A fan fault diagnosis method based on CVA and PLSR

Abstract

The application provides a fan fault diagnosis method based on CVA and PLSR, which combines PLSR with CVA to extract a new statistic quantity to identify the fault of the fan, the identification accuracy is greatly improved after the method is used, and the method has the advantages of clear mechanism, simple threshold judgment, easy realization and strong practicability. It comprises the following steps: step 1, obtaining operation data and performing typical variable analysis; step 2, using PLSR to model the typical state and the calculated new feature; step 3, calculating two statistics according to the typical state and the obtained new feature; and step 4, determining an alarm threshold, and setting the alarm threshold for different faults.

Description

Technical Field

[0001] This invention relates to the field of wind power technology, specifically to a wind turbine fault diagnosis method based on CVA and PLSR. Background Technology

[0002] Canonical variable analysis (CVA) has become increasingly popular with the development of computer technology and is now widely validated in the Tenness Eastman (TE) process.

[0003] When applied to other industrial processes, such as directly applying canonical variable analysis (CVA) to the fault diagnosis of wind turbines, many problems arise. For example, using CVA alone makes it difficult to identify blade icing faults, and the probability of false positives and false negatives for other faults is relatively high.

[0004] Therefore, how to improve the typical scalar analysis (CVA) for identifying wind turbine faults by using other algorithms has been a long-term research problem in the field of wind power technology. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a wind turbine fault diagnosis method based on CVA and PLSR. It combines PLSR and CVA to extract a new statistical quantity to identify wind turbine faults. Using this method, the identification accuracy is greatly improved. Furthermore, this method has the advantages of clear mechanism, simple threshold judgment, easy implementation, and strong practicality.

[0006] A wind turbine fault diagnosis method based on CVA and PLSR, characterized by comprising the following steps:

[0007] Step 1: Obtain runtime data and perform typical variable analysis;

[0008] Step 2: Use PLSR to model typical states and the newly calculated features;

[0009] Step 3: Calculate two statistics based on the typical state and the newly obtained characteristics;

[0010] Step 4: Determine alarm thresholds and set alarm thresholds for different faults.

[0011] Its further features are:

[0012] Step 1: Collect normal operating data in the industrial process. Before processing, the data is denoised. Then, according to the algorithm of canonical variable analysis, the data at any time point is transformed into a typical state containing canonical variable c and typical residual e. If a fault occurs, these states can be used to extract fault characteristics, thus providing a basis for fault diagnosis.

[0013] In the canonical variable analysis of step 1, let y be the observation vector, y∈Rm×N Let y represent m variables and N observation points, with past and future time lags of p and f, respectively. For each past data point y... p (k) will collect data from the past p observation points, and for each future data point y... f (k) Collect data from f future observation points, and then arrange all past and future data points in chronological order to form a past and future matrix. The past and future matrix has M observation points, where M = Np-f+1.

[0014] Where the past matrix is: Y p =[y p (p+1), y p (p+2), ...y p (p+M)]∈R mp×M ;

[0015] Future matrix: Y f =[y f (p+1), y f (p+2), ...y f (p+M)]∈R mf×M ;

[0016] The data is normalized, as shown in the formula below. and These are the average values ​​of the past and future matrices, respectively, σ(y) p ) and σ(y f If ) are their standard deviations, then there exist matrices J and L such that J∑ppJ T =I,L∑ffL T =I,J∑pfL T = diag(r1, r2, ... r) r ,0,...,0), where ∑pp,∑ff,J∑pf----covariance matrices of p and p, f and f, p and f; I is a block diagonal matrix, the first block is the identity matrix, the second block is a diagonal matrix, and the elements in the diagonal matrix have the following relationship: r1≥r2≥…≥r r ;

[0017] Singular value decomposition yields J and L. The transformation matrices J and L can convert process data into uncorrelated canonical variables c and canonical residuals e, i.e., c(t) = J. r *Y p (t), e(t) = J q *Y p (t); where J r It is the first r lines of J, J q It is the last q lines of J, where q = mp - r;

[0018] Based on the canonical variable c and the canonical residual e, the statistic T is summarized. s 2 and T r 2 Statistic T s 2 and T r 2 They are defined as follows: i is the i-th feature of the residual;

[0019] Here, a new variable related to the typical variable c or the typical residual e is introduced, which is more sensitive to faults. The new variable is defined as b = PCA(c T c), d = PCA(e) T e) b and d are just symbols and will not be used later. They will be given more specific names in the specific implementation.

[0020] PCA() indicates that after dimensionality reduction through principal component analysis, the statistical measures obtained from these two new variables will significantly increase the discriminative power of subsequent faults. After obtaining the statistical measures of the new variables, the corresponding threshold can be calculated. When the measures exceed the threshold, it indicates that a fault has occurred.

[0021] In step 2, PLSR is used to perform regression modeling on the typical state and the new feature z, and a mapping model is established to obtain the regression matrix B. The new feature z at any time point i can be obtained by using the regression matrix to calculate the typical residual e. i z = PCA(e T .e), Z i =e i As shown in *B+F; similarly, the typical variable c is processed in the same way;

[0022] In step 2, during partial least squares regression, let X be the observation matrix, Y be the prediction matrix, and p1 and q1 be the first principal component axis vectors of X and Y, respectively. The optimization objective is Maximize(Xp1, Yq1); Subject to ||p1‖|=1||q1||=1. After solving, we can obtain the principal component axis vectors p1 and q1. We can also calculate the first pair of related principal components u1 and v1 of X and Y, where u1=Xp1 and v1=Yq1.

[0023] Using the least squares regression algorithm, model X and Y with their principal components u1 and v1 respectively, as shown in the following formula: X = u1c1 T +E;Y=v1d1 T +G; E and G are residual matrices. Then, based on the correlation between u1 and v1, the principal component u1 is used to replace v1 in the model as follows: Y = u1r1 T +F,

[0024] Then, the remaining part E in X excluding principal component u1 is taken as the new X, and the remaining part F in Y excluding principal component v1 is taken as the new Y. This process is repeated using the formula for partial least squares regression. The iteration stops when the residual F reaches the required accuracy. Finally, we have X = u1c1 T +u2c2 T +...u n c n T +E, Y=u1r1 T +u2r2 T +…u n r n T +F.

[0025] Written in matrix form as X = UC T +E, Y = UR T +F=XPR T +F=XB+F, where B=pR T This formula is the regression equation from X to Y. During the review process, only the values ​​of p and r need to be recorded.

[0026] In step 3, after the data from any observation point are processed by the transformation matrix J and the regression matrix B, a new feature z can be obtained. i Based on the typical state and the new features, two statistics T can be calculated. r 2 and T sr 2 For future use.

[0027] In step 4, different methods are used to select alarm thresholds for different faults in the baseline model. When rotor speed sensor faults and torque actuator faults occur, the fluctuation of statistics is small, and the difference between statistics during fault and non-fault periods is obvious. Therefore, no filter is selected. However, a low-pass filter is required in other fault cases because the statistics fluctuate greatly and have peaks. The specific settings of filtering, thresholds and dead zones need to be calibrated according to the actual situation for different wind fields.

[0028] After adopting this method, a new detection quantity T was proposed. sr 2 T relative to CVA r 2 In terms of statistics, it not only improves the detection capability of common faults, but also shows significant effectiveness in diagnosing some difficult-to-detect faults in wind turbine units. It employs a new statistical measure, T, proposed by combining canonical variable analysis and partial least squares regression. sr 2It can effectively improve the ability to identify wind turbine faults and has a certain degree of versatility. In simulation verification, CVA-PLSR reduced the false detection rate, false negative rate and delay time by more than half. In general, the newly proposed statistics have higher sensitivity, stronger ability to extract fault information and higher detection accuracy. Users can use the method proposed in this paper to perform status detection and fault diagnosis on the running data of their wind turbine generators, transmit the wind turbine's operating data to the data center, and output the results to the management personnel after analysis and diagnosis by this method. If a fault is found, a warning will be issued. Detailed Implementation

[0029] A wind turbine fault diagnosis method based on CVA and PLSR includes the following steps:

[0030] Step 1: Acquire operational data and perform typical variable analysis. The specific operation of Step 1 is as follows: Collect normal operational data in the industrial process, reduce noise in the data before processing, and then transform the data at any point in time into typical states of typical variable c and typical residual e according to the algorithm of typical variable analysis. If a fault occurs, these states can be used to extract fault characteristics, thereby providing a basis for fault diagnosis.

[0031] In the canonical variable analysis of step 1, let y be the observation vector, y∈R m×N Let y represent m variables and N observation points, with past and future time lags of p and f, respectively. For each past data point y... p (k) will collect data from the past p observation points, and for each future data point y... f (k) Collect data from f future observation points, and then arrange all past and future data points in chronological order to form a past and future matrix. The past and future matrix has M observation points, where M = Np-f+1.

[0032] Where the past matrix is: Y p =[y p (p+1), y p (p+2), ...y p (p+M)]∈R mp×M ;

[0033] Future matrix: Y f =[y f (p+1), y f (p+2), ...y f (p+M)]∈R mf×M ;

[0034] The data is normalized, as shown in the formula below. and These are the average values ​​of the past and future matrices, respectively, σ(y) p ) and σ(y f If ) are their standard deviations, then there exist matrices J and L such that J∑ppJ T =I,L∑fffL T =I,J∑pfL T = diag(r1, r2, ... r) r Let I be a block diagonal matrix, where ∑pp, ∑ff, J∑pf are the covariance matrices of p and p, f and f, and p and f. The first block is the identity matrix, and the second block is a diagonal matrix. The elements of the diagonal matrix have the following relationship: r1≥r2≥…≥r r .

[0035] Singular value decomposition yields J and L. The transformation matrices J and L can convert process data into uncorrelated canonical variables c and canonical residuals e, i.e., c(t) = J. r *Y p (t), e(t) = J q *Y p (t). Among them, J r It is the first r lines of J, J q It is the last q lines of J, where q = mp - r.

[0036] Based on the canonical variable c and the canonical residual e, the statistic T is summarized. s 2 and T r 2 Statistic T s 2 and T r 2 They are defined as follows: i is the i-th feature of the residual.

[0037] Here, a new variable related to the typical variable c or the typical residual e is introduced, which is more sensitive to faults. The new variable is defined as b = PCA(c T c), d = PCA(e) T e) b and d are just symbols and will not be used later. They will be given more specific names in the specific implementation.

[0038] PCA() indicates that after dimensionality reduction through principal component analysis, the statistical measures obtained from these two new variables will significantly increase the discriminative power of subsequent faults. After obtaining the statistical measures of the new variables, the corresponding threshold can be calculated. When the measures exceed the threshold, it indicates that a fault has occurred.

[0039] Step 2: Use PLSR to model the typical state and the calculated new feature; use PLSR to perform regression modeling on the typical state and the new feature z, and establish a mapping model to obtain the regression matrix B. Using the regression matrix to calculate the typical residual e, the new feature Z at any time point i can be obtained. i z = PCA(e T .e), z i =e i As shown in *B+F, the same process applies to the typical variable c;

[0040] In step 2, during partial least squares regression, let X be the observation matrix, Y be the prediction matrix, and p1 and q1 be the first principal component axis vectors of X and Y, respectively. The optimization objective is Maximize(Xp1, Yq1); Subject to ||p1||=1||q1||=1. After solving, we can obtain the principal component axis vectors p1 and q1. We can then calculate the first pair of correlated principal components u1 and v1 of X and Y, where u1=Xp1 and v1=Yq1.

[0041] Using the least squares regression algorithm, model X and Y with their principal components u1 and v1 respectively, as shown in the following formula: X = u1c1 T +E;Y=v1d1 T +G. E and G are residual matrices; then, based on the correlation between u1 and v1, the principal component u1 is used to replace v1 in the model as follows: Y = u1r1 T +F,

[0042] Then, the remaining part E in X excluding principal component u1 is taken as the new X, and the remaining part F in Y excluding principal component v1 is taken as the new Y, where F is the residual matrix. This process is repeated iteratively according to the formula for partial least squares regression. The iteration stops when the residual F reaches the required accuracy. Finally, we have X = u1c1 T +u2c2 T +…u n c n T +E, Y=u1r1 T +u2r2 T +…u n r n T +F.

[0043] Written in matrix form as X = UC T +E, Y = UR T +F=XPR T +F=XB+F, where B=PR T This formula is the regression equation from X to Y. During the review process, only the values ​​of p and r need to be recorded.

[0044] Step 3: Calculate two statistics based on the typical state and the obtained new features; after the data of any observation point are processed by the transformation matrix J and the regression matrix B, the new feature Z can be obtained. i Based on the typical state and the new features, two statistics T can be calculated. r 2 and T sr 2 For future use;

[0045] Step 4: Determine the alarm threshold. Alarm thresholds are set for different faults. Different methods are used to select alarm thresholds for different faults in the baseline model. When rotor speed sensor and torque actuator faults occur, the fluctuation of statistics is small, and the difference between statistics during fault and non-fault periods is significant. Therefore, no filter is selected. However, in other fault cases, a low-pass filter is required because the statistics fluctuate greatly and have peaks. Specific settings for filtering, thresholds, and dead zones need to be calibrated according to the actual conditions of different wind farms.

[0046] The meanings of the English or Chinese abbreviations in the text are as follows:

[0047] CVA: Canonical Variate Analysis; Canonical Variable Analysis

[0048] PLSR: Partial Least Squares Regression.

[0049] PCA: Principal Component Analysis.

[0050] TE: Tennessee Eastman. Eastman Chemical Company in the United States developed the Tennessee Eastman (TE) simulation platform, an open and challenging chemical engineering modeling platform. The data it generates has time-varying, strongly coupled, and nonlinear characteristics, and is widely used to test control and fault diagnosis models of complex industrial processes.

[0051] Data center: A place with many computing resources used to store and compute wind farm return data; it can also be called a computing center or information center.

[0052] In practical implementation, several common faults were considered, including sensor and actuator failures in different locations. The table lists their fault types, each varying in severity. Some are extremely serious, requiring a safe and rapid shutdown of the wind turbine; others are less severe, and the controller can adapt to these faults during operation, although performance may be reduced. The baseline model considers additive and multiplicative faults, as well as faults causing dynamic changes in various parts of the system. These fault descriptions are derived from actual wind turbine failures, but due to confidentiality reasons, detailed descriptions of actual faults cannot be provided. In summary, the faults selected in the baseline model cover different parts of the wind turbine, different types, and different degrees of severity.

[0053] Fault Description

[0054]

[0055] Different methods are used to select alarm thresholds for different faults in the baseline model. For faults 4 and 7, the statistical fluctuations are small, and the difference between the statistical values ​​during fault and non-fault periods is significant; therefore, no filter is selected. However, a low-pass filter is needed for other fault cases because the statistical fluctuations are large and have peaks. Looking at the statistics for fault 1, the difference between fault and non-fault time periods is very obvious, making threshold selection relatively simple; the maximum value of the fault-free time period can be directly selected (mode Max). However, for pitch actuator faults 2, 5, and 6, the Euclidean distance is small, requiring consideration of multiple factors when selecting the threshold. Finally, a threshold of 1% is chosen to achieve the alarm rate during the fault-free time period (mode Falserate1%). To directly compare the advantages and disadvantages of the two methods, no dead zone is set here. In actual wind turbine blade icing faults, the statistical fluctuations are severe, so a low-pass filter is also used, and the threshold is set to achieve the alarm rate of 5% during the fault-free time period (mode Falserate5%), again without a dead zone. The table below details the threshold selection modes and parameters.

[0056] Selection modes and parameters for different fault thresholds

[0057]

[0058] The principle is as follows, and it proposes a new detection quantity T. sr 2 T relative to CVA r 2 In terms of statistics, it not only improves the detection capability of common faults, but also shows significant effectiveness in diagnosing some difficult-to-detect faults in wind turbine units. It employs a new statistical measure, T, proposed by combining canonical variable analysis and partial least squares regression. sr 2It can effectively improve the ability to identify wind turbine faults and has a certain degree of versatility. In simulation verification, CVA-PLSR reduced the false detection rate, false negative rate and delay time by more than half. In general, the newly proposed statistics have higher sensitivity, stronger ability to extract fault information and higher detection accuracy. Users can use the method proposed in this paper to perform status detection and fault diagnosis on the running data of their wind turbine generators, transmit the wind turbine's operating data to the data center, and output the results to the management personnel after analysis and diagnosis by this method. If a fault is found, a warning will be issued.

[0059] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0060] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A wind turbine fault diagnosis method based on CVA and PLSR, characterized in that, It includes the following steps: Step 1: Acquire operational data and perform canonical variable analysis. Collect normal operational data in the industrial process. Before processing, the data is denoised. Then, according to the algorithm of canonical variable analysis, the data at any point in time is transformed into typical states containing canonical variables c and typical residuals e. If a fault occurs, these states can be used to extract fault characteristics, thereby providing a basis for fault diagnosis. Step 2: Use PLSR to model the typical state and the calculated new feature. Use PLSR to perform regression modeling on the typical state and the new feature z, and establish a mapping model to obtain the regression matrix B. Use the regression matrix B to calculate the new feature z at any time point i from the typical residual e. i z = PCA(e T ·e),z i =e i *B+F; Similarly, using regression matrix B to calculate the canonical variable c, we can obtain the new feature z at any time point i. i z = PCA(c T ·c),Z i =c i *B+F; Step 3: Calculate two statistics based on the typical state and the obtained new features. After the data of any observation point is processed by the transformation matrix J and the regression matrix B, the new feature Z can be obtained. i ; Based on the typical state and the new features, two statistics T are calculated. r 2 and T r 2 For future use; Step 4: Determine the alarm threshold. Set alarm thresholds for different faults. There are different methods for selecting alarm thresholds for different faults in the baseline model. When rotor speed sensor faults and torque actuator faults occur, the fluctuation of statistics is small, and the difference between statistics during fault and non-fault periods is obvious. Therefore, no filter is selected. However, in other fault cases, a low-pass filter needs to be used because the statistics fluctuate greatly and have peaks. The specific settings of filtering, thresholds and dead zones need to be calibrated according to the actual situation of different wind fields. In the canonical variable analysis of step 1, let y be the observation vector, y∈R m×N Let y represent m variables and N observation points, with past and future time lags of p and f, respectively. For each past data point y... p (k) will collect data from the past p observation points, and for each future data point y... f (k) Collect data from f future observation points, and then arrange all past and future data points in chronological order to form a past matrix and a future matrix. The number of observation points in the past matrix and the future matrix is ​​M, where M = Np-f+1. Where the past matrix is: Y p =[y p (p+1), y p (p+2), ...y p (p+M)]∈R mp×M ; Future matrix: Y f =y f (p+1), y f (p+2), ...y f (p+M)]∈R mf×M ; After normalizing the data, there exist matrices J and L such that J∑ppJ T =I,L∑ffL T =I,J∑pfL T = diag[r1, r2, ... r r ,0,...,0), where ∑pp,∑ff,j∑pf----covariance matrices of p and p, f and f, and p and f; I is a block diagonal matrix, the first block is the identity matrix, the second block is a diagonal matrix, and the elements in the diagonal matrix have the following relationship: r1≥r2≥…≥r r ; Singular value decomposition yields J and L. The transformation matrices J and L can convert process data into uncorrelated canonical variables c and canonical residuals e, i.e., c(t) = J. r *Y p (t),e(t)=J q *Y p (t); where J r It is the first r lines of J, J q It is the last q lines of J, where q = mp - r; Based on the canonical variable c and the canonical residual e, the statistic T is summarized. s 2 and T r 2 Canonical variable T s 2 and T r 2 They are defined as follows: i is the i-th feature of the residual; Here, a new variable related to the typical variable c or the typical residual e is introduced, which is more sensitive to faults. The new variable is defined as b = PCA(c T c), d = PCA(e) T e) b and d are just symbolic representations and will not be used later. They will be given more specific names in the specific implementation. PCA() indicates that after dimensionality reduction through principal component analysis, the statistical measures obtained from these two new variables will significantly increase the discriminative power of subsequent faults. After obtaining the statistical measures of the new variables, the corresponding threshold can be calculated. When the measures exceed the threshold, it indicates that a fault has occurred.

2. The wind turbine fault diagnosis method based on CVA and PLSR as described in claim 1, characterized in that: Let X be the observation matrix and Y be the prediction matrix, and p1 and q1 be the first principal component axis vectors of X and Y, respectively; the optimization objective is Maximize(Xp1, Yq1); Subject to ||p1||=1||q1||=1, after solving, we can obtain the principal component axis vectors p1 and q1, and calculate the first pair of related principal components u1 and v1 of X and Y, u1=Xp1, v1=Yq1; Using the least squares regression algorithm, model X and Y with their principal components u1 and v1 respectively, as shown in the following formula: X = u1c1 T +E;Y=v1d1 T +G; E and G are residual matrices; then, based on the correlation between u1 and v1, the principal component u1 is used to replace v1 in the model as follows: Y = u1r1 T +F, Then, the remaining part E in X excluding principal component u1 is taken as the new X, and the remaining part F in Y excluding principal component v1 is taken as the new Y, where F is the residual matrix. This process is repeated iteratively according to the formula for partial least squares regression. The iteration stops when the residual F reaches the required accuracy. Finally, we have X = u1c1 T +u2c2 T +…u n c n T +E, Y=u1r1 T +u2r2 T +…u n r n T +F; Written in matrix form as X = UC T +E, Y = UR T +F=XPR T +F=XB+F, where B=PR T This formula is the regression equation from X to Y. During the review process, only the values ​​of p and r need to be recorded.

Images