Wind turbine generator gearbox fault monitoring method and system

Abstract

The invention provides a wind turbine generator gearbox fault monitoring method and system, and the method comprises the following steps: obtaining sample data of a wind turbine generator gearbox, and carrying out the preprocessing of the sample data, and obtaining the preprocessed sample data; performing local score calculation on the preprocessed sample data based on KNN-MST to obtain a local score; constructing a dual fault detection model, and calculating to obtain a final anomaly group and a corresponding anomaly group distance matrix based on the normalized local score of the sample and the dual fault detection model; and collecting real-time operation data of the wind turbine gearbox, calculating an abnormal score of the operation data in combination with the final abnormal group and the abnormal distance matrix, and obtaining an equipment operation state diagnosis result, thereby monitoring whether the wind turbine gearbox is abnormal or not. According to the method, the fault point is subjected to double diagnosis by calculating the local score of the data progress and combining the improved CUSUM algorithm and the improved Euclidean distance method, so that the occurrence of false alarm of the fault point is avoided, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of fault monitoring of wind power generation systems, and more specifically, to a method and system for fault monitoring of a wind turbine gearbox under power market conditions. Background Art

[0002] Under the goals of carbon peak and carbon neutrality, wind power generation will become one of the main forces in the growth of clean energy. However, wind turbines are usually installed in remote suburbs or coastal areas, greatly affected by environmental conditions, and fault shutdowns occur frequently, with relatively high operation and maintenance costs. The gearbox is a key device in the transmission system of wind turbines, with the longest average downtime for maintenance and relatively high repair costs. Conducting condition monitoring on it can detect potential faults as early as possible, reduce the fault frequency, and thus save operation and maintenance costs.

[0003] In the field of wind power generation, operation and maintenance data resources mainly come from the real-time data measured by various sensors installed on the wind turbines through the Supervisory Control and Data Acquisition (SCADA) system collected by the dispatching automation system. Each piece of data usually contains hundreds of parameter information that can reflect the working state of the wind turbine at that moment. These parameters include environmental parameters such as temperature and wind speed, and wind turbine state parameters such as output power, wind turbine speed, and bearing temperature. The SCADA system generates a large amount of historical data every year. However, at present, the wind power industry in China mainly conducts real-time detection and statistical analysis on this data. The massive historical data is stored in the database without being deeply mined, and the value of the data resources has not been fully utilized. In addition, at present, most wind farms in China still adopt an operation and maintenance strategy mainly based on corrective maintenance and regular maintenance. This traditional "passive" maintenance strategy is prone to "insufficient maintenance" or "excessive maintenance", and it has been difficult to meet the requirements of actual production. Therefore, the demand for intelligent operation and maintenance is particularly urgent.

[0004] The gearbox is one of the important mechanical transmission components of wind turbines, and its operating state will have a huge impact on the overall operation of wind turbines. If the fault state can be predicted in advance and the system can be maintained in time, it can prevent the system from occurring serious faults or even shutdowns, which will greatly reduce the operation and maintenance costs of the wind power system. Summary of the Invention

[0005] To solve the deficiencies in the prior art, the present invention provides a method for fault monitoring of a wind turbine gearbox, which can solve the problems of inaccurate and untimely fault monitoring and high monitoring costs in the prior art.

[0006] The present invention adopts the following technical solutions.

[0007] A method for fault monitoring of a wind turbine gearbox includes the following steps:

[0008] Obtain the sample data of the wind turbine gearbox and perform preprocessing to obtain the preprocessed sample data;

[0009] Calculate the local scores of the preprocessed sample data based on KNN-MST to obtain the normalized local scores of the samples;

[0010] Construct a double fault detection model, and calculate the final abnormal group and its corresponding abnormal group distance matrix based on the normalized local scores of the samples and the double fault detection model;

[0011] Collect the real-time operation data of the wind turbine gearbox, calculate the abnormal scores of the operation data in combination with the final abnormal group and the abnormal distance matrix, and obtain the equipment operation status diagnosis result, so as to monitor whether the wind turbine gearbox is abnormal.

[0012] Preferably, the sample data includes the operation data of the wind turbine gearbox under different equipment states. Among them, the SCADA data of the wind turbine is collected as the operation data, including the gearbox oil temperature, the gearbox bearing temperature, the nacelle temperature, the rotor speed, the environmental wind direction, and the active power; the equipment state includes that the fault state of the current wind turbine gearbox is the normal state or the fault state.

[0013] Preferably, preprocess the sample data to obtain the preprocessed sample data, specifically including:

[0014] Perform missing value processing, duplicate value processing, resolution processing, and normalization processing on the collected sample data in sequence to obtain the preprocessed sample data represents the number of preprocessed samples, represents the number of features after preprocessing.

[0015] Preferably, calculate the local scores of the preprocessed sample data based on KNN-MST to obtain the normalized local scores of the samples, specifically including:

[0016] Calculate the adjacency matrix D of the preprocessed sample data based on KNN;

[0017] Construct a local minimum spanning tree based on the adjacency matrix D, and calculate the minimum spanning tree scores of each sample in the preprocessed sample data based on the local minimum spanning tree;

[0018] Calculate the average minimum spanning tree scores of all neighboring points of each sample in the preprocessed sample data;

[0019] Calculate according to the minimum spanning tree scores of the samples and the average minimum spanning tree scores of the neighboring points to obtain the normalized local scores of the samples.

[0020] Preferably, calculating the adjacency matrix D of the preprocessed sample data based on KNN specifically includes:

[0021] Using the weighted Euclidean distance with exponential decay to measure the distance of all samples in the preprocessed sample data, and calculating the distance and between samples

[0022]

[0023] In the formula, represents the data of the th row in the preprocessed sample data represents the data of the t-th row in the preprocessed sample data , and

[0024] α represents the control parameter of the exponential function decay rate, υ represents the adjustment parameter of the decay starting point, and both α and υ are preset positive numbers;

[0025] The intermediate parameter z i satisfies:

[0026]

[0027] Among them, respectively represent the data of the th row and the i-th column, and the data of the t-th row and the i-th column in the preprocessed sample data

[0028] According to the calculation result of the distance measurement, select the first K nearest samples of each sample and store them in the nearest neighbor point matrix :

[0029]

[0030] Among them, and respectively represent the first nearest neighbor sample and the K-th nearest neighbor sample of the distance ;

[0031] Calculate the distance between each sample and its first K nearest neighbors based on the nearest neighbor point matrix and store it in the adjacency matrix :

[0032]

[0033] In the formula, represents the distance between the sample and the sample , Represents the distance of the k-th neighboring sample.

[0034] Preferably, the local minimum spanning tree is constructed based on the adjacency matrix D, and the minimum spanning tree scores of each sample in the preprocessed sample data are calculated based on the local minimum spanning tree, specifically including:

[0035] Based on the adjacency matrix D, for the preprocessed sample data construct a local minimum spanning tree for each sample and its first K neighboring points;

[0036] Based on the local minimum spanning tree of each sample in the preprocessed sample data calculate the minimum spanning tree scores of each sample in the preprocessed

[0037] sample data of each sample.

[0038] Preferably, the normalized local score of the sample is calculated according to the minimum spanning tree score of the sample and the average minimum spanning tree score of the neighboring points, specifically including:

[0039] For the m-th sample in the preprocessed sample data average the MST scores of all neighboring points of according to the following formula to obtain the average minimum spanning tree score of the neighboring points

[0040] Calculate the local score of the sample according to the minimum spanning tree score of the sample and the average minimum spanning tree score of the neighboring points

[0041]

[0042]

[0043] Calculate the local scores o of all samples in the preprocessed sample data

[0044]

[0045] Normalize each sample in the local scores o of all samples to obtain the normalized local scores of all samples in the preprocessed sample data

[0045] ​Preferably, a dual-fault detection model is constructed, and the final abnormal group and its corresponding abnormal group distance matrix are calculated based on the normalized local scores of the samples and the dual-fault detection model; specifically including:

[0046] Perform probability density analysis on the normalized local scores and obtain the values of the parameters μ and λ in the improved CUSUM model according to the analysis results;

[0047] Construct an improved CUSUM model based on the obtained parameters μ and λ, and perform preliminary diagnosis of the fault points based on the improved CUSUM model to obtain a temporary abnormal group;

[0048] Perform re-diagnosis of the fault points on the temporary abnormal group based on the improved Euclidean distance to obtain the final abnormal group and the abnormal group distance matrix.

[0049] Preferably, the preliminary diagnosis of the fault points based on the improved CUSUM model to obtain a temporary abnormal group specifically includes:

[0050] Preset a sliding window Calculate the normalized local scores within the sliding window The error Δ of the i-th sample : i :

[0051]

[0052] Accumulate the errors to obtain the cumulative sum at the current moment

[0053]

[0054] In the formula, η represents the offset, and its value range is [0, 0.01]; represents the cumulative sum at the previous moment;

[0055] During the process of accumulating the errors, set a time window w so that the cumulative sum is reset to zero every time a time window w is filled;

[0056] During the calculation process of the cumulative sum, if the cumulative sum obtained in a certain calculation is higher than the set alarm threshold, the sample point corresponding to the error at this time is an abnormal data point and is stored in the temporary abnormal group in.

[0057] Preferably, the re-diagnosis of the fault points on the temporary abnormal group based on the improved Euclidean distance to obtain the final abnormal group and the abnormal group distance matrix specifically includes:

[0058] Perform re-diagnosis on the temporary abnormal group Perform distance measurement on each abnormal data point in it, calculate the distance between any two abnormal data points, and obtain an abnormal distance matrix based on the distance measurement result

[0059]

[0060] In the formula represents a temporary abnormal group in the th th abnormal data point and the

[0061] Obtain an abnormal distance matrix Each row element in represents the distance value between the current point and other elements in the temporary abnormal group in. Sum the elements of the th row separately to obtain the th abnormal data point corresponding abnormal score

[0062]

[0063] Based on the above method, calculate and obtain The abnormal group score g of all sample data:

[0064]

[0065] Perform normalization processing on the abnormal group score g, and eliminate the data points with normalized abnormal scores greater than the preset score threshold in the abnormal group score g to obtain the final abnormal group and its corresponding abnormal group distance matrix

[0066] Preferably, the method for collecting the real-time operation data of the wind turbine gearbox, calculating the abnormal score by combining the obtained final abnormal group and abnormal distance matrix, and monitoring whether the wind turbine gearbox is abnormal specifically includes:

[0067] Collect the real-time SCADA data of the wind turbine gearbox as operation data and perform preprocessing on it;

[0068] Based on KNN-MST, calculate the local score of the preprocessed real-time operation sample data to obtain the local score corresponding to the real-time sample data;

[0069] Calculate the abnormal score of the local score based on the dual-fault monitoring model and determine whether the wind turbine gearbox is abnormal.

[0070] The present invention also provides a fault monitoring system for a wind turbine gearbox, which is used to implement the fault monitoring method for the wind turbine gearbox, and includes: a data acquisition module, a preprocessing module, a local score calculation module, a dual fault monitoring module, and an anomaly monitoring module;

[0071] The data acquisition module is used for sample data of the wind turbine gearbox and real-time operation data of the wind turbine gearbox to be monitored;

[0072] The preprocessing module is used to preprocess the data obtained by the data acquisition module to obtain preprocessed data;

[0073] The local score calculation module calculates the local score based on the KNN-MST algorithm for the preprocessed data;

[0074] The dual fault monitoring module is used to construct a dual fault detection model and calculate the normalized local score of the data based on the dual fault detection model to obtain the final abnormal group and its corresponding abnormal group distance matrix;

[0075] The anomaly monitoring module is used to calculate the anomaly score of the operation data according to the real-time operation data of the wind turbine gearbox collected, and combine the final abnormal group and the abnormal distance matrix to obtain the equipment operation status diagnosis result, so as to monitor whether the wind turbine gearbox is abnormal.

[0076] The beneficial effects of the present invention are as follows. Compared with the prior art, the present invention constructs a fault monitoring model for the wind turbine gearbox based on the operation data of the wind turbine gearbox and the equipment operation status, and realizes the real-time monitoring of the gearbox operation status based on the constructed model and the real-time collected operation data. Through the improvement of data processing and calculation methods, the present invention can accurately monitor the operation status of the wind turbine gearbox, realizing the saving of the overall operation and maintenance cost and the improvement of the power generation efficiency. The present invention has at least the following beneficial effects:

[0077] (1) The present invention designs and implements an exponentially decaying weighted Euclidean distance for distance measurement of data. The distance result calculated by this method can reduce the influence of too large difference in a certain dimension between data on the measurement result;

[0078] (2) The present invention designs and implements an improved double sliding window cumulative sum (CUSUM) algorithm. By obtaining the mean value in the cumulative sum calculation process through a sliding window and monitoring and resetting the cumulative sum through a window during the monitoring process after the cumulative sum calculation is completed, it can quickly respond to the dynamic changes of the data stream, improving the robustness of the algorithm and the timeliness of fault detection;

[0079] (3) The design of the present invention realizes a dual - fault monitoring method. First, an improved CUSUM is used to initially diagnose the fault points in the data stream. When an anomaly is detected, the improved Euclidean distance is used to measure the distance between the fault points and the anomaly group saved in the training stage and calculate the scores. Then, the fault points are re - diagnosed based on the scores, avoiding false alarms of fault points and reducing the operation and maintenance costs. Description of the Drawings

[0080] Figure 1 is the flow strategy diagram of the gearbox fault detection method proposed by the present invention;

[0081] Figure 2 is the schematic diagram of probability density analysis of data in the gearbox fault detection method proposed by the present invention;

[0082] Figure 3 is the structural diagram of the gearbox fault detection system in the present invention. Detailed Embodiment

[0083] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only a part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0084] As Figure 1 shown, the present invention proposes a gearbox fault monitoring method for wind turbines, which specifically includes:

[0085] Step 1: Obtain the operation data of the gearbox of the wind turbine under normal and fault states as sample data X respectively, including the collected operation data of the gearbox of the wind turbine, and pre - process the sample data X to obtain the pre - processed sample data

[0086] Specifically, collect the SCADA data of the gearbox of the wind turbine as the operation data, including the gearbox oil temperature, gearbox bearing temperature, nacelle temperature, rotor speed, ambient wind direction, and active power;

[0087] Pre - process the collected sample data X, which specifically includes:

[0088] Form the collected sample data into a matrix with M rows and N columns, where M represents the number of samples and N represents the number of features;

[0089] Preferably, the sampling frequency is set to 10 minutes per strip, and each piece of data includes the gearbox oil temperature, gearbox bearing temperature, nacelle temperature, rotor speed, ambient wind direction, and active power value of the wind turbine gearbox at the same sampling moment. A total of M pieces of data are collected.

[0090] For the sample data perform missing value processing, duplicate value processing, resolution processing, and normalization processing in sequence; among them, the missing value processing adopts the forward filling method. When performing duplicate value processing, only the samples that are repeated in all columns are deleted. When performing resolution processing, the sampling frequency of the data is modified from the original 10 minutes per strip to 1 hour per strip by taking the average value. When performing normalization processing, the linear normalization (Normalization) method is adopted;

[0091] Let the preprocessed sample data be a matrix with rows and columns, where represents the number of preprocessed samples, and represents the number of features after preprocessing. The feature points in the normalized data sample

[0092]

[0093] are calculated as follows: where represents the sample data in the th row and

[0094] after missing value processing, duplicate value processing, and resolution processing. The feature point at the column of the

[0095] Step 2: Calculate the local score of the preprocessed sample data based on KNN-MST to obtain the normalized local score of the sample data

[0096] Step 2-1: Calculate the adjacency matrix of the preprocessed sample data and using KNN, specifically including:

[0097]

[0098] where each row of data in the preprocessed sample data is used as a sample. Represents the preprocessed sample data The th row of data in Represents the preprocessed sample data The t-th row of data in, and

[0099] α represents the control parameter of the exponential function decay rate, υ represents the adjustment parameter of the decay starting point, both α and υ are positive numbers, and are preset by those skilled in the art according to specific data and corresponding manual experience;

[0100] The intermediate parameter z i Satisfies:

[0101]

[0102] Wherein, Respectively represent the preprocessed sample data The th row and the i-th column data, the t-th row and the i-th column data in, and

[0103] The distance result calculated by this method can reduce the influence on the measurement result when the difference in a certain dimension between samples is too large. According to step 2-1-1, the preprocessed sample data The distance between each row of samples and other rows of samples in can be calculated.

[0104] Step 2-1-2, according to the calculation result of the distance metric, select the first K nearest neighbor samples of each sample in the preprocessed sample data And store them in the nearest neighbor point matrix :

[0105]

[0106] Wherein, Represents the distance The k-th nearest neighbor sample, k ∈ [1, K], the size of K is preset by the technician, and

[0107] And Respectively represent the 1st nearest neighbor sample and the K-th nearest neighbor sample of the distance ;

[0108] Step 2-1-3, based on the nearest neighbor point matrix E, calculate the distance between each sample and its first K nearest neighbors and store it in the adjacency matrix :

[0109]

[0110] In the formula, represents a sample and the distance from the k-th nearest neighboring sample The distance calculation method refers to Step 2-1-1.

[0111] Step 2-2: Based on the adjacency matrix D, construct a local minimum spanning tree (MST), and use the total edge length of the MST as the MST score of the sample;

[0112] Step 2-2-1: Based on the adjacency matrix D, for the preprocessed sample data construct a local MST (Minimum Spanning Tree) for each sample and its first K neighboring points;

[0113] For the sample in the preprocessed sample data the generation of its local MST is as follows:

[0114] Take the data of the th row in the neighboring point matrix E and the corresponding sample and its neighboring points as the minimum edge connection Generate the minimum edge connection based on the adjacency matrix D to form a connected undirected graph

[0115] Among them, the minimum edge connection is:

[0116]

[0117] The generated connected undirected graph where is the vector composed of all edges in the connected undirected graph; takes values as the vector composed of the elements in the th row of the adjacency matrix D.

[0118] Arrange all the edge lengths in in ascending order;

[0119] Select each sample point in the minimum edge connection in turn to form a loop until all the sample

[0120] points in

[0121] are connected to form a local minimum spanning tree containing K edges Based on the above steps, the local minimum spanning trees constructed by each sample point in the preprocessed sample data

[0122] Step 2-2-2, based on preprocessed sample data The local minimum spanning tree of each sample in is used to calculate the preprocessed sample data. The MST score of each sample in;

[0123] For samples MST score The calculation method is to The local minimum spanning tree of The weighted sum of all the side lengths in is calculated as follows:

[0124]

[0125] In the formula, the minimum spanning tree is Representation sample The local minimum spanning tree of The length of the i-th side in i is the edge length weight, p i The calculation formula is as follows:

[0126]

[0127] In the formula, Traversing the minimum spanning tree The cth edge length in the minimum spanning tree is c∈[1,K-1]; the edge lengths in the minimum spanning tree are arranged in ascending order.

[0128] The preprocessed sample data is calculated based on the above method The MST scores of all samples in .

[0129] Step 2-3, calculate the preprocessed sample data The average MST score of all neighboring points of each sample in;

[0130] For samples According to the following formula The MST scores of all neighboring points Taking the average, we get The average MST score of neighboring points

[0131]

[0132] In the formula, Representation sample The MST score of the kth neighboring point. The calculation method of the MST score refers to step 2-2-2.

[0133] Step 2-4: Calculate the sample data based on the MST scores of each sample and the average MST scores of the neighboring points of each sample Calculate the local scores of each sample in

[0134] Specifically, the local score of sample is calculated as follows:

[0135]

[0136] Based on the above method, calculate the local scores of each sample in the sample data to obtain the local scores of all sample data

[0137] Furthermore, normalize the local scores o of all sample data to obtain the normalized local scores of all samples in the sample data

[0138] Specifically, for the local score of each sample among them perform normalization to obtain

[0139]

[0140] Integrate the normalized local scores of each sample to obtain the normalized local scores of all samples in the sample data

[0141] Step 3: Construct a dual-fault detection model, and calculate the final abnormal group and its corresponding abnormal group distance matrix based on the normalized local scores of the samples and the dual-fault detection model;

[0142] The constructed dual-fault monitoring model includes calculating the parameters of the model based on the normalized local scores of the samples, constructing an improved CUSUM model based on the parameters for the initial diagnosis of fault points; and the re-diagnosis of fault points based on the improved Euclidean distance to obtain the final abnormal group and its corresponding abnormal group distance matrix. Step 3 specifically includes:

[0143] Step 3-1: Perform probability density analysis on the normalized local scores and obtain the values of the parameters μ and λ in the improved CUSUM model according to the analysis results;

[0144] Use Gaussian kernel density estimation to perform probability density analysis on the normalized local scores of the data sample The analysis function is as follows: ​​

[0145]

[0146] In the formula, the parameter σ is the standard deviation of the normalized local score ;

[0147] Based on the probability density analysis results, draw The probability density distribution curve of is as shown in Figure 2 , and the parameter h satisfies:

[0148]

[0149] In Figure 2 , μ represents the mean value when the normalized local score follows a normal distribution, and λ represents the minimum threshold corresponding to the long tail when the normalized local score shows a long tail beyond natural fluctuations. The values of μ and λ are determined according to the probability density analysis results of the normalized local score .

[0150] Step 3-2: Construct an improved CUSUM model based on the obtained parameters μ and λ, and perform an initial diagnosis of the fault point based on the improved CUSUM model to obtain an adjacent time abnormal group, specifically including:

[0151] Step 3-2-1: Calculate the error Δ of the i-th sample in : i :

[0152]

[0153] In the formula, the sliding window needs to be set according to the data distribution and corresponding manual experience;

[0154] Step 3-2-2: Accumulate the errors to obtain the cumulative sum s t at the current moment:

[0155]

[0156] In the formula, η represents the offset, and its value range is [0, 0.01]; represents the cumulative sum at the previous moment.

[0157] During the process of accumulating the errors, set a time window w so that every time a time window w is filled, the cumulative sum is reset to zero. The setting of the time window w needs to be set according to the specific data and corresponding manual experience;

[0158] Step 3-2-3, during the calculation of the cumulative sum s, if the cumulative sum s obtained in a certain calculation is higher than the set warning threshold θ, record the sample point corresponding to the error at this time as an abnormal data point and store it in the temporary abnormal group ;

[0159] wherein, represents the number of temporary abnormal points, and the setting of the threshold θ needs to be set according to specific data and corresponding manual experience.

[0160] Step 3-3, based on the improved Euclidean distance, re-diagnose the fault points in the temporary abnormal group to obtain the final abnormal group and the abnormal group distance matrix, so as to realize the double detection of the fault points;

[0161] Perform distance measurement on each abnormal data point in the temporary abnormal group Calculate the distance between any two abnormal data points, and obtain the abnormal distance matrix according to the distance measurement result

[0162]

[0163] In the formula, represents the distance between the th abnormal data point and the th

[0164] Obtain the abnormal distance matrix Each row element in represents the distance value between the current point and other elements in the temporary abnormal group . Sum the values of the th row alone to obtain the abnormal score corresponding to the

[0165]

[0166] Based on the above method, calculate the abnormal group score g of all sample data in the temporary abnormal group :

[0167]

[0168] And perform normalization processing on g. Eliminate the data points with normalized abnormal scores greater than the score threshold in g to obtain the final abnormal group and its corresponding abnormal group distance matrix where the setting of the score threshold needs to be set according to specific data and corresponding manual experience.

[0169] Step 4: Collect the real-time operation data of the wind turbine gearbox, calculate the anomaly score of the operation data in combination with the final anomaly group and the anomaly distance matrix, and obtain the device operation status diagnosis result, so as to monitor whether the wind turbine gearbox is abnormal.

[0170] Specifically, Step 4 includes:

[0171] Step 4-1: Collect the real-time operation data of the wind turbine gearbox and preprocess it;

[0172] Collect the real-time SCADA data of the wind turbine gearbox as the operation data, including the gearbox oil temperature, gearbox bearing temperature, nacelle temperature, rotor speed, ambient wind direction, and active power; the preprocessing method refers to the way of preprocessing the collected real-time sample data in Step 1.

[0173] Step 4-2: Anomaly score calculation stage: Calculate the local score of the preprocessed real-time sample data based on KNN-MST to obtain the local score corresponding to the real-time sample data;

[0174] Specifically, referring to the calculation method in Step 2, the data point x in the preprocessed real-time operation data test is successively subjected to KNN calculation with the final anomaly group obtained in Step 3 to find its K nearest neighbors, and then a local minimum spanning tree is generated Based on the local minimum spanning tree the local score of x test is obtained

[0175] Step 4-3: Fault detection stage: Calculate the anomaly score of the local score based on the dual fault monitoring model and judge whether the wind turbine gearbox is abnormal; The anomaly score of is calculated and it is judged whether the wind turbine gearbox is abnormal;

[0176] Specifically, Step 4-3 includes:

[0177] Step 4-3-1: Calculate the error of the local score and accumulate it. If the current cumulative sum s test is greater than the set warning threshold θ, then construct the anomaly distance matrix between the current data and the data points in the final anomaly group Based on calculate the anomaly score of the current data, update the anomaly group score vector g, and obtain the normalized anomaly score g of the current data after normalizing g ; if the current cumulative sum s test ; if the current cumulative sum s testIf it is less than or equal to the set alarm threshold θ, return to step 4-1 to continue collecting real-time data and accumulating it.

[0178] Step 4-3-2, combine the normalized anomaly score g of the current data test and the score threshold to determine whether the gearbox of the wind turbine is abnormal:

[0179] If the normalized anomaly score g of the current data test is less than the score threshold it indicates that the gearbox of the wind turbine is in an abnormal state and an alarm is issued;

[0180] If the anomaly score g of the current data test is greater than or equal to the score threshold θ, it indicates that the gearbox of the wind turbine is in a normal state and continue monitoring.

[0181] As Figure 3 shown, the present invention also proposes a fault monitoring system for the gearbox of a wind turbine, which is used to implement the above-mentioned fault monitoring method for the gearbox of a wind turbine. The system specifically includes: a data acquisition module, a preprocessing module, a local score calculation module, a dual fault monitoring module, and an anomaly monitoring module;

[0182] Specifically, the data acquisition module is used for sample data of the gearbox of the wind turbine and for collecting real-time operation data of the gearbox of the wind turbine to be monitored;

[0183] The preprocessing module is used to preprocess the data obtained by the data acquisition module to obtain preprocessed data;

[0184] The local score calculation module calculates the local score based on the KNN-MST algorithm for the preprocessed data;

[0185] The dual fault monitoring module is used to construct a dual fault detection model and calculate the normalized local score of the data based on the dual fault detection model to obtain the final abnormal group and its corresponding abnormal group distance matrix;

[0186] The anomaly monitoring module is used to calculate the anomaly score according to the real-time operation data of the gearbox of the wind turbine collected and to monitor whether the gearbox of the wind turbine is abnormal in combination with the final abnormal group and the abnormal distance matrix obtained by the dual fault monitoring module.

[0187] The beneficial effect of the present invention is that, compared with the prior art, the present invention initially diagnoses the fault points of the data stream by improving CUSUM. When an anomaly is detected, the fault points will be measured for distance and score calculation with the abnormal group saved in the training stage through the improved Euclidean distance, and the fault points will be rediagnosed through the score, avoiding the occurrence of false alarms of the fault points and reducing the operation and maintenance costs.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or make equivalent substitutions, and any modification or equivalent substitution that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A method for monitoring a wind turbine gearbox fault, characterized in that: The steps include: Acquire sample data of a wind turbine gearbox and perform preprocessing to obtain preprocessed sample data; Based on KNN-MST, local scores are calculated for the preprocessed sample data to obtain the normalized local scores of the samples; A dual fault detection model is constructed, and the final anomaly group and its corresponding anomaly group distance matrix are calculated based on the normalized local score of the sample and the dual fault detection model; The real-time operation data of the wind turbine gearbox is collected, and the abnormal score of the operation data is calculated by combining the final abnormal group and the abnormal distance matrix to obtain the equipment operation status diagnosis result, so as to monitor whether the wind turbine gearbox is abnormal.

2. The wind turbine gearbox fault monitoring method according to claim 1, characterized in that: The sample data includes the operating data of the wind turbine gearbox under different equipment states, wherein the SCADA data of the wind turbine is collected as the operating data, including the gearbox oil temperature, gearbox bearing temperature, cabin temperature, rotor speed, ambient wind direction and active power; the equipment state includes whether the fault state of the current wind turbine gearbox is a normal state or a fault state.

3. The wind turbine gearbox fault monitoring method according to claim 2, characterized in that: The sample data is preprocessed to obtain preprocessed sample data, specifically including: The collected sample data is processed by missing value processing, duplicate value processing, resolution processing and normalization processing in turn to obtain the preprocessed sample data represents the number of samples after preprocessing, Represents the number of features after preprocessing.

4. The wind turbine gearbox fault monitoring method according to claim 1, characterized in that: Based on KNN-MST, local scores are calculated for the preprocessed sample data to obtain normalized local scores of the samples, including: Calculate the adjacency matrix D of the preprocessed sample data based on KNN; Construct a local minimum spanning tree based on the adjacency matrix D, and calculate the minimum spanning tree score of each sample in the preprocessed sample data based on the local minimum spanning tree; Calculate the average minimum spanning tree score of all neighboring points of each sample in the preprocessed sample data; The normalized local score of the sample is calculated based on the minimum spanning tree score of the sample and the average minimum spanning tree score of the neighboring points.

5. The wind turbine gearbox fault monitoring method according to claim 3, characterized in that: The adjacency matrix D of the sample data after preprocessing is calculated based on KNN, specifically includes: The exponentially decayed weighted Euclidean distance is used to measure the distance of all samples in the preprocessed sample data and calculate the sample and The distance between In the formula, Represents the sample data after preprocessing The Row data, Represents the sample data after preprocessing The tth row of data in α represents the control parameter of the exponential function attenuation speed, υ represents the adjustment parameter of the attenuation starting point, and both α and υ are preset positive numbers; Intermediate parameter z i satisfy: in, Represent the sample data after preprocessing Middle The data of the i-th row and the i-th column, and the data of the t-th row and the i-th column, and According to the calculation results of the distance metric, the first K neighboring samples of each sample are selected and stored in the neighboring point matrix middle: in, and Respectively indicate distance The 1st neighboring sample and the Kth neighboring sample of ; Based on the neighboring point matrix, the distance between each sample and its first K neighboring points is calculated and stored in the adjacency matrix middle: In the formula, Representation sample With sample The distance between Indicates distance The kth neighboring sample of .

6. The wind turbine gearbox fault monitoring method according to claim 3, characterized in that: The method of constructing a local minimum spanning tree based on the adjacency matrix D and calculating the minimum spanning tree score of each sample in the preprocessed sample data based on the local minimum spanning tree specifically includes: Based on the adjacency matrix D, the preprocessed sample data Construct a local minimum spanning tree for each sample in and its first K neighboring points; Based on the preprocessed sample data The local minimum spanning tree of each sample in is used to calculate the preprocessed sample data. The minimum spanning tree score for each sample in .

7. The wind turbine gearbox fault monitoring method according to claim 3, characterized in that: The calculation based on the minimum spanning tree score of the sample and the average minimum spanning tree score of the neighboring points to obtain the normalized local score of the sample specifically includes: For the preprocessed sample data The mth sample in According to the following formula The MST scores of all neighboring points Taking the average, we get Average minimum spanning tree score of neighboring points According to the sample The minimum spanning tree score of the sample is calculated by the average minimum spanning tree score of the neighboring points The local score Calculate the preprocessed sample data The local score o of all samples in: Normalize each sample in the local score o of all samples to obtain the preprocessed sample data Normalized local score of all samples 8. The wind turbine gearbox fault monitoring method according to claim 1, characterized in that: A dual fault detection model is constructed, and the final abnormal group and its corresponding abnormal group distance matrix are calculated based on the normalized local score of the sample and the dual fault detection model; specifically, the following steps are included: Normalized local score Conduct probability density analysis and obtain the values ​​of parameters μ and λ in the improved CUSUM model based on the analysis results; An improved CUSUM model is constructed based on the obtained parameters μ and λ, and the initial diagnosis of the fault point is performed based on the improved CUSUM model to obtain the adjacent time abnormal group; Based on the improved Euclidean distance, the fault points of adjacent anomaly groups are re-diagnosed to obtain the final anomaly groups and the anomaly group distance matrix.

9. The wind turbine gearbox fault monitoring method according to claim 8, characterized in that: The initial diagnosis of the fault point based on the improved CUSUM model to obtain the adjacent time abnormal group specifically includes: Preset Sliding Window In the sliding window Calculate the normalized local score internally The error Δ of the i-th sample i : Accumulate the error to get the current cumulative sum s t : Where η represents the offset, and its value range is [0,0.01]; Indicates the cumulative sum of the previous moment; In the process of accumulating errors, a time window w is set so that the accumulated sum is reset to zero every time a time window w is accumulated. During the cumulative sum calculation process, if the cumulative sum obtained in a certain calculation is higher than the set alarm threshold, the sample point at the current moment corresponding to the error at this time is regarded as an abnormal data point and stored in the temporary abnormal group. middle.

10. The wind turbine gearbox fault monitoring method according to claim 8, characterized in that: The method of re-diagnosing the fault points of adjacent abnormal groups based on the improved Euclidean distance to obtain the final abnormal groups and the abnormal group distance matrix specifically includes: Adjacent time abnormal group The distance between any two abnormal data points is calculated, and the abnormal distance matrix is ​​obtained based on the distance measurement results. In the formula, Indicates a temporary exception group The The abnormal data point and The distance between abnormal data points, and Get the outlier distance matrix Each row element in represents the current point and the temporary anomaly group The distance value between other elements in The rows are summed individually to get Outlier data points The corresponding anomaly score Based on the above method, we can calculate The abnormal group score g of all sample data is: Normalize the abnormal group score g, and set the normalized abnormal score in the abnormal group score g greater than the preset score threshold The data points are removed to obtain the final abnormal group And its corresponding abnormal group distance matrix 11. The wind turbine gearbox fault monitoring method according to claim 1, characterized in that: The real-time operation data of the wind turbine gearbox is collected, and the abnormality score is calculated by combining the obtained final abnormal group and the abnormal distance matrix, and monitoring whether the wind turbine gearbox is abnormal, specifically including: Collect real-time SCADA data of wind turbine gearbox as operation data and pre-process it; Based on KNN-MST, local scores are calculated for the preprocessed real-time running sample data to obtain local scores corresponding to the real-time sample data; Based on the dual fault monitoring model, the abnormal score of the local score is calculated and whether the wind turbine gearbox is abnormal is determined.

12. A wind turbine gearbox fault monitoring system, used to implement the wind turbine gearbox fault monitoring method according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, preprocessing module, local score calculation module, dual fault monitoring module, abnormality monitoring module; The data acquisition module is used for the sample data of the wind turbine gearbox and the real-time operation data of the wind turbine gearbox to be monitored; The preprocessing module is used to preprocess the data acquired by the data acquisition module to obtain preprocessed data; The local score calculation module performs local score calculation on the preprocessed data based on the KNN-MST algorithm; The dual fault monitoring module is used to construct a dual fault detection model, and calculate the normalized local score of the data based on the dual fault detection model to obtain the final abnormal group and its corresponding abnormal group distance matrix; The abnormality monitoring module is used to calculate the abnormality score of the operating data based on the real-time operating data of the wind turbine gearbox collected, combined with the final abnormal group and the abnormal distance matrix, and obtain the equipment operating status diagnosis result, so as to monitor whether the wind turbine gearbox is abnormal.

Images