Fault judgment method of wind driven generator and computer equipment

Abstract

The invention provides a fault judgment method of a wind driven generator and computer equipment. The fault judgment method comprises the following steps: acquiring a first sample data set of the wind driven generator of a to-be-judged wind driven generator set; based on the data distribution of each piece of first sample data in the first sample data set, screening the first sample data set to obtain a second sample data set; comparing the second sample data set with a reference data set to obtain a fault discrimination condition for discriminating the fault of the wind driven generator; and based on the operation data of the wind driven generator and the fault discrimination condition, carrying out fault discrimination on the wind driven generator. The problem that efficient and accurate fault judgment of a lightweight algorithm is difficult to realize is solved, fault judgment can be carried out on the wind driven generator by screening and optimizing sample data and flexibly determining fault judgment conditions, and efficient and accurate fault judgment can be carried out through the lightweight algorithm.

Description

Technical Field

[0001] The present disclosure relates to the field of wind power generation, and more particularly, to a method for fault discrimination of a wind turbine and a computer device. Background Art

[0002] As the main force of new energy power, the healthy development of the wind power industry will play a crucial role in the growth of electric power. Among them, the stable operation of wind turbine generator equipment has become an important guarantee for promoting the process of the wind power industry.

[0003] In the field of wind power generation, timely and accurate prediction and discrimination of the operating status of key components of wind turbine generators can provide effective technical support for formulating maintenance operation plans on site, and play a key role in reducing maintenance costs and time costs. In particular, with the full-speed advancement of wind power technology from onshore to offshore, the single-unit capacity of wind turbines has increased year by year, and the difficulty of carrying out maintenance operations on offshore wind turbines is high. The operating status of key components of wind turbines has become an issue that has been increasingly emphasized in the wind power field. Therefore, how to achieve fast and accurate fault discrimination of wind turbine components is the key to ensuring the operational reliability and economic reliability of wind turbines.

[0004] In related technologies, in some methods, algorithms such as machine learning and neural networks can be used to discriminate faults of wind turbine components. However, in such methods, the requirements for the magnitude and granularity of data are relatively high, the operation space is complex, and the computing power of the analysis device is required to be relatively high; in other methods, lightweight algorithms can be used for fault discrimination. However, in such methods, the quality requirements for sample data are relatively high, and in the case of a large amount of sample data, it will also cause low operation efficiency. Summary of the Invention

[0005] In view of the problem that it is difficult to achieve efficient and accurate fault discrimination of lightweight algorithms in related technologies, the present disclosure provides a method for fault discrimination of a wind turbine and a computer device.

[0006] The first aspect of the present disclosure provides a method for fault discrimination of a wind turbine, and the fault discrimination method: obtaining a first sample data set of a wind turbine of a wind turbine generator set to be discriminated; screening the first sample data set based on the data distribution of each first sample data in the first sample data set to obtain a second sample data set; comparing the second sample data set with a reference data set to obtain a fault discrimination condition for discriminating faults of the wind turbine, where the reference data set includes data representing the wind turbine in a non-fault state; and discriminating faults of the wind turbine based on the operation data of the wind turbine and the fault discrimination condition.

[0007] Optionally, the first sample data in the first sample data set is determined as follows: Obtain a plurality of generator parameters characterizing the operating state of the wind turbine, where the plurality of generator parameters include the operating parameters and / or energy efficiency parameters of the wind turbine generator set; Determine the correlation coefficient between each pair of generator parameters among the plurality of generator parameters, where the correlation coefficient represents the correlation between two generator parameters; Delete one generator parameter in the parameter pair whose correlation coefficient satisfies a preset correlation condition from the plurality of generator parameters to obtain the filtered generator parameters; Use the sample data corresponding to the filtered generator parameters as the first sample data.

[0008] Optionally, the operating parameters include the three-phase current of the generator, and include a plurality of parameters among the maximum deviation of the three-phase current, wind speed, active power, ambient temperature, nacelle temperature, generator temperature, and generator speed, where: Use the other parameters in the operating parameters except the three-phase current of the generator as the filtered generator parameters.

[0009] Optionally, the first sample data set is filtered as follows: Perform distance measurement on each first sample data in the first sample data set to obtain the distribution distance of each first sample data relative to the sample representative data, where the sample representative data characterizes the data central tendency of the first sample data set; Based on the distribution distance of each first sample data and a preset distance filtering condition, filter the first sample data set to obtain the second sample data set.

[0010] Optionally, the distribution distance of each first sample data is obtained as follows: Based on each first sample data in the first sample data set, the preset supremum of the measurement distances between the same type of sample data, and the preset infimum of the measurement distances between different types of sample data, optimize the preset measurement matrix to obtain an optimized measurement matrix, where the measurement matrix is used to determine the measurement distance between samples; Based on the first sample data, the sample representative data, and the optimized measurement matrix, determine the distribution distance.

[0011] Optionally, the distance filtering condition is determined as follows: Based on the distribution distances of all first sample data in the first sample data set, determine the minimum distance and the maximum distance in the distribution distances; Based on the minimum distance and the maximum distance, determine the filtering step; Based on the filtering step and the minimum distance, determine the upper limit distance of the distribution distance of the first sample data; Based on the upper limit distance, determine the distance filtering condition, where the distance filtering condition means filtering out the first sample data whose distribution distance is less than or equal to the upper limit distance from the first sample data set.

[0012] Optionally, the fault discrimination condition is obtained by: inputting the second sample data set and the reference data set into a non-linear state estimation model, comparing the data features of the second sample data set and the reference data set to obtain a residual value; and obtaining the fault discrimination condition based on the residual value.

[0013] Optionally, the operating data includes a first operating data set, and the fault discrimination of the wind turbine is performed by: weighting the second sample data in the second sample data set based on the first operating data set of the wind turbine to obtain a third sample data set; comparing the second sample data set and the third sample data set to obtain a fault parameter; and performing fault discrimination on the wind turbine based on the fault parameter and the fault discrimination condition.

[0014] Optionally, the third sample data set is obtained by: analyzing the data features of the first operating data set according to time sequence, determining the weights of the second sample data corresponding to each first operating data; and weighting the second sample data in the second sample data set based on the weights to obtain the third sample data set.

[0015] Optionally, the performing fault discrimination on the wind turbine based on the fault parameter and the fault discrimination condition includes: determining that the wind turbine is in a normal state in response to the fault parameter satisfying a first fault value range; determining that the wind turbine has a magnet steel shedding fault in response to the fault parameter satisfying a second fault value range; and determining that the wind turbine has a three-phase current imbalance fault in response to the fault parameter satisfying a third fault value range, where the upper limit value of the first fault value range is less than the lower limit value of the second fault value range, and the upper limit value of the second fault value range is less than the lower limit value of the third fault value range.

[0016] A second aspect of the present disclosure provides a computer device, which includes: at least one processor; and at least one memory storing computer-executable instructions, where when the computer-executable instructions are run by the at least one processor, the at least one processor is caused to execute the fault discrimination method of the wind turbine according to the embodiments of the present disclosure.

[0017] A third aspect of the present disclosure provides a computer-readable storage medium, when the instructions in the computer-readable storage medium are run by at least one processor, the at least one processor is caused to execute the fault discrimination method of the wind turbine according to the embodiments of the present disclosure.

[0018] According to the fault discrimination method and computer device of a wind turbine of the present disclosure, the first sample data set of the wind turbine can be screened to obtain a second sample data set, and the screened second sample data set can be compared with a reference data set to determine a fault discrimination condition for fault discrimination. Thus, based on this fault discrimination condition and the operation data of the wind turbine, fault discrimination of the wind turbine can be performed. In this way, the sample data can be optimized by screening and the fault discrimination condition can be flexibly determined to perform fault discrimination on the wind turbine. This method is applicable to algorithms with high requirements for the quality of sample data and small-scale sample data, so that efficient and accurate fault discrimination can be performed through a lightweight algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 FIG. is a schematic flow chart showing a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0020] Figure 2 FIG. is a schematic flow chart showing the step of determining the first sample data in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0021] Figure 3 FIG. is a schematic diagram showing an example of calculating parameter correlation in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0022] Figure 4 FIG. is a schematic flow chart showing the step of screening the first sample data set in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0023] Figure 5 FIG. is a schematic flow chart showing the step of obtaining the distribution distance of each first sample data in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0024] Figure 6 FIG. is a schematic flow chart showing the step of determining a distance screening condition in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0025] Figure 7 FIG. is a schematic framework diagram showing an example of fault discrimination in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0026] Figure 8 FIG. is a schematic flow chart showing the step of obtaining a fault discrimination condition in the fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0027] Figure 9It is a flowchart showing the steps of fault discrimination for a wind turbine in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0028] Figure 10 It is a schematic diagram showing an example of sample data in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0029] Figure 11 It is a schematic diagram showing an example of the trend change of the metric learning objective function value in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0030] Figure 12 and Figure 13 It is a schematic diagram showing examples of sample data before and after optimization in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0031] Figure 14 and Figure 15 It is a schematic diagram showing examples of the Mahalanobis distance distribution of sample data before and after optimization in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0032] Figure 16 It is a schematic diagram showing a DTW path of an example in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0033] Figure 17 It is a schematic diagram showing an example of verifying the residual trend of sample data in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure.

[0034] Figure 18 and Figure 19 It is a schematic diagram showing examples of the generator state residual trend of sample data before and after optimization in a fault discrimination method of a wind turbine according to an exemplary embodiment of the present disclosure. Detailed Embodiments

[0035] The following detailed embodiments are provided to help the reader obtain a comprehensive understanding of the methods, devices, and / or systems described herein. However, various changes, modifications, and equivalents of the methods, devices, and / or systems described herein will be apparent after understanding the disclosure of the present application. For example, the order of operations described herein is merely exemplary and is not limited to those set forth herein, but may be changed as will be apparent after understanding the disclosure of the present application, except for operations that must occur in a specific order. Additionally, descriptions of features known in the art may be omitted for greater clarity and conciseness.

[0036] The features described herein can be implemented in different forms and should not be construed as limited to the examples described herein. Instead, the examples described herein are provided only to illustrate some of the many viable ways of implementing the methods, apparatuses, and / or systems described herein, which will be apparent after understanding the disclosure of the present application.

[0037] As used herein, the term "and / or" includes any one of the associated listed items and any combination of any two or more of them.

[0038] Although terms such as "first", "second", and "third" may be used herein to describe various components, elements, regions, layers, or parts, these components, elements, regions, layers, or parts should not be limited by these terms. Instead, these terms are only used to distinguish one component, element, region, layer, or part from another. Thus, a first component, a first element, a first region, a first layer, or a first part as referred to in the examples described herein may also be referred to as a second component, a second element, a second region, a second layer, or a second part without departing from the teachings of the examples.

[0039] In the specification, when an element (such as a layer, a region, or a substrate) is described as "on" another element, "connected to" or "coupled to" another element, the element may be directly "on" another element, directly "connected to" or "coupled to" another element, or there may be one or more other elements therebetween. In contrast, when an element is described as "directly on" another element, "directly connected to" or "directly coupled to" another element, there may be no other elements therebetween.

[0040] The terms used herein are only for describing various examples and are not intended to limit the disclosure. Unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. The terms "comprising", "including", and "having" specify the presence of the recited features, quantities, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof.

[0041] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs after understanding the present disclosure. Unless explicitly defined herein, terms (such as those defined in a general dictionary) should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and should not be interpreted in an idealized or overly formal manner.

[0042] In addition, in the description of the examples, when the detailed description of the relevant structures or functions that are considered to be well-known will cause ambiguous interpretation of the present disclosure, such detailed description will be omitted.

[0043] As described above, in the related art, with the development of wind power generation technology, the requirements for fault discrimination of the operating states of key components of wind turbines are getting higher and higher. As the single-unit capacity of the turbines increases year by year, the operating states of the key components of the turbines have become an issue that is increasingly emphasized in the wind power field. Therefore, how to achieve fast and accurate fault discrimination of the turbine components is the key to ensuring the operational reliability and economic reliability of the turbines.

[0044] In particular, in order to pursue better wind resources, the location selection of wind farms is gradually extending from land to the sea. Offshore wind farms are usually far from land, have a harsh natural environment, and have a large number of power generation devices. Affected by these factors, compared with traditional power generation devices, the deterioration or functional failure of the key components of wind turbines will bring higher maintenance time costs and operation difficulties.

[0045] In the related art, algorithms such as machine learning and neural networks can be used for fault discrimination. These algorithms are relatively mature and have certain application fields. However, such algorithms usually have high requirements for the magnitude and granularity of data, have complex operation spaces, have high requirements for the computing power of analysis devices, and moreover, the characteristics of "black box" and "high dimensionality" of such algorithms are obvious, resulting in the lack of interpretability of the analysis results.

[0046] In this regard, some lightweight algorithms can also be used for fault discrimination. However, due to the certain limitations of the computing power of lightweight algorithms, they have high requirements for the quality of sample data, and in the case of a large amount of sample data, it will also cause low operation efficiency.

[0047] In view of the above problems, the present disclosure provides a fault discrimination method and a computer device for a wind turbine to solve or at least alleviate the above problems.

[0048] According to a first aspect of an exemplary embodiment of the present disclosure, a fault discrimination method for a wind turbine is provided. The fault discrimination method can be executed by a computer device having computing capabilities. The computer device can be, for example, a terminal device or a server. Among them, the terminal device can be, for example, a tablet computer, a laptop computer, a digital assistant, etc.; the server can be an independent server, a server cluster, a cloud computing platform, or a virtualization center.

[0049] In an example application scenario, a computer device can obtain a first sample data set of a wind turbine to be discriminated, and based on the data distribution of each first sample data in the first sample data set, screen the first sample data set to obtain a second sample data set. The computer device can also compare the second sample data set with a reference data set to obtain a fault discrimination condition for discriminating faults of the wind turbine, where the reference data set includes data characterizing the wind turbine in a non-fault state. The computer device can further perform fault discrimination on the wind turbine based on the operation data of the wind turbine and the fault discrimination condition.

[0050] Here, the computer device can be disposed, for example, at a wind turbine generator set or a wind farm, and can be communicatively connected to, for example, a measurement device, a control system, or a data center of the wind turbine generator set or the wind farm, so as to obtain data such as unit operation data required for performing the above method.

[0051] The fault discrimination method for a wind turbine according to the present disclosure can screen and optimize sample data, flexibly determine a fault discrimination condition, and reasonably screen the sample data, so as to ensure efficient and accurate fault discrimination of the wind turbine, and this method is applicable to algorithms with high requirements for the quality of sample data and small-volume sample data, thereby enabling fault discrimination based on a lightweight algorithm.

[0052] The fault discrimination method for a wind turbine according to the present disclosure can include the following steps:

[0053] As Figure 1 shown, in step S110, a first sample data set of the wind turbine of the wind turbine generator set to be discriminated can be obtained.

[0054] As an example, the wind turbine generator set according to the example of the present disclosure can be a direct-drive unit such as a 6.0 MW direct-drive wind turbine generator set. The cut-in wind speed of the unit can be, for example, 3 m / s, the rated wind speed can be, for example, 11.5 m / s, and the rated speed of the generator can be, for example, 17.3 rpm. To ensure the low-speed output capacity and the compactness of the body structure, the stator can adopt a design with two sets of three-phase windings with a phase difference, the generator has an "outer-rotor and inner-stator" structure, and the generator adopts a passive air-cooling method to achieve natural regulation of the temperature of the generator body through wind speed changes. The generator and the wind wheel are directly connected by a nested main shaft, that is, the wind wheel directly drives the generator to generate electricity, and the whole mechanical structure is complete and tight, and the reliability of the unit equipment is high. Here is only an example of the wind turbine generator set to which this method can be applied, and its application is not limited thereto, and it can also be applied to other types of wind turbine generator sets.

[0055] In this step S110, the first sample dataset may include multiple pieces of first sample data. As an example, the first sample data may be data representing the operating state of a wind turbine. For example, it can be analyzed from two dimensions of the operation and energy efficiency of the generator to extract key feature indicators as data information representing the operating state of the generator.

[0056] Specifically, due to the relatively complex operating mechanism of the wind turbine, there may be various combinations of the parameters used to describe the operating state of the generator. The Supervisory Control And Data Acquisition (SCADA) system for the unit records more than thirty parameters representing the generator state. In addition, there are also some environmental parameters indirectly related to them. Reasonably selecting the input parameters of the discrimination model is the primary task of improving the discrimination result and reducing the redundancy of the model. According to the embodiments of the present disclosure, in order to improve the completeness of the description of the operating state of the generator of the unit, the input parameters can be divided into two categories: operating parameters and energy efficiency parameters.

[0057] As an example, in this step S110, the first sample data in the first sample dataset can be determined in the following manner:

[0058] As Figure 2 shown, in step S210, multiple generator parameters representing the operating state of the wind turbine can be obtained.

[0059] Here, the multiple generator parameters may include the operating parameters and / or energy efficiency parameters of the wind turbine.

[0060] As an example, the operating parameters of the generator can be basic data describing the operation of the generator. For example, it may include wind speed v, active power P, ambient temperature T s , nacelle temperature T c , generator temperature T G , generator speed R, three-phase current of the generator (such as I A , I B and I C three phases) and at least one of the maximum deviation E d of the three-phase current.

[0061] Here, the maximum deviation of the three-phase current can be expressed by the following formula (1):

[0062]

[0063] Where E max is the maximum value of the three-phase current, and the unit can be, for example, ampere (A); is the average value of the three-phase current, and the unit can be, for example, A.

[0064] As an example, the energy efficiency parameter can be a parameter that describes the energy transfer and conversion ability of the generator. For example, for a direct-drive permanent magnet generator, considering the structure and operating characteristics of the generator, the energy efficiency parameter can include the copper loss and / or iron loss of the generator.

[0065] Here, the copper loss of the generator is the power loss generated by the current in the winding, which can be expressed by the following formula (2):

[0066] P Cu =cI 2 u (2)

[0067] Where c is the number of phases; I is the single-phase current of the stator, and the unit can be, for example, A; u is the stator resistance, and the unit can be, for example, ohm (Ω).

[0068] The iron loss of the generator is mainly caused by the molecular friction of ferromagnetic materials and the circulating current in the conductive sheet. For example, it can include hysteresis loss, classical eddy current loss, and additional eddy current loss. The iron loss can be expressed by the following formula (3):

[0069]

[0070] Where K h is the hysteresis loss coefficient; K c is the static eddy current loss coefficient; K e is the additional eddy current loss coefficient; M is the peak magnetic flux density, and the unit can be, for example, tesla (T); f is the magnetic flux frequency, and the unit can be, for example, hertz (Hz).

[0071] In step S220, the correlation coefficient between parameter pairs formed by every two of the multiple generator parameters can be determined.

[0072] Here, the correlation coefficient can represent the correlation between two generator parameters. Specifically, by judging the relationship between generator parameters, the description ability of different parameter pairs for the operating state characteristics of the generator can be evaluated, so that it can be used to screen out the parameters that can effectively describe the operating state of the generator as the first sample data.

[0073] As an example, based on the existing correlation calculation method, the correlation between every two of the multiple generator parameters can be determined by analyzing the parameter characteristics or physical meanings of each generator parameter. The analysis result can be expressed, for example, by Figure 3 where the value range of the correlation coefficient is [-1, 1], and the closer the absolute value of the correlation coefficient is to 1, the stronger the correlation between the two parameters.

[0074] In this step, it is possible to clearly determine whether each parameter to be involved in subsequent analysis meets the relevance requirements. Among them, the mutual relevance of all parameters should not be too strong (or rather, the homogeneity is too high), which will cause some parameters not to play a role in the actual budget and make the operation redundant; nor should all parameters have no mutual relevance (or rather, the heterogeneity is too high), which will make it difficult to interpret and analyze the operation results.

[0075] In step S230, one generator parameter in the parameter pair whose correlation coefficient meets the preset correlation condition can be deleted from multiple generator parameters to obtain the filtered generator parameters.

[0076] Here, the preset correlation condition can be, for example, that the correlation coefficient meets a preset numerical range, and this numerical range can be determined according to actual needs. For example, taking Figure 3 the correlation analysis method shown as an example, the value range of the correlation coefficient is [-1, 1], and the preset numerical range can be within this value range. For example, it can be [-0.6, -0.4] and [0.4, 0.6]. However, it is not limited to this, and other ranges can also be set according to actual needs.

[0077] In this step, if the two parameters in the parameter pair meet the above preset correlation condition, then one parameter in the parameter pair can be removed, and the other parameter in the parameter pair is used as the parameter of the first sample data.

[0078] In step S240, the sample data corresponding to the filtered generator parameters can be used as the first sample data.

[0079] After obtaining the filtered generator parameters, the corresponding sample data can be used as the above first sample data.

[0080] Here, the operating parameters include the three-phase current of the generator, and include multiple parameters among the maximum deviation of the three-phase current, wind speed, active power, ambient temperature, nacelle temperature, generator temperature, and generator speed. Among them, other parameters except the three-phase current of the generator in the operating parameters are used as the filtered generator parameters.

[0081] Specifically, taking Figure 3 the correlation analysis result shown as an example, the correlation coefficient between the three-phase current value of the generator and the maximum deviation of the three-phase current is relatively high, which indicates that the corresponding comparison of these dimensional parameters for the change of the generator state is relatively similar. At the same time, compared with the three-phase current value of the generator, the correlation coefficient between the maximum deviation of the three-phase current and the active power is relatively small. Therefore, through the correlation analysis of the generator operating parameters, the three-phase current data of the generator can be removed, and the maximum deviation of the three-phase current and other data are retained as the parameters describing the generator operating state and used as the first sample data.

[0082] Here, in the case of finding some parameters with relatively high correlation coefficients, if there is a large overlap in the physical properties of these parameters or the physical phenomena they represent, it will affect the operation efficiency. In addition, it may also interfere with the final calculation results. Therefore, when the correlation coefficients between the three-phase current and the maximum deviation of the three-phase current with respect to the active power are relatively close, the three parameters of the three-phase current can be discarded, and the maximum deviation of the three-phase current can be retained.

[0083] Through the above method, based on the correlation between parameters, multiple generator parameters related to the operating state of the wind turbine can be screened to minimize the amount of data in the first sample dataset as much as possible, which is beneficial to improving the operation speed of subsequent algorithms and the fault discrimination efficiency.

[0084] Although the process of determining the first sample data is described by way of example above, it is not limited thereto, and the relevant data of one or more specified unit operating parameters can also be directly used as the first sample data.

[0085] In step S120, based on the data distribution of each first sample data in the first sample dataset, the first sample dataset can be screened to obtain a second sample dataset.

[0086] Here, the data distribution can be used to characterize the data feature distribution among the first sample data and can reflect the overall data situation of the first sample dataset. As an example, the data distribution can be represented by the Mahalanobis distance, but it is not limited thereto, and other methods such as the Euclidean distance and the Manhattan distance can also be used.

[0087] In this step S120, based on the data distribution, the first sample data with a relatively deviated data feature distribution, or in other words, the first sample data whose data features deviate from the overall distribution of the dataset, can be removed. As an example, the first sample dataset can be screened in the following way:

[0088] As Figure 4 shown, in step S410, the distance metric can be performed on each first sample data in the first sample dataset to obtain the distribution distance of each first sample data relative to the sample representative data.

[0089] Here, the distance metric can be, for example, the Mahalanobis distance metric, but it is not limited thereto, and it can also be the Euclidean distance, the Manhattan distance, etc. The sample representative data can characterize the central tendency of the data in the first sample dataset. For example, it can be the data center of gravity of the dataset (or it can also be called the "data center").

[0090] As an example, in this step S410, the distribution distance of each first sample data can be obtained in the following way: As Figure 5As shown, in step S510, the preset metric matrix can be optimized based on each first sample data in the first sample dataset, the preset supremum of the metric distances between the same type of sample data, and the preset infimum of the metric distances between different types of sample data, to obtain an optimized metric matrix; in step S520, the distribution distance can be determined based on the first sample data, the sample representative data, and the optimized metric matrix.

[0091] Here, the metric matrix can be used to determine the metric distance between samples. In this step, the optimization of the metric matrix can be achieved through metric learning.

[0092] The main role of metric learning is to determine the number field category. Taking the Mahalanobis distance as an example, there can be a metric matrix H. When there is no participation of the metric learning algorithm, this matrix represents the covariance matrix of the variables involved in the calculation. In the process of calculating the Mahalanobis distance, it is default that all the digital characteristics involved in the calculation belong to the same number field category (that is, the rules for the same or different categories are the same). However, in practice, for actual data (or sample data), it cannot be strictly guaranteed whether these data characteristics truly belong to one or more categories (usually considered as one category).

[0093] Therefore, under the intervention of metric learning, the "supremum and infimum" of the data for calculating the Mahalanobis distance can be clarified, such as the supremum (that is, clearly defined as the same category) and the infimum (that is, clearly defined as different categories). In most cases, due to objective reasons, it is difficult to traverse all situations of the observation system for the data. The actual data or sample data only partially describes the system state. Therefore, the supremum and infimum can be different values.

[0094] In this step, the preset metric matrix can be optimized through the procedural iteration of the metric algorithm to obtain an optimized metric matrix. For example, the metric matrix can be continuously corrected to make the measured value of the metric matrix converge to reach the optimal metric matrix, so as to ensure that the calculation of the Mahalanobis distance of the data obtained under the given conditions is accurate, that is, this distance objectively describes the relationship between the data.

[0095] Based on the above concept, as an example, in the above step S510, the optimized metric matrix can be obtained in the following way: traverse the sample pairs formed by the first sample data in the first sample dataset, and determine the candidate metric matrix by performing the following steps S511 to S512 until the candidate metric matrix meets the preset iteration condition, and take the candidate metric matrix that meets the preset iteration condition as the optimized metric matrix.

[0096] Here, the preset iteration condition can be, for example, the convergence condition of the metric matrix. For example, it can be iterated until the related function of the metric matrix is a convergence function, then the current metric matrix can be determined as the finally optimized metric matrix.

[0097] Specifically, in step S511, the current metric distance between the current sample pairs can be determined based on any unvisited current sample pair in the first sample dataset and the current metric matrix.

[0098] As an example, taking the Mahalanobis distance as the metric distance, assume that the n-dimensional sample x = (x1, x2,..., x n ) T and y = (y1, y2,..., y n ) T , the Mahalanobis distance between the two samples can be expressed by the following formula (4):

[0099]

[0100] where H ∈ R n×n is the metric matrix, and there can be symmetric positive semi-definiteness. Perform eigenvalue decomposition on it, H = Q V ΛQ, where Λ is the diagonal matrix for stretching and transforming the original digital space; Q V can linearly reconstruct the original digital space, and V is the rank of H.

[0101] The metric matrix H corresponding to the observed samples under different systems is different. Therefore, the metric matrix H reasonably trained from the sample data can improve the discrimination ability of the metric algorithm for the nature of the data itself.

[0102] In step S512, the current metric matrix can be optimized according to the types of the two first sample data in the current sample pair.

[0103] In this step, in response to the candidate metric matrix not satisfying the above preset iteration condition, the candidate metric matrix is used as the current metric matrix for the next traversal. Among them, in the initial traversal process, the current metric matrix is the preset metric matrix, the current supremum is the preset supremum, and the current infimum is the preset infimum.

[0104] Specifically, in response to the two first sample data in the current sample pair being the same type of sample data, the current metric matrix is optimized based on the current metric distance and the current supremum to determine the candidate metric matrix; or, in response to the two first sample data in the current sample pair being different types of sample data, the current metric matrix is optimized based on the current metric distance and the current infimum to determine the candidate metric matrix.

[0105] As an example, assume that the sample dataset is {x k}, and x k ∈R W , k = 1, 2,..., n, where n is the number of sample data in the sample dataset. x i and x jThe Mahalanobis distance can be expressed by the following formula (5):

[0106]

[0107] Describe different types of samples in the form of sample constraint pairs, that is, it is considered that the Mahalanobis distance calculated by the metric matrix H reflects the constraint pair relationship of the samples. According to the convention of the Information-Theoretic Metric Learning (ITML): If x i and x j are sample data of the same type, then their Mahalanobis distance needs to be less than a preset supremum; otherwise, it is greater than a preset infimum. Here, ITML is a method for evaluating the effect of feature selection algorithms (this attributive is also known as the information gain algorithm or mutual information algorithm). By continuously increasing the sample constraint pairs, finally determine H ∈ R w×w to satisfy the following constraint relationship shown in formula (6):

[0108]

[0109] where S is the set of samples of the same type, W is the set of samples of different types, s is the supremum, and w is the infimum.

[0110] Affected by the relevance of each constraint pair, the metric matrix H that satisfies the constraint conditions is not unique. To ensure the stability of the algorithm, the metric matrix H can be normalized to form a preset matrix H0, and the distance between the two can be represented by the Bregman divergence:

[0111]

[0112] where tr(H) is the trace of the metric matrix H, and φ(H) is a strictly convex differentiable function. The characteristics of different convex differentiable functions are different. Here, φ(H) is taken as Log(det(H)), that is, the Bregman divergence is converted into the LogDet divergence. It is known that the value of the LogDet divergence is constant when performing an invertible linear transformation K:

[0113] D LD (H, H0) = D LD (K T HK, K T H0K) (8)

[0114] At this time, the metric learning is converted into the optimization problem of LogDet, that is and satisfy the following constraint relationship:

[0115]

[0116] The metric matrix H constrained by the samples can be calculated by using, for example, the gradient descent iterative method to ensure the existence of a feasible solution. As an example, the metric matrix can be optimized through the following iterative steps S1 to S9:

[0117] Step S1: Initialize the relevant data, let H t = H0, λ(i,j) t = 0;

[0118] Step S2: For the sample (i,j) ∈ x k , when (i,j) t ∈ S, m(i,j) t = s, l(i,j) t = 1; when the sample pair (i,j) t ∈ W, m(i,j) t = w, l(i,j) t = -1;

[0119] Step S3:

[0120] Step S4:

[0121] Step S5:

[0122] Step S6:

[0123] Step S7:

[0124] Step S8: H t+1 = H t + μH t (x i - x j )(x i - x j ) T H t ;

[0125] Step S9: Repeat the process of (2) to (8) for the sample pairs of the rows until convergence.

[0126] In the above process, β is a slack variable used to balance the objective function and the linear constraints. By iteratively calculating by successively adding sample pairs until the objective function converges, the final metric matrix H is obtained.

[0127] In some lightweight algorithms that can be used for fault discrimination, the more observation vectors of the sample data, the better the algorithm's performance. However, too many observation vectors may weaken the algorithm's sensitivity to noise; while the fewer the observation vectors of the sample data, the higher the computing speed, but it will affect the accuracy of the final result. Taking the non-linear state estimation technique (NSET) as an example, NSET is an analysis method for capturing the characteristics of non-linear data (this attributive is also known as non-linear estimation or non-linear calculation). In the process of building a model using the NSET method, as a model sample, the construction of the process memory matrix is the key to ensuring the accuracy of the algorithm results. Theoretically, the more state vectors in the process memory matrix, the better the model analysis effect, but too many state vectors will weaken the sensitivity of the sample to noise. At the same time, the model operation efficiency will also decrease significantly; on the contrary, if the number of state vectors in the process memory matrix is too small, although the model operation speed is increased, the result accuracy will be greatly reduced due to poor data correlation.

[0128] Regarding the above problems, according to the embodiments of the present disclosure, the model operation efficiency and operation accuracy can be balanced, and the information theory metric learning algorithm can optimize the calculation of the metric distance such as the Mahalanobis distance of the sample data in the digital space, and improve the rationality of the construction of the observation matrix such as the process memory matrix, so as to realize the improvement of the lightweight algorithm such as the NSET method.

[0129] In step S420, based on the distribution distance of each first sample data and a preset distance screening condition, the first sample data set can be screened to obtain a second sample data set.

[0130] Here, the first sample data near the center of the sample set can be determined according to the distribution distance of each first sample data, and a certain distance range can be determined as the distance screening condition, and the first sample data outside this distance range can be removed, so as to realize the screening of the data.

[0131] As an example, in this step S420, the distance screening condition can be determined in the following way:

[0132] As Figure 6 shown, in step S610, based on the distribution distances of all the first sample data in the first sample data set, the minimum distance and the maximum distance in the distribution distances can be determined.

[0133] In step S620, based on the minimum distance and the maximum distance, the screening step size can be determined.

[0134] In step S630, based on the screening step size and the minimum distance, the upper limit distance of the distribution distance of the first sample data can be determined.

[0135] In step S640, a distance screening condition can be determined based on the upper limit distance, where the distance screening condition indicates screening out the first sample data with a distribution distance less than or equal to the upper limit distance from the first sample dataset.

[0136] As an example, combined with Figure 7 , the original normal data of the unit (i.e., the first sample data) X n×m can be divided into u working conditions and normalized. This is only an example, and it can also be divided based on other parameters such as power. Perform ITML iterative analysis on the data of each working condition, that is, the process of obtaining the optimized metric matrix described in step S510 above, so as to determine the optimal metric matrix of the data vector under each working condition, denoted as {H k}, k = 1, 2,..., u. The Mahalanobis distance between the vector X k (e) and the centroid of the data population X k can be expressed by the following formula (10):

[0137]

[0138] where H k is the optimal metric matrix under working condition k, g k is the centroid (which can also be called the "center") of X k , and the centroid g k can be expressed by the following formula (11):

[0139]

[0140] where is the value of the nth row and the ith column in the kth working condition data population, c is the number of vectors in this working condition, and c ≤ m.

[0141] The Mahalanobis distances under working condition k can be determined, denoted as M k . Sort all the data vectors under this working condition according to the Mahalanobis distance with a certain step size h k . The expression of h k is as follows:

[0142] h k = [max(d k ) - min(d k )] / z (12)

[0143] where z is the number of steps. Here, the step size h k and the number of steps z can be determined according to actual needs.

[0144] As can be seen from the definition of Mahalanobis distance: the smaller the Mahalanobis distance from a data vector to the population, the closer the vector is to the overall characteristics of the data. Therefore, data vectors that are relatively close to the data population under each working condition can be selected for subsequent operations. For example, the upper limit of the distance of the data vector under working condition k can be selected as dm k , which can be expressed by the following formula (13):

[0145] dm k = min(d k ) + jh k (13)

[0146] In this formula (13), the coefficient j of the step size h k can be set according to actual needs, for example, it can be 5. In this regard, taking a two-dimensional point set as an example, after the metric learning process, in the original first sample data set, points close to the center of the data set will gather towards the center, and points far from the data set will diverge towards the surroundings. Therefore, under working condition k, the value of max(d k ) - min(d k ) will be larger than the value without metric learning. In this case, on the premise of determining the step size h k , the value of the step size h k will also increase. In addition, after metric learning, the distribution form of the data set changes. When screening data, more data that are close to the center will also be retained. After these processes, the data with a distance greater than the coefficient "5" of the step size h k is actually relatively small. Therefore, here j = 5 can be selected.

[0147] In addition, although the upper limit of the distance for screening data is given above with reference to formula (13), it is not limited to this. Points far from the center of the data set can also be excluded by other means. For example, to ensure the completeness of data feature description, within the determined wind speed range, on the basis that the actual Mahalanobis distance is greater than the upper limit of the distance dm k , the first sample data with a metric distance in the top 30% at this step size distance can also be further retained.

[0148] Based on the above formulas (10) to (13), the selection of the observation vectors under all working conditions can be determined to obtain the second sample data set for subsequent calculations. In this way, data deviating from the center of the data set can be quickly and simply excluded by distance screening, so as to obtain a relatively concentrated second sample data set that can better represent the overall characteristics of the data set.

[0149] Generally speaking, through the above method, the first sample data set can be reconstructed to obtain the second sample data set, so that the data features of the second sample data in the reconstructed second sample data set are sharpened, and the disturbance terms deviating from the sample center are removed, which is beneficial to improving the subsequent operation speed.

[0150] In step S130, the second sample data set can be compared with the reference data set to obtain a fault discrimination condition for discriminating the faults of the wind turbine.

[0151] Here, the reference data set can include data characterizing the wind turbine in a non-fault state. The reference data set can include labeled verification data, and the verification data can include data in the generator fault state and data in the generator non-fault state, and each data has been labeled with the corresponding generator state, that is, the fault state or the non-fault state. These verification data can be used to compare with the second sample data set, so that the algorithm can learn the data features of the wind turbine when various faults occur.

[0152] The fault discrimination condition can be, for example, a fault threshold determined after comparing the second sample data set and the reference data set. For example, the NSET algorithm can be used as the fault discrimination algorithm to determine the fault threshold.

[0153] As an example, in this step S130, the fault discrimination condition can be obtained in the following way: as Figure 8 shown, in step S810, the second sample data set and the reference data set can be input into the non-linear state estimation model to compare the data features of the second sample data set and the reference data set to obtain a residual value; in step S820, based on the residual value, a fault discrimination condition can be obtained.

[0154] In the above steps, the non-linear state estimation model can be the NSET algorithm model. As Figure 7 shown, the NSET algorithm model can obtain a residual value based on the input second sample data set and reference data set. The fault discrimination condition can be determined based on this residual value. The NSET algorithm model will be described exemplarily below.

[0155] Specifically, assume that there are n data for describing the operating state of the wind turbine generator set, and the data vector obtained at time i is the observation vector X(i). By collecting the observation vectors of the unit in the normal state for a period of time, a process memory matrix can be formed:

[0156]

[0157] Through reasonable screening, the process memory matrix can describe the motion process of the unit in the normal state. Input the observation vector X into D obs(For example, it can be the second sample data set), the prediction vector X can be obtained est (For example, it can be the reference data set), where X obs and X est can have a linear relationship as shown in the following formula (15):

[0158] X est = D·W = w1X(1)+w2X(2)+…+w m X(m) (15)

[0159] where, W = [w1,w2,…,w n T is the weight vector. W is the independent variable of X obs and X est and the residual ε. By taking the partial derivative of the dot product of ε and performing extreme value calculation, the following formula (16) is obtained:

[0160]

[0161] where, is a general matrix operation. Here, for example, it can represent the Euclidean distance operation, which can be expressed by the following formula (17):

[0162]

[0163] Substituting formula (16) into formula (15), the prediction vector X est can be obtained, which can be expressed by the following formula (18):

[0164]

[0165] When the unit is in the normal state, the similarity between the observation vector and the vector in the process memory matrix is relatively high, the characteristics of the prediction vector are relatively close to those of the observation vector, that is, the residual value ε is small. On the contrary, when the unit is in the abnormal state, the similarity between the observation vector and the vector characteristics in the process memory matrix is low, the characteristics of the prediction vector are significantly different from those of the observation vector, and the residual value ε will become larger.

[0166] Here, by artificially setting or calculating the residual threshold and comparing the size relationship between the actual residual value and the threshold, it is judged whether there is a fault in the unit equipment. For example, the residual value calculated by the NSET algorithm model can be used as the residual threshold.

[0167] By determining the fault discrimination condition in the above manner, different fault discrimination conditions can be determined according to different units, so as to make the fault discrimination more accurate.

[0168] In step S140, based on the operation data of the wind turbine and the fault discrimination condition, the wind turbine can be subjected to fault discrimination. ​

[0169] In this step, the operating data can be the operating data of the wind turbine during the period to be discriminated. Based on the fault discrimination conditions obtained above, the operating data can be judged to evaluate the state of the wind turbine. For example, as Figure 7 shown, the operating data can also be referred to as analysis data.

[0170] As an example, the operating data can include a first operating data set. The following method can be used to perform fault discrimination on the wind turbine:

[0171] As Figure 9 shown, in step S910, based on the first operating data set of the wind turbine, the second sample data in the second sample data set can be weighted to obtain a third sample data set.

[0172] Specifically, as described above, the second sample data set can be obtained based on the first sample data set divided by bins such as wind speed or power. Therefore, the binning method of the second sample data set can be the same as that of the first sample data set. For example, as Figure 7 shown, when the first sample data set is binned by wind speed segments, the second sample data set is also binned by wind speed segments.

[0173] In this case, in this step S910, the second sample data set can be weighted according to the actual situation of the first operating data set to be discriminated for fault, and a third sample data set containing the data characteristics of the first operating data set is obtained. Here, the first operating data set can be subjected to data normalization processing, and then segmented by time series to perform time series analysis of its data characteristics.

[0174] As an example, the second sample data set can be weighted according to time series. Specifically, the third sample data set can be obtained in the following way: analyze the data characteristics of the first operating data set according to time series, determine the weights of the second sample data corresponding to each first operating data; based on the weights, weight the second sample data in the second sample data set to obtain a third sample data set.

[0175] For example, the weights of the second sample data can be determined in the following way: perform time series arrangement on the second sample data in the second sample data set to obtain a sample time series; perform time series segmentation on the first operating data in the first operating data set to obtain multiple operating time series, where each operating time series includes at least one first operating data; determine the sequence weight of each operating time series according to the data characteristics of each operating time series; based on the sequence weights, determine the weights of the second sample data corresponding to each operating time series in the sample time series.

[0176] As an example, as Figure 7As shown, the locally weighted dynamic time wrapping (LDTW) algorithm can be used to analyze the data characteristics of the first running dataset for time series weighting. Here, LDTW is an algorithm for time series matching and alignment (this attributive can also be called weighted dynamic time warping).

[0177] In the LDTW algorithm, when assigning values to the time series, it can be assumed that the overall sample time series of the first running dataset is X, and the running time series X of the first running dataset is obtained test ={x1,…x q} T . X test has q subsequences, and the lengths T of each sequence may not be the same. First, taking X as the reference sequence, DTW calculations can be performed on each time subsequence of X text to obtain the standard alignment arrangement and alignment set of the reference sequence with the data at different moments of each subsequence. Then, the positive example set cNN + and the negative example set cNN - of X can be calculated. The two sets are respectively composed of c time series of the same class and different classes that are closest to X. Each segment of x i forms a weight vector W it for the sample data X, and this weight vector can be expressed by the following formula (19):

[0178]

[0179] where A it can be expressed by the following formula (20):

[0180]

[0181] where π ii' represents the standard alignment arrangement of the i-th analysis sequence and the sample sequence under the standard DTW calculation, and x it represents the alignment set of x i at time t; q is the number of subsequences; and represent the positive example set and negative example set of x i ; ξ is a very small value defined to prevent the denominator from being 0 and can be set according to actual needs.

[0182] Based on the above formulas (19) and (20), all subsequence weight calculations can be completed for X text to obtain the weight (or weights) set for weighting the second sample dataset. Each second sample data can be weighted based on these weights to obtain the third sample dataset.

[0183] Here, for the time-series weighting process, after analyzing through the time-series analysis method, a weight value is multiplied to each dimension data of the data within the calculation time period. If two sets of data are exactly the same, then the weight value of each dimension is the same, that is, it exists in a form of proportional scaling with the sample data. If there are significant differences between the two sets of data, the weight value on one or several dimensions will be amplified to increase the difference in features between the actual data and the sample data (for example, for the same state under different excitation conditions, the dominant degree of different dimension data features for this state is different). The whole process is actually to reconcile the instability of lightweight algorithms such as NSET itself and the "sluggish" change of results caused by the overly large sample volume.

[0184] In lightweight algorithms such as the NSET method, the calculation method of pairwise comparison is adopted during data analysis, which may lack the correlation analysis of sample data for the entire time-series data. According to the embodiments of the present disclosure, by adopting a time-series analysis method such as the LDTW method, for analyzing each time series, corresponding weight values of subsequences can be assigned to each dimension data of the overall first running data set X with respect to the second sample data set X text Thereby, algorithms such as NSET can be improved, and the perception ability of the second sample data set for different time-series data features during the calculation process of the algorithm can be enhanced. Such a time-series analysis method assigns different weight values to each data point of the input time series through the behavioral differences of intra-class and inter-class data characteristics, reduces the intra-class time-series metric value, and simultaneously increases the inter-class time-series metric value, improving the data classification accuracy.

[0185] At the underlying logic of the calculation, the actual data is input into the analysis model one by one at each moment for calculation, but in fact, the data has a time dimension. Here, introducing time-series analysis is to perform feature mapping between the actual data and the sample data in the time dimension, that is, to adjust the weight values of the actual data of each dimension within a period of time to strengthen the perception ability of the algorithm for the features of the actual data. In actual calculation, time-series analysis calculation is performed on the sample data of each wind speed section and the actual data of the same time section.

[0186] In step S920, the second sample data set and the third sample data set can be compared to obtain the fault parameter.

[0187] For example, as Figure 7As shown, after obtaining the third sample data set, the second sample data set and the third sample data set can be input into a lightweight algorithm such as NSET to compare the two, and a residual result that can be used as a fault parameter is obtained. The calculation process taking NSET as an example has been described above with reference to Equations (14) to (18). Here, Equations (14) to (18) can be similarly applied to compare the second sample data set and the third sample data set to calculate the fault parameter.

[0188] Although described herein by taking NSET as an example, the embodiments of the present disclosure are not limited thereto, and other algorithms can also be used. The calculation of the fault parameter here can adopt the same method as the calculation of the fault threshold in the fault discrimination condition described above, so as to facilitate fault discrimination.

[0189] In step S930, the wind turbine can be subjected to fault discrimination based on the fault parameter and the fault discrimination condition.

[0190] As an example, the wind turbine can be subjected to fault discrimination based on the fault parameter and the fault discrimination condition in the following manner: in response to the fault parameter satisfying the first fault value range, it is determined that the wind turbine is in a normal state; in response to the fault parameter satisfying the second fault value range, it is determined that the wind turbine has a magnet shedding fault; in response to the fault parameter satisfying the third fault value range, it is determined that the wind turbine has a three-phase current imbalance fault.

[0191] Here, the upper limit value of the first fault value range can be less than the lower limit value of the second fault value range, and the upper limit value of the second fault value range can be less than the lower limit value of the third fault value range.

[0192] As an example, when the fault parameter is a residual value, the first fault value range can be [0, 0.24], the second fault value range can be [0.3, 0.40], and the third fault value range can be [0.50, 0.65].

[0193] According to the method of the embodiments of the present disclosure, it is not only possible to determine whether the wind turbine is abnormal, but also possible to further analyze the cause of the fault, which is beneficial to quickly troubleshooting possible faults and performing maintenance in a timely manner.

[0194] Through the above-described fault discrimination method, the correlation of data in the time dimension can be strengthened, the time influence factor can be introduced, and it is also possible to allow a difference between the length of the actual operation data (or analysis data) and the length of the sample data, without being strictly the same.

[0195] Specifically, when dividing the sample data into wind speed segments, the relationship between wind speed changes and time is actually implied, that is, it is assumed that there is a strong correlation between wind speed changes and time. In this case, using time series data and wind speed segment sequences for time series analysis and calculation can, on the one hand, strengthen the correlation of data in the time dimension; on the other hand, it does not require a strict match between the actual operating data and the sample data length. This is because in actual calculations, it is difficult to ensure that the data length within each determined calculation cycle is the same. Time series analysis such as the LDTW algorithm can achieve matching between data features of different time lengths. Therefore, fault identification based on the above method can improve the accuracy of fault identification and reduce the requirements for input data.

[0196] The above describes the fault determination method of the wind turbine generator according to the embodiment of the present disclosure. The application of the fault determination method will be described below in conjunction with a specific example.

[0197] Taking a 6.0MW direct-drive wind turbine unit in an offshore wind farm as an example, the generator status of the unit was analyzed using the NSET method improved by the fault identification method according to the embodiment of the present disclosure and the original NSET method.

[0198] It is known that the generator magnet of unit 11 of the wind farm fell off on July 15, 2022. The operation data of the unit from May 1, 2022 to July 20, 2022 are collected, and the communication frequency is 10s. Among them, May 1, 2022 to June 30, 2022 can be used as the original sample data (the first sample data set mentioned above), and July 1, 2022 to July 20, 2022 can be used as the analysis data (the first operation data set mentioned above).

[0199] The original sample data was preprocessed to remove interference items such as shutdown, power overrun and power limit. The sample data was divided into working conditions with a wind speed of 2m / s as the step size, and the data distribution under each sub-working condition of Unit 11 was obtained as follows: Figure 10 shown.

[0200] In the improved method, the supremum of the Mahalanobis distance between samples of the same type is s=1, and the infimum of the Mahalanobis distance between different samples is w=10. To avoid overfitting, the slack variable β=10 -2 , ITML is used to calculate the metric matrix of each wind speed segment for the original data, and the change trend of the objective function of each wind speed segment is as follows Figure 11 shown.

[0201] Depend on Figure 11It can be seen that the objective function values for each wind speed segment start to converge after a finite number of iterations, indicating the existence of an optimal metric matrix H. Taking the overall wind speed data optimized by the metric matrix as a sample and using the wind speed as the partitioning variable, the factor analysis method is adopted to project the overall data onto a two-dimensional digital space. The results are as Figure 12 and Figure 13 shown, where Figure 12 is the distribution of the sample data for each wind speed segment before data optimization, Figure 13 is the distribution of the sample data for each wind speed segment after data optimization. Here, factor analysis can reasonably separate each point set according to the correlation analysis of the influence degree of data characteristic factors, and at the same time delete some data that has little relationship with the characteristics of the point set, improving the data representation ability.

[0202] From the comparison between Figure 12 and Figure 13 , it can be seen that the data points of the original data are more compactly distributed in the digital space, the boundaries of the data point sets for each wind speed segment are relatively blurred, and there is obvious overlap of the point sets in some intervals; the distances between the data point sets for each wind speed segment after optimization by metric learning are significantly increased in the digital space, and the boundaries of each point set are relatively clear. Thus, it can be seen that through metric learning, the data of the generator set has enhanced the data characteristics of each wind speed segment, and at the same time provided a better analysis environment for the reconstruction of the process memory matrix.

[0203] The process memory matrix is constructed for the data of each wind speed segment using the improved method and the original method respectively. The Mahalanobis distance distributions under each wind speed condition are as Figure 14 and Figure 15 shown, where Figure 14 is the distribution of the Mahalanobis distance under each wind speed condition calculated using the improved method, Figure 15 is the distribution of the Mahalanobis distance under each wind speed condition calculated using the original method.

[0204] From the comparison between Figure 14 and Figure 15 , it can be seen that Figure 14 the data distribution is more uniform than Figure 15 . In Figure 14 , the number of data vectors distributed at steps 1, 2, 8, 9, and 10 is significantly higher than that in Figure 15 . This shows that when calculating the Mahalanobis distance using the metric matrix optimized by metric learning, the metric matrix "pulls closer" the data vectors that are relatively close to the overall distance data and "pushes away" the data vectors that are relatively far from the overall distance data. In terms of data characteristics, it strengthens the characteristic attributes representing the overall sample data and weakens the interference of data with weak characteristics on the overall characteristics of the sample.

[0205] After completing the sample reconstruction, perform a time series correlation analysis between the unit analysis data and the sample data of the reconstructed process memory matrix. Segment the analysis data on a daily basis and input it into the DTW model with the reconstructed data as the benchmark. The results are as Figure 16 shown. In Figure 16 , the abscissa represents the sample data taking values along the time series, where 0 to 50 indicates the number of times of taking values, and the ordinate represents the analysis data taking values along the time series, where 0 to 50 indicates the number of times of taking values. The curve in this figure represents the path. Under normal circumstances, when the characteristics of the analysis data are similar to those of the sample data, the path is a relatively straight diagonal line. If the characteristics of the two are quite different, the path will become tortuous and deviate significantly from the diagonal path. Through this Figure 16 , it can be seen that the LDTW method can demonstrate the intervention ability of data in the time dimension.

[0206] In Figure 16 , the baseline is the theoretical path under the normal state of the unit. Among them, the path of the unit in the normal state fluctuates less near the baseline. The other lines are the paths of the unit generator with faults, and these paths all deviate significantly from the baseline, indicating that there are certain differences between the characteristics of the fault time series data and the sample. From the expression of the time series weight vector, it can be known that when there are obvious differences in characteristics between the actual data and the sample data, the process of weight assignment will be relatively "aggressive", that is, the standard deviation of the weights assigned to each dimension data in the time series will be significantly greater than that of the normal data. After LDTW calculation, the standard deviations of the weights of each dimension data in different time series of this unit from July 1, 2022 to July 20, 2022 are shown in Table 1 below:

[0207] Table 1

[0208] Time period Standard deviation Time period Standard deviation Time period Standard deviation Time period Standard deviation Day1 0.0151 Day6 0.0258 Day11 0.0411 Day16 0.0684 Day2 0.0170 Day7 0.0269 Day12 0.0429 Day17 0.0714 Day3 0.0178 Day8 0.0274 Day13 0.0485 Day18 0.0798 Day4 0.0184 Day9 0.0354 Day14 0.0541 Day19 0.0821 Day5 0.0235 Day10 0.0399 Day15 0.0570 Day20 0.0825

[0209] As can be seen from Table 1 above, as time goes by, the standard deviation of the weights of each dimension data in the time series shows an obvious increasing trend, indicating that the difference between the current operation data of the unit and the characteristics of the normal data is gradually becoming obvious, the difference in the weights of each dimension data in the time series increases, that is, the current unit is moving away from the normal operation state.

[0210] Take the normal data of Unit 9 in the same wind farm from May 1, 2022 to June 30, 2022 as the verification data, perform NSET analysis on the sample data of the two methods, and obtain the trend of the verification residuals of the sample data as Figure 17 shown.

[0211] From Figure 17It can be seen that the state residual curves of both methods fluctuate within a certain numerical range, indicating that the improved algorithm can accurately represent the normal operating state of the unit. At the same time, the fluctuation degree of the residual curve obtained by the improved algorithm is significantly smaller than that of the original algorithm. This shows that for the representation of the operating characteristics of the generator, the overall characteristics of the sample data of the improved algorithm are more completely represented and the anti-data interference ability is stronger. Affected by the uncertainty in the process of nonlinear model analysis, twice the maximum value of the residual is taken as the threshold under the normal state of the unit. Among them, the residual threshold of the improved method is 0.24, and the residual threshold of the original method is 0.35.

[0212] Apply the two methods to perform NSET analysis on Unit 11 of the wind farm, and obtain the generator state residual trends before and after sample data optimization as Figure 18 and Figure 19 shown, where Figure 18 is the generator state residual trend calculated by the improved method, Figure 19 is the generator state residual trend calculated by the original method.

[0213] From Figure 18 and Figure 19 comparison, it can be seen that Figure 18 's residual value curve exceeds the threshold on July 5, 2022, Figure 19 's residual curve exceeds the threshold on July 15, 2022. At the same time, the rising trend of the residual curve obtained by the improved algorithm is obvious, and the data fluctuation is small; the trend change of the residual curve of the original algorithm is slow, the data fluctuation is large, and the residual curve shows a "repeated jump" near the threshold. This shows that compared with the original algorithm, the improved algorithm can detect the change of the generator state of the unit earlier and make an accurate judgment.

[0214] Here, taking the generator of an offshore direct-drive wind turbine as the research object and using SCADA data as the information source, the operation of the generator is comprehensively described from the perspectives of operation and energy efficiency. Using NSET based on nonlinear modeling as the fault discrimination method ensures the interpretability of the analysis results; aiming at the problem of memory matrix reconstruction in the NSET process, a metric learning method is introduced to improve it. By obtaining the optimal metric matrix, the characteristics of the sample data are improved and the volume of the sample data is reduced; the LDTW method is used to assign sample weights through the correlation analysis of the same and different types of sample data and analysis data at the time series level, improving the traditional NSET's perception ability of data in the time series. Through the analysis of a 6.0MW unit generator fault example, compared with the original algorithm, the improved algorithm can detect the generator fault earlier and more accurately, proving the effectiveness of the improved method.

[0215] The fault discrimination method of a wind turbine according to an embodiment of the present disclosure can improve the quality of sample data without increasing the volume of sample data, so as to meet the requirements of lightweight algorithms such as NSET and improve the operation efficiency.

[0216] In addition, in the related art, lightweight algorithms such as NSET are difficult to process complex time series with different intra-class global features and similar inter-class global features. On the one hand, it is impossible to reasonably screen samples, and it is impossible to ensure the accuracy of NSET results through the reconstruction of the process memory matrix; on the other hand, the traditional algorithm adopts a point-by-point analysis strategy, which is very insensitive to the timing characteristics of data and ignores the correlation analysis of the same or different time series. In response to this, the fault discrimination method of a wind turbine according to an embodiment of the present disclosure can reasonably screen sample data, reconstruct the process memory matrix, ensure the accuracy of the operation result, and can also introduce the timing characteristics of data and ignore the correlation analysis of the same or different time series.

[0217] Specifically, the fault discrimination method of a wind turbine according to an embodiment of the present disclosure can introduce an information theory metric learning algorithm for the reconstruction of the process memory matrix of NSET, perform correlation iterative analysis on the original sample data in the digital space, find the optimal metric matrix, optimize the distance distribution characteristics between data, improve the data characteristics of the reconstructed process memory matrix, and reduce the volume of sample data. In addition, the fault discrimination method of a wind turbine according to an embodiment of the present disclosure can adopt a locally weighted dynamic time warping method to analyze the correlation between the sample data and the actual data in terms of timing, assign weights to each time series time point, and improve the perception and recognition ability of NSET for the timing characteristics of the actual data.

[0218] According to a second aspect of an embodiment of the present disclosure, a computer device is provided. Specifically, the above-mentioned fault discrimination method of a wind turbine can be executed by the computer device, and the computer device includes: at least one processor; at least one memory storing computer-executable instructions, wherein when the computer-executable instructions are run by the at least one processor, the at least one processor is caused to execute the fault discrimination method of a wind turbine according to the exemplary embodiment of the present disclosure.

[0219] As an example, the computer device can be a PC computer, a tablet device, a personal digital assistant, a smart phone, or other devices capable of executing the above instruction set. Here, the computer device does not have to be a single electronic device, but can also be any assembly of devices or circuits that can execute the above instructions (or instruction sets) alone or jointly. The computer device can also be a part of an integrated control system or a system manager, or can be configured as a portable electronic device that is interconnected with a local or remote (e.g., via wireless transmission) interface.

[0220] In a computer device, a processor may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, the processor may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, and the like.

[0221] The processor may execute instructions or code stored in a memory, where the memory may also store data. The instructions and data may also be sent and received over a network via a network interface device, where the network interface device may employ any known transmission protocol.

[0222] The memory may be integrated with the processor, for example, by arranging RAM or flash memory within an integrated circuit microprocessor or the like. Additionally, the memory may include stand-alone devices such as external disk drives, storage arrays, or other storage devices usable by any database system. The memory and the processor may be operatively coupled or may communicate with each other, for example, via I / O ports, network connections, etc., such that the processor can read files stored in the memory.

[0223] Furthermore, the computer device may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.). All components of the computer device may be connected to each other via a bus and / or a network.

[0224] According to a third aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided. Specifically, the fault discrimination method of a wind turbine according to the embodiments of the present disclosure can be written as a computer program and stored on a computer-readable storage medium. When the instructions in the computer-readable storage medium are run by at least one processor, at least one processor is caused to execute the fault discrimination method of the wind turbine according to the exemplary embodiments of the present disclosure. Examples of computer-readable storage media include: read-only memory (ROM), programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disc memory, hard disk drive (HDD), solid state drive (SSD), card memory (such as, multimedia card, secure digital (SD) card or extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer such that the processor or computer can execute the computer program. In one example, the computer program and any associated data, data files, and data structures are distributed across a networked computer system such that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner by one or more processors or computers.

[0225] The specific embodiments of the present disclosure have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that these embodiments can be modified and varied without departing from the principles and spirit of the present disclosure as defined by the claims and their equivalents, and these modifications and variations should also be within the protection scope of the claims of the present disclosure.

Claims

1. A fault discrimination method for a wind turbine, characterized in that, The fault discrimination method: Obtain a first sample data set of the wind turbine of the wind power generation unit to be discriminated; Based on the data distribution of each first sample data in the first sample data set, screen the first sample data set to obtain a second sample data set; Compare the second sample data set with a reference data set to obtain a fault discrimination condition for discriminating the faults of the wind turbine, wherein the reference data set includes data representing the wind turbine in a non-fault state; Based on the operation data of the wind turbine and the fault discrimination condition, discriminate the faults of the wind turbine.

2. The fault discrimination method according to claim 1, characterized in that Determine the first sample data in the first sample data set by the following method: Obtain a plurality of generator parameters characterizing the operation state of the wind turbine, wherein the plurality of generator parameters include the operation parameters and / or energy efficiency parameters of the wind power generation unit; Determine the correlation coefficient between each pair of generator parameters formed by every two of the plurality of generator parameters, wherein the correlation coefficient represents the correlation between two generator parameters; Delete one generator parameter in the parameter pair whose correlation coefficient satisfies a preset correlation condition from the plurality of generator parameters to obtain the screened generator parameters; Use the sample data corresponding to the screened generator parameters as the first sample data.

3. The fault discrimination method according to claim 2, wherein, The operation parameters include the three-phase current of the generator, and include a plurality of parameters among the maximum deviation of the three-phase current, wind speed, active power, ambient temperature, nacelle temperature, generator temperature, and generator speed, wherein: Use the other parameters except the three-phase current of the generator in the operation parameters as the screened generator parameters.

4. The fault discrimination method according to claim 1, wherein Screen the first sample data set by the following method: Perform distance measurement on each first sample data in the first sample data set to obtain the distribution distance of each first sample data relative to the sample representative data, wherein the sample representative data characterizes the central tendency of the data set of the first sample data set; Based on the distribution distance of each first sample data and a preset distance screening condition, screen the first sample data set to obtain the second sample data set.

5. The fault discrimination method according to claim 4, wherein Obtain the distribution distance of each first sample data by the following method: Based on each first sample data in the first sample data set, the preset supremum of the measurement distances between the same type of sample data, and the preset infimum of the measurement distances between different types of sample data, optimize a preset measurement matrix to obtain an optimized measurement matrix, wherein the measurement matrix is used to determine the measurement distance between samples; Based on the first sample data, the sample representative data, and the optimized measurement matrix, determine the distribution distance.

6. The fault discrimination method according to claim 4, wherein, Determine the distance screening condition by the following method: Based on the distribution distances of all the first sample data in the first sample data set, determine the minimum distance and the maximum distance in the distribution distances; Based on the minimum distance and the maximum distance, determine the screening step size; Based on the screening step size and the minimum distance, determine the upper limit distance of the distribution distance of the first sample data. Based on the upper limit distance, determine the distance screening condition, where the distance screening condition represents screening out the first sample data with a distribution distance less than or equal to the upper limit distance from the first sample data set.

7. The fault discrimination method according to any one of claims 1 to 6, characterized in that, Obtain the fault discrimination condition in the following manner: Input the second sample data set and the reference data set into a non-linear state estimation model, compare the data characteristics of the second sample data set and the reference data set, and obtain a residual value; Based on the residual value, obtain the fault discrimination condition.

8. The fault discrimination method according to any one of claims 1 to 6, characterized in that The operating data includes a first operating data set, and the fault discrimination of the wind turbine is performed in the following manner: Based on the first operating data set of the wind turbine, weight the second sample data in the second sample data set to obtain a third sample data set; Compare the second sample data set and the third sample data set to obtain a fault parameter; Based on the fault parameter and the fault discrimination condition, perform fault discrimination on the wind turbine.

9. The fault discrimination method according to claim 8, wherein Obtain the third sample data set in the following manner: Analyze the data characteristics of the first operating data set according to time sequence, and determine the weights of the second sample data corresponding to each first operating data; Based on the weights, weight the second sample data in the second sample data set to obtain the third sample data set.

10. The fault discrimination method according to claim 8, wherein The performing fault discrimination on the wind turbine based on the fault parameter and the fault discrimination condition includes: In response to the fault parameter satisfying the first fault value range, determine that the state of the wind turbine is normal; In response to the fault parameter satisfying the second fault value range, determine that the wind turbine has a magnet steel shedding fault; In response to the fault parameter satisfying the third fault value range, determine that the wind turbine has a three-phase current imbalance fault, where the upper limit value of the first fault value range is less than the lower limit value of the second fault value range, and the upper limit value of the second fault value range is less than the lower limit value of the third fault value range.

11. A computer device, characterized in that, Includes: At least one processor; At least one memory storing computer-executable instructions, where when the computer-executable instructions are run by the at least one processor, the at least one processor is caused to execute the fault discrimination method of the wind turbine according to any one of claims 1-10.

12. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are run by at least one processor, the at least one processor is caused to execute the fault discrimination method of the wind turbine according to any one of claims 1-10.

Images