An abnormal online detection method and system for wind turbine generators using only normal samples

Abstract

The present invention discloses a method and system for abnormal online detection of wind turbine generators using only normal samples, belonging to the technical field of wind turbine generator fault detection. Specifically, it couples a multi-level quantitative Poincaré map and support vector data description. The method includes the following steps: collecting historical and online operation vibration signals of the core components of the wind turbine generator; using the multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the component vibration signals; constructing a hypersphere space of SVDD based on the characteristics of normal samples to obtain an abnormal detection decision domain; using this hypersphere space to detect the online measured signals to determine whether the generator set has an abnormal state and issue an alarm. The present invention provides a method that can accurately and quickly detect the abnormal conditions of wind turbine generators and issue an alarm, which can be simply and economically realized, provides strong guidance for timely discovering the abnormalities of wind turbine generators, and ensures the reliable and stable operation of wind turbine generators.

Description

Technical Field

[0001] The present invention belongs to the technical field of on-line fault detection of wind turbines. More specifically, it relates to a method and system for on-line abnormal detection of wind turbines using only normal samples, specifically coupling a multi-level quantitative Poincaré map and SVDD. Background Art

[0002] As an effective means to reduce the impact of fossil energy on climate change, wind energy plays an important role in the global energy structure. However, with the large-scale construction, operation, and production of wind turbines, a series of new problems have emerged. Accidental faults and shutdowns will lead to a significant increase in the operation and maintenance costs of wind turbines. To solve this problem, effective condition monitoring of wind turbines is extremely urgent.

[0003] In recent years, extracting fault information from vibration signals to reflect the performance degradation of the core components of wind turbines and combining the powerful learning ability of machine learning to detect early anomalies have become the focus of research. However, most wind turbine abnormal detection methods generally consider normal samples and fault samples equally important in machine learning training. It ignores that the unit is in a normal state most of the time, and the collected normal samples are far more than the fault samples. In addition, fault information based on time-domain and frequency-domain methods requires professional knowledge or prior knowledge of the system, and sometimes it is very difficult or even impossible to obtain. At the same time, this information is usually designed for specific degradation processes, and it is difficult to obtain the multi-level and essential non-linear dynamic characteristics of wind turbines, and it is difficult to generalize them well to other units. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement requirements of the prior art, the present invention proposes a method and system for on-line abnormal detection of wind turbines using only normal samples, which can achieve high-efficiency and high-precision state abnormal detection, and can provide strong guidance for the operation and maintenance of wind farms.

[0005] To achieve the above object, according to one aspect of the present invention, there is provided a method for on-line abnormal detection of wind turbines using only normal samples, including:

[0006] Collecting the on-line operation vibration signal of the target component of the wind turbine;

[0007] Using a multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the on-line operation vibration signal;

[0008] Calculating the distance from the non-linear dynamic characteristics of the on-line operation vibration signal to the center of the SVDD hypersphere space, and comparing it with the radius of the pre-constructed hypersphere space to determine whether the unit is in an abnormal state and issuing an alarm for the abnormal state.

[0009] Among them, the target components include the main bearing of the wind turbine, the gearbox, the high-speed shaft bearing, the generator, etc.

[0010] In some alternative embodiments, the use of the multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the on-line operation vibration signal includes:

[0011] Decompose the on-line operation vibration signal x = {x(i), i = 1, 2,..., N} of the target component of the wind turbine with length N into multi-level sub-signals y (s) = {y (s) (i)}, where s is the number of levels of the multi-level quantitative Poincaré map;

[0012] Define two vectors m and n: k = N - s;

[0013] The Poincaré map of each level is represented by the point (m(k), n(k)). The Poincaré map appears as a point cloud elongated along the main line (m(k) = n(k)). The distribution of points along the main line is called the long-term correlation, and the distribution perpendicular to the main line is called the short-term correlation;

[0014] Calculate the short-term correlation SD1 and the long-term correlation SD2 according to the vectors m and n;

[0015] The non-linear dynamic characteristics captured by the multi-level quantitative Poincaré map are obtained by merging SD1 and SD2 of each level through the Concat splicing operation.

[0016] In some alternative embodiments, the multi-level sub-signal y (s) = {y (s) (i)} in where s is the number of levels of the multi-level quantitative Poincaré map. When s = 1, y (1) is equal to the on-line operation vibration signal x of the target component of the wind turbine. When s > 1, the on-line operation vibration signal x of the target component of the wind turbine is decomposed into multiple sub-signals.

[0017] In some alternative embodiments, where, represents the average value of all values in the vector m, represents the average value of all values in the vector n.

[0018] In some alternative embodiments, GMPOP(x, s) = Concat[SD1 (1) , SD2 (1) , SD1 (2) , SD2 (2) ,..., SD1 (s) , SD2 (s)Obtain the nonlinear dynamic characteristics captured by the multi-level quantitative Poincaré map.

[0019] In some alternative embodiments, calculating the distance from the nonlinear dynamic characteristics of the online running vibration signal to the center of the hypersphere space of SVDD includes:

[0020] From Obtain the distance from the nonlinear dynamic characteristics of the online running vibration signal to the center of the hypersphere space of SVDD, where K(·) is the kernel function, xz is the nonlinear dynamic characteristics of the online running vibration signal captured by the multi-level quantitative Poincaré map, α is the Lagrange multiplier, z is the nonlinear dynamic characteristics of the historical normal running vibration signal captured by the multi-level quantitative Poincaré map, and Nz is the length of the feature z.

[0021] In some alternative embodiments, the method for constructing the radius of the hypersphere space includes:

[0022] Collect the historical normal running vibration signals of the target components of the wind turbine, and capture the multi-level nonlinear dynamic characteristics in the historical normal running vibration signals by using the multi-level quantitative Poincaré map;

[0023] Train the SVDD model based on the multi-level nonlinear dynamic characteristics in the historical normal running vibration signals, construct the hypersphere space of SVDD, and obtain the radius of the hypersphere space.

[0024] In some alternative embodiments, from Obtain the radius R of the hypersphere space, where K(·) is the kernel function, z sv is the support vector point learned by the SVDD model from the nonlinear dynamic characteristics of the historical normal running vibration signals, Nz is the length of the feature z, z is the nonlinear dynamic characteristics of the historical normal running vibration signals captured by the multi-level quantitative Poincaré map, and α is the Lagrange multiplier.

[0025] According to another aspect of the present invention, there is provided an abnormal online detection system for a wind turbine using only normal samples, including:

[0026] A data acquisition module for collecting the online running vibration signals of the target components of the wind turbine;

[0027] A feature mining module for capturing the multi-level nonlinear dynamic characteristics in the online running vibration signals by using the multi-level quantitative Poincaré map;

[0028] A detection module for calculating the distance from the nonlinear dynamic characteristics of the online running vibration signal to the center of the hypersphere space of SVDD, comparing it with the radius of the hypersphere space pre-constructed by normal samples, determining whether the unit is in an abnormal state, and issuing an alarm for the abnormal state.

[0029] According to another aspect of the present invention, there is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method described in any one of the above.

[0030] Generally speaking, compared with the prior art by the above technical solution conceived by the present invention, the following beneficial effects can be achieved:

[0031] The present invention specifically includes a multi-level quantitative Poincaré map and SVDD. The multi-level quantitative Poincaré map is innovatively proposed, which captures the state evolution characteristics in the signal from the perspective of nonlinear dynamics, can effectively characterize the internal change patterns of complex nonlinear wind turbines, realizes the differentiation of normal and abnormal samples, and provides high-quality information for subsequent anomaly detection using only normal samples. Based on the multi-level quantitative Poincaré map of positive samples, the machine learning detector SVDD is introduced to automatically learn and construct an anomaly detection decision domain, avoiding the threshold setting of manual experience and improving the decision accuracy. The wind turbine fault detection method proposed by the present invention reduces the dependence on abnormal data, significantly improves the adaptability and scalability, has a fast processing speed and high dynamic detection accuracy, and can send alarm information in time if an abnormal situation occurs during operation. Therefore, the wind turbine fault detection method integrating the multi-level quantitative Poincaré map and SVDD can improve the equipment status monitoring efficiency and realize high-precision, high-reliability and strong-adaptability online anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flowchart of the implementation of a method for online anomaly detection of a wind turbine using only normal samples provided by an embodiment of the present invention;

[0033] Figure 2 is the structure of the multi-level quantitative Poincaré map provided by an embodiment of the present invention;

[0034] Figure 3 is the result of a method for online anomaly detection of a wind turbine using only normal samples provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0036] The overall anomaly detection process of the present invention can be divided into two stages: the first stage uses multi-level quantitative Poincare maps to extract key fault dynamic features; the second stage uses support vector data description (SVDD) to construct anomaly detection decision criteria based on the features of positive samples to achieve anomaly detection tasks. This provides a new method for online detection of wind turbine anomalies with higher detection efficiency, accuracy and adaptability.

[0037] Embodiment 1

[0038] like Figure 1 As shown, a flow chart of a method for online abnormality detection of a wind turbine using only normal samples provided by an embodiment of the present invention includes the following steps:

[0039] Step 1: Collect historical vibration signals of the core bearings of the wind turbine, use the data for the initial period as normal operation data, and use the remaining data as real-time verification data to verify the online detection capability of the detection method;

[0040] Step 2: Use the multi-level quantitative Poincare map to analyze the vibration signal of the wind turbine bearing collected, capture its dynamic evolution trend, and construct the feature space, where the level of the multi-level quantitative Poincare map is set to 10;

[0041] Step 3: Based on the features of normal samples, the hypersphere space of SVDD is trained to obtain the anomaly detection decision domain, wherein the radial basis kernel function is selected as the kernel function; the hyperparameters of SVDD are obtained by Bayesian optimization, and thus the construction of the anomaly detection decision hypersphere space is realized;

[0042] Step 4: Input the measured sample signal, use the multi-level quantitative Poincare map to obtain the characteristics of the measured signal, calculate the distance from the characteristic to the center of the hypersphere space, and compare it with the radius of the constructed hypersphere space. If the distance of the sample is greater than the radius of the hypersphere, it is judged as an abnormal sample and an abnormal alarm is issued. Compare the detection result with the actual operation status to determine the online detection performance of the model and realize online anomaly detection.

[0043] In the embodiment of the present invention, Figure 2 As shown, the specific method of analyzing and processing the vibration signal using the multi-level quantitative Poincare map in step 2 includes:

[0044] Decompose a vibration signal x = {x(i), i = 1, 2, ..., N} of length N into multi-level sub-signals y (s) ={y (s) (i)}, so we have:

[0045]

[0046]

[0047] where s is the number of levels of the multi-level quantitative Poincaré plot; when s = 1, y (1) is equal to the original signal x; when s > 1, the original signal x is decomposed into multiple sub-signals.

[0048] Subsequently, two vectors m and n are defined:

[0049] m = (m(1), m(2),..., m(k)) = (y (s) (1), y (s) (2),..., y (s) (N - s))

[0050] n = (n(1), n(2),..., n(k)) = (y (s) (2), y (s) (3),..., y (s) (N - s + 1))

[0051] where k = N - s.

[0052] The Poincaré plot of each level can be represented by the points (m(k), n(k)). The Poincaré plot usually appears as a cloud of points elongated along the main line (m(k) = n(k)). The distribution of points along the main line is called long-term correlation, and the distribution perpendicular to the main line is called short-term correlation.

[0053] Calculate the short-term correlation SD1 and the long-term correlation SD2:

[0054]

[0055]

[0056] where the overbar indicates taking the average.

[0057] The Concat concatenation operation combines the SD1 and SD2 of each level to obtain the features captured by the multi-level quantitative Poincaré plot:

[0058] GMPOP(x, s) = Concat[SD1 (1) , SD2 (1) , SD1 (2) , SD2 (2) ,..., SD1 (s) , SD2 (s)

[0059] In the embodiment of the present invention, step 3 trains the SVDD model based on the multi-level quantitative Poincaré plot parameters of normal samples to construct a hypersphere space, and its radius is:

[0060] ​

[0061] where K(·) is the kernel function; Nz is the length of feature z; z is the nonlinear dynamic feature captured from normal operation data using multi-level quantitative Poincare maps, z sv is the support vector point learned by SVDD from the nonlinear dynamic feature z; α is the Lagrange multiplier.

[0062] In the embodiment of the present invention, step 4 is specifically as follows:

[0063] Execute step 2 to obtain the nonlinear dynamic characteristics xz of the real-time verification data captured by the multi-level quantitative Poincare map, and calculate the distance from the sample xz to the center of the hypersphere space:

[0064]

[0065] If d>R, the sample is judged as an abnormal sample and an abnormal alarm is issued. The detection results are compared with the actual operation situation to determine the online detection performance of the detection model and realize online anomaly detection.

[0066] according to Figure 3 The results show that the wind turbine fault detection method integrating multi-level quantitative Poincare map and SVDD can detect all abnormal conditions including performance degradation, and give abnormal warnings, and there is only one false alarm point in the normal stage. The results show the advancement of the wind turbine abnormality online detection method of the present invention.

[0067] Embodiment 2

[0068] In another embodiment of the present invention, there is also provided a wind turbine generator set abnormality online detection system using only normal samples, comprising:

[0069] A data acquisition module is used to collect online operation vibration signals of target components of a wind turbine;

[0070] Feature mining module, used to capture multi-level nonlinear dynamic features in online vibration signals using multi-level quantitative Poincare maps;

[0071] The detection module is used to calculate the distance from the nonlinear dynamic characteristics of the online vibration signal to the center of the SVDD hypersphere space, and compare it with the radius of the hypersphere space constructed in advance through normal samples to determine whether the unit is in an abnormal state and issue an alarm for the abnormal state.

[0072] The specific implementation of each module can refer to the description of the above method embodiment, and will not be repeated in this embodiment.

[0073] It should be noted that according to the needs of implementation, each step / component described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0074] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims

1. An abnormal online detection method for wind turbine generators using only normal samples, characterized in that, Including: Collecting the on-line operation vibration signals of the target components of the wind turbine generator set; Using a multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the on-line operation vibration signals; Calculating the distance from the non-linear dynamic characteristics of the on-line operation vibration signals to the center of the SVDD hypersphere space, comparing it with the radius of the hypersphere space pre-constructed by normal samples, determining whether the unit is in an abnormal state, and sending an alarm for the abnormal state; The using of a multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the on-line operation vibration signals includes: Decompose the on-line operation vibration signal x = {x(i), i = 1, 2, …, N} of the target component of the wind turbine with length N into multiple levels of sub-signals y (s) ={y (s) (i)}, where s is the number of levels of the multi-level quantitative Poincaré plot; Define two vectors m and n: k = N - s; Each level of the Poincaré map is represented by points (m(k), n(k)), and the Poincaré map shows a point cloud elongated along the main line (m(k)=n(k)). The distribution of points along the main line is called long-term correlation, and the distribution perpendicular to the main line is called short-term correlation; Calculating the short-term correlation SD1 and the long-term correlation SD2 according to vectors m and n; Using the Concat splicing operation to merge the SD1 and SD2 of each level to obtain the non-linear dynamic characteristics captured by the multi-level quantitative Poincaré map.

2. The method according to claim 1, wherein Multi - level sub - signal y (s) ={y (s) (i)} in where s is the number of levels of the multi - level quantitative Poincaré diagram. When s = 1, y (1) is equal to the online operation vibration signal x of the target component of the wind turbine. When s>1, the online operation vibration signal x of the target component of the wind turbine is decomposed into multiple sub - signals.

3. The method according to claim 2, wherein Among them, represents the average value of all values in vector m, represents the average value of all values in vector n.

4. The method according to claim 3, characterized in that The non - linear dynamic characteristics captured by the multi - level quantitative Poincaré plot are obtained from GMPOP(x,s) = Concat[SD1 (1) , SD2 (1) , SD1 (2) , SD2 (2) ,..., SD1 (s) , SD2 (s) .

5. The method according to claim 4, wherein The calculating of the distance from the non-linear dynamic characteristics of the on-line operation vibration signals to the center of the SVDD hypersphere space includes: From obtain the distance from the non - linear dynamic feature of the online - running vibration signal to the center of the sphere in the SVDD hypersphere space, where K(·) is the kernel function, xz is the non - linear dynamic feature of the online - running vibration signal captured by using the multi - level quantitative Poincaré map, α is the Lagrange multiplier, z is the non - linear dynamic feature of the historical normal - running vibration signal captured by using the multi - level quantitative Poincaré map, and Nz is the length of the feature z.

6. The method according to any one of claims 1 to 5, characterized in that The construction method of the hypersphere space radius includes: Collecting the historical normal operation vibration signals of the target components of the wind turbine generator set, and using a multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the historical normal operation vibration signals; Training an SVDD model based on the multi-level non-linear dynamic characteristics in the historical normal operation vibration signals, constructing the hypersphere space of SVDD, and obtaining the hypersphere space radius.

7. The method according to claim 6, characterized in that, From obtain the hypersphere space radius R, where K(·) is the kernel function, z sv is the support vector point learned by the SVDD model from the nonlinear dynamic characteristics of the historical normal operation vibration signals, z is the nonlinear dynamic characteristics of the historical normal operation vibration signals captured by using the multi-level quantitative Poincaré map, Nz is the length of the feature z, and α is the Lagrange multiplier.

8. An abnormal online detection system for wind turbine generators that only uses normal samples, characterized in that, Including: A data acquisition module for collecting the on-line operation vibration signals of the target components of the wind turbine generator set; A feature mining module for using a multi-level quantitative Poincaré map to capture multi-level non-linear dynamic characteristics in the on-line operation vibration signals; A detection module for calculating the distance from the non-linear dynamic characteristics of the on-line operation vibration signals to the center of the SVDD hypersphere space, comparing it with the radius of the hypersphere space pre-constructed by normal samples, determining whether the unit is in an abnormal state, and sending an alarm for the abnormal state; The feature mining module is used to decompose the on-line operation vibration signal x = {x(i), i = 1, 2,..., N} of the target component of the wind turbine with a length of N into multi-level sub-signals y (s) ={y (s) (i)}, where s is the number of levels of the multi-level quantitative Poincaré map; define two vectors m and n: k = N - s; each level of the Poincaré map is represented by the point (m(k), n(k)). The Poincaré map shows a point cloud elongated along the main line (m(k) = n(k)). The distribution of points along the main line is called long-term correlation, and the distribution perpendicular to the main line is called short-term correlation; according to the vectors m and n, calculate the short-term correlation SD1 and the long-term correlation SD2; the SD1 and SD2 of each level are merged by the Concat splicing operation to obtain the non-linear dynamic characteristics captured by the multi-level quantitative Poincaré map.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images