria and successfully triggering an alert, indicating that the turbine's operating state had transitioned from a stable state to a critical state. From the 28th to the 100th time window, the ICI value remained stable within the 0–0.2 range, indicating that the wind turbine had entered a new stable state, namely a critical state. By the 101st time window, the ICI value increased significantly again, indicating that the turbine had transitioned from a critical state to a fault state, consistent with the actual fault occurrence time recorded in the turbine operation log. After the fault occurred, the ICI value fluctuated back to the 0–0.2 range, indicating that the system had entered a new stable state, namely, a shutdown fault state. These results indicate that the warning signal was issued 780 minutes before the actual fault time, validating the effectiveness of this method in providing early warning of hydraulic system failures.

[0113] Can be obtained from Figure 11 Variables 19, 21, 23, and 24 have significantly higher scores than other variables. Therefore, these variables constitute a key subnetwork. Upon investigation, they were found to correspond to the average hydraulic oil temperature, average nacelle temperature, average blade pitch angle, and average power generation at the grid access point, respectively. The abnormal changes in these variables closely match the characteristics of the shutdown failure caused by hydraulic system oil leakage recorded in the wind turbine operation log. Therefore, wind turbine fault location can be achieved based on the key variable network.

[0114] See also Figure 12 FIG. 1 is a structural diagram of a device according to an embodiment of the present invention, comprising:

[0115] The sample collection module 1201 collects the data of the continuous monitoring variables covered by the SCADA system of the wind turbine generator set itself, and selects The data of consecutive sampling moments are used as experimental samples and preprocessed;

[0116] The time window construction module 1202 sets the length and step size parameters of the sliding window and constructs the time window data matrix using the preprocessed experimental samples;

[0117] A differential network construction module 1203 calculates the Pearson correlation of the variable pair in the current time window and the previous time window, and constructs a differential network for the current time window based on the Pearson correlation of the variable pair; and constructs differential networks for all time windows;

[0118] Node score calculation module 1204 evaluates the variability and correlation characteristics of each node in the differential network of the current time window, and calculates the score of each node in the differential network of the current time window based on the variability and correlation characteristics; and calculates the scores of the nodes in the differential network of all time windows;

[0119] The node score visualization module 1205 maps the scores of the nodes corresponding to all variables in all time windows into a three-dimensional coordinate graph and draws a three-dimensional visualization graph of the entire process;

[0120] Fault prediction module 1206 selects the one with the highest score nodes, define the The accumulated score of each node is used as the comprehensive status index ICI of the wind turbine set. According to the dynamic change of the ICI value, early fault warning of the wind turbine set is realized. It can be seen that the present invention can send out a warning signal before a fault occurs, and perform regular and fixed-point detection of the critical changes in the operating status. It has been verified that it has certain engineering practical value.

[0121] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A wind turbine generator fault warning method based on dynamic network markers, characterized in that: The steps include: Collect the data of continuous monitoring variables covered by the wind turbine's own SCADA system, select The data of consecutive sampling moments are used as experimental samples and preprocessed; Set the length and step size parameters of the sliding window and use the preprocessed experimental samples to construct the time window data matrix; Calculating the Pearson correlation of the variable pair in the current time window and the previous time window, and constructing a differential network of the current time window based on the Pearson correlation of the variable pair; Construct a differential network for all time windows; Evaluate the variability and correlation characteristics of each node in the differential network of the current time window, and calculate the score of each node in the differential network of the current time window based on the variability and correlation characteristics; Calculate the scores for nodes in the difference network of all time windows; Map the scores of the nodes corresponding to all variables in all time windows to a three-dimensional coordinate graph, and draw a three-dimensional visualization of the entire process; Select the highest score nodes, define the The accumulated score of each node is used as the comprehensive status index ICI of the wind turbine, and the early fault warning of the wind turbine is realized according to the dynamic change of the ICI value.

2. The wind turbine generator set fault early warning method based on dynamic network markers according to claim 1 is characterized in that: The data of continuous monitoring variables covered by the SCADA system of the wind turbine generator set are collected, and the selected The data of consecutive sampling moments are used as experimental samples and preprocessed. The preprocessing refers to normalizing the data, which is expressed as: ； in, is the collected variable value, For variable values Normalized data of is the average value of each variable collected; is the standard deviation of the collected variable values.

3. The wind turbine generator set fault early warning method based on dynamic network markers according to claim 1, characterized in that: The method of setting the length and step size parameters of the sliding window and constructing a time window data matrix using the preprocessed experimental samples includes the following steps: Set the length of the sliding window to , the step size is , get the number of variables as ;Continuous acquisition variables in The variable value at each moment is used to obtain the time window data matrix , expressed as: ； in, Indicates the The variable in the current time window variable values; Repeat the above steps times, obtained A time window data matrix.

4. The wind turbine generator set fault early warning method based on dynamic network markers according to claim 1, characterized in that: The step of calculating the Pearson correlation of the variable pair in the current time window and the previous time window, and constructing the differential network of the current time window according to the Pearson correlation of the variable pair, comprises the following steps: Calculate the Pearson correlation between any pair of variables in the previous and current time windows, expressed as: ； in, Indicates the Variables in a time window and variables Pearson correlation of ; Representing variables In the The time window variable values, Representing variables In the The time window variable values; Indicates the variables within a time window The mean of the variable values, Indicates the variables within a time window The mean of the variable values; For the same pair of variables, the specific Pearson correlation coefficient is calculated based on the Pearson correlation of the previous time window and the current time window, which is expressed as: ； in, Representing variables and variables In the Specific Pearson correlation coefficient in the time window; Taking the variables as nodes, we determine whether there is an edge connection between the two nodes corresponding to the pair of variables based on whether the value of the specific Pearson correlation coefficient reaches the threshold. The above operations are performed on all variable pairs to construct the differential network of the current time window.

5. The wind turbine generator set fault early warning method based on dynamic network markers according to claim 1, characterized in that: The calculation process of the node score includes the following steps: The differential expression deviation CV of the calculation node is defined as: ； in, Representation node The first-order field is composed of nodes It is composed of nodes connected by edges. express The nodes in Represents a first-order domain The number of midpoints, Representation node the standard deviation of the values ​​in the current time window, Representation node The average of the values ​​in the current time window, Represents a first-order domain The standard deviation of the node value in the current time window, In the first-order domain The average value of the node in the current time window; Calculate the nodes of a node and its first-order neighbors The mean value is defined as: ； in, Representation node With first-order fields The average value of the specific correlation coefficients of all nodes in the current time window; Calculate the relationship between the nodes in the first-order neighborhood of the node and the nodes in the second-order neighborhood of the node average value , expressed as: ； in, Representation node The second-order neighborhood of The nodes of are composed of nodes connected by edges. Represents the second-order neighborhood The number of nodes, Represents the second-order neighborhood Nodes in Based on the above calculation results, the score of each node in the differential network under the current time window is calculated, which is expressed as: ； in, Representation node score.

6. The wind turbine generator set fault early warning method based on dynamic network markers according to claim 1, characterized in that: Mapping the scores of the nodes corresponding to all variables in all time windows to a three-dimensional coordinate graph and drawing a three-dimensional visualization graph of the entire process includes the following steps: Construct a three-dimensional coordinate system with the X-axis representing the variable, the Y-axis representing the time window, and the Z-axis representing the node score; The calculated scores of the nodes corresponding to each variable in each time window are placed in the three-dimensional coordinate axis to obtain a three-dimensional visualization diagram of the whole process.

7. The wind turbine generator set fault early warning method based on dynamic network markers according to claim 1, characterized in that: The wind turbine comprehensive status index ICI is expressed as: ； in, Representing variables The corresponding node score.

8. A wind turbine fault warning device based on dynamic network markers, characterized in that: include: The sample collection module collects the data of continuous monitoring variables covered by the wind turbine's own SCADA system and selects The data of consecutive sampling moments are used as experimental samples and preprocessed; The time window construction module sets the length and step size parameters of the sliding window and constructs the time window data matrix using the preprocessed experimental samples; A differential network construction module calculates the Pearson correlation of the variable pair in the current time window and the previous time window, and constructs the differential network of the current time window based on the Pearson correlation of the variable pair; Construct a differential network for all time windows; The node score calculation module evaluates the variability and correlation characteristics of each node in the differential network of the current time window, and calculates the score of each node in the differential network of the current time window based on the variability and correlation characteristics; Calculate the scores for nodes in the differential network of all time windows; The node score visualization module maps the node scores corresponding to all variables in all time windows into a three-dimensional coordinate graph and draws a three-dimensional visualization of the entire process; Fault prediction module, select the one with the highest score nodes, define the The accumulated score of each node is used as the comprehensive status index ICI of the wind turbine, and the early fault warning of the wind turbine is realized according to the dynamic change of the ICI value.

Images