Wind driven generator bearing fault diagnosis method and system

Abstract

The invention discloses a wind driven generator bearing fault diagnosis method and system, and relates to the technical field of wind power generation equipment fault diagnosis. The method comprises the following steps: collecting bearing vibration data by using a specific sensor and preprocessing; de-noising is carried out through an enhanced combined differential morphological filter ECGMF; decomposing the denoised signal by using an intrinsic time decomposition (ITD) algorithm, extracting a component instantaneous amplitude and performing spectral analysis to obtain a fault feature; constructing a deep support vector machine DSVM model in combination with a support vector machine SVM and a neural network; and inputting the feature vector into a deep support vector machine DSVM model for training, and judging a fault output result. The system comprises a vibration data acquisition module, an enhanced combined differential morphological filtering module, an inherent time scale decomposition module, a depth support vector machine module and a diagnosis result output module, the functions of all the steps are correspondingly achieved, fault features are effectively extracted and accurately diagnosed through cooperation of all the links, and the accuracy and efficiency of fault diagnosis of the wind driven generator bearing can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation equipment fault diagnosis, and more specifically, to a wind power generator bearing fault diagnosis method and system. Background Art

[0002] As an important renewable energy power generation equipment, wind turbines are often in harsh working environments during operation, facing interference from strong background noise and electromagnetic noise. These interference factors make it difficult to accurately detect fault characteristic information of generator bearings, which in turn brings many challenges to the normal operation and maintenance of wind turbines.

[0003] In the existing fault diagnosis technology, there are the following problems:

[0004] First, for the detection of rolling bearing fault feature information, traditional methods are difficult to effectively overcome the influence of strong noise interference. Although some filters in mathematical morphology can be used for signal processing, the existing kurtosis criterion and signal-to-noise ratio method are not accurate enough when selecting the scale of the structural element, making it difficult to accurately extract weak fault feature information from strong background noise.

[0005] Secondly, the fault vibration signal of the spherical roller bearing of the wind turbine is non-stationary and nonlinear. The traditional fault feature extraction method is difficult to effectively adapt to this complex signal characteristic and cannot fully and accurately extract the feature information that can reflect the bearing fault status.

[0006] Finally, as a commonly used classification tool, support vector machine (SVM) cannot fully explore the deep laws of data when processing data. For complex wind turbine bearing fault diagnosis scenarios, its diagnostic effect has certain limitations.

[0007] Therefore, there is an urgent need for a wind turbine bearing fault diagnosis method and system that is more accurate, efficient, and able to adapt to complex environments and diverse failure modes. Summary of the invention

[0008] In view of the deficiencies in the prior art, an object of the present invention is to provide a method and system for diagnosing bearing faults of a wind turbine generator.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] In a first aspect, a method for diagnosing a bearing fault of a wind turbine generator comprises the following steps:

[0011] The wind turbine bearing vibration data is collected and preprocessed through D / SENL020 and D / SENN020 sensors;

[0012] The collected wind turbine bearing vibration signal is denoised by using enhanced combined differential morphological filter (ECGMF).

[0013] The intrinsic time decomposition (ITD) algorithm is used to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing.

[0014] The support vector machine SVM is combined with the neural network model to construct a deep support vector machine DSVM model;

[0015] The feature vector is input into the deep support vector machine (DSVM) model for training, and the fault of the wind turbine bearing is judged and the diagnosis result is output.

[0016] Preferably, the vibration data includes main bearing axial vibration data, main bearing horizontal vibration data, main bearing vertical vibration data, rear bearing vertical vibration data, stator axial vibration data, and stator horizontal vibration data; the preprocessing includes data visualization inspection and data normalization.

[0017] Preferably, the collected vibration signal is subjected to denoising processing by using an enhanced combined differential morphological filter ECGMF, comprising:

[0018] Determine the multi-scale range to be adopted based on data visualization analysis and understanding of the characteristics of the vibration source; select a flat structural element with zero height;

[0019] For each selected scale value, the corresponding structural element size is set; the combination of various mathematical morphological operations involved in the enhanced combined differential morphological filter ECGMF transformation at the current scale is clarified; mathematical morphological operations are performed on the preprocessed vibration signal in turn;

[0020] The mathematical morphology is composed of four basic operators: dilation, erosion, opening and closing. Assume that the original one-dimensional signal f(n) is defined as F=(0,1,…,N-1), and the structure element g(m) is defined as G=(0,1,…,M-1), N≥M. Where n is an independent variable representing the position index or serial number of the discrete sampling points of the one-dimensional signal f(n); m is a variable representing the position index or serial number of the element in the domain of the structure element g(m). The four basic operators are defined as follows:

[0021] Dilation operator:

[0022] Corrosion operator: (fΘg)(n)=min[f(n+m)-g(m)]

[0023] Open operator:

[0024] Closing operator:

[0025] in, ——dilation operator; Θ——erosion operator; ——Open operator; ·——Close operator;

[0026] The corrosion operator can reduce the peak value of the signal and expand the trough of the signal, and the closing operator can suppress the negative pulse of the signal; on the contrary, the expansion operator can reduce the valley value of the signal and expand the peak value of the signal, and the opening operator can suppress the positive pulse of the signal;

[0027] According to this property, two cascade operators consisting of dilation operator and closing operator are defined as:

[0028]

[0029] A cascade operator consisting of an erosion operator and an opening operator is defined as:

[0030]

[0031] F DC and F CD The ability to suppress negative pulses and extract positive pulse signals is enhanced; on the contrary, F EO The ability to suppress positive pulses and extract negative pulses has been enhanced, so new difference operators can be defined:

[0032]

[0033] Considering that the early fault information of wind turbine bearings is often interfered by severe background noise, in order to extract weak fault feature information from strong background noise, combined with F CD_EO and F DC_EO Based on the characteristics of the operator, a new enhanced combined differential morphology operator is proposed:

[0034] ECGMF(n)=F CD_EO (n)·F DC_EO (n)

[0035] At each scale, the spectrum analysis is performed on the vibration signal after the current scale enhanced combined differential morphological filter ECGMF transformation operation;

[0036] Determine the characteristic frequencies associated with equipment failures based on the failure mechanism of the target equipment, previous experience, and related theoretical research;

[0037] For the spectrum at each scale, calculate the characteristic frequency intensity coefficient, then the characteristic frequency intensity coefficient C CFI The definition is as follows:

[0038]

[0039] in, ——the amplitude of the ith fault frequency; ——the amplitude of each j point; M, N——fault frequency and total spectrum quantity;

[0040] Since the fault characteristic frequency is relatively less affected by interference, the characteristic frequency intensity coefficient C is selected CFI To determine the optimal scale of the structural element; the required characteristic frequency intensity coefficient C CFI The larger the value is, the more obvious the fault characteristic signal extraction is. Therefore, the maximum characteristic frequency intensity coefficient C CFI The scale corresponding to the value is taken as the optimal scale of the structural element;

[0041] According to the determined optimal scale of the structural element, the corresponding structural element parameters are set; the various mathematical morphological operations involved in the enhanced combined differential morphological filter ECGMF transformation under the optimal scale are re-determined; the pre-processed original vibration signal is substituted into the enhanced combined differential morphological filter ECGMF for mathematical morphological operations to complete the denoising process.

[0042] Preferably, the inherent time decomposition (ITD) algorithm is used to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing, including:

[0043] The denoised vibration signal is decomposed using intrinsic time decomposition (ITD). The vibration signal can be decomposed into a series of intrinsic rotation components with different characteristic scales and the sum of a trend component, specifically:

[0044] For signal X t , define ξ as the baseline extraction operator, so that from the signal X t The residual signal remaining after extracting a baseline from the signal X becomes an inherent rotation component; t A decomposition of is:

[0045] X t =ξX t +(1-ξ)X t =L t +H t

[0046] Among them, L t =ξX t is the baseline signal, H t =(1-ξ)X t is the intrinsic rotation component;

[0047] Assume {T K, K=1,2,…} is the signal X t The local pole of K and L K represents X(T K ) and L(T K ); Assume that in [0,T K ]Define L t and H t , and X t At t∈[0,T K+2 ] is defined; in the interval between consecutive poles (T K ,T K+1 ] defines the piecewise linear baseline extraction factor ξ:

[0048]

[0049] Among them, α∈(0,1), take 0.5; a baseline signal L is obtained by one decomposition. t and an intrinsic rotation component H t , where the intrinsic rotation component H t Represents the local relatively high-frequency component in the signal;

[0050] The baseline signal obtained from the first decomposition is used as the new input signal to continue decomposing until a monotonic signal is obtained; this converts the original vibration signal X t Decomposed into the sum of several inherent rotation components of different frequency bands from high to low and a monotonic trend component. The whole process is expressed as:

[0051]

[0052] Among them, ψξ K X t is the intrinsic rotation component of the K+1th layer, ξ P X t It is the monotonic trend component or the baseline signal extracted to the lowest frequency by stopping the decomposition before the monotonic trend component is obtained;

[0053] Perform spectrum analysis on the instantaneous amplitude of the first K inherent rotation components containing obvious periodic impact components to extract the bearing fault feature vector Where i = 1, 2, ..., k, specifically:

[0054] The intrinsic time decomposition (ITD) algorithm takes the full wave as the unit and proposes a new method to define the instantaneous amplitude, instantaneous phase and instantaneous frequency.

[0055] The instantaneous phase, full wave refers to the signal between two adjacent upper zero crossing points; half wave refers to the signal between two zero crossing points; monotonic interval refers to the signal between two adjacent extreme value points; based on this, the instantaneous phase θt It can be defined as:

[0056]

[0057] Among them, A1>0, A2>0, respectively represent the amplitude of the positive half-wave and negative half-wave parts of a full wave, t1 and t5 are two upper zero-crossing moments, t2 is the maximum point (A1) of the positive half-wave, t3 is the lower zero-crossing moment, and t4 is the minimum point (-A2) of the negative half-wave;

[0058] The instantaneous frequency, instantaneous frequency f t is the instantaneous phase θ t The differential of , that is:

[0059]

[0060] The instantaneous amplitude, instantaneous amplitude A t Defined in half-wave units, that is, the value of the extreme point of the signal between two zero crossings is a constant in each half-wave, that is:

[0061]

[0062] After completing the spectrum analysis, the characteristic information related to the bearing fault is extracted from the spectrum of each inherent rotation component to form the bearing fault feature vector Where i = 1, 2, ..., k, that is, the above feature extraction operation is performed on the first K selected inherent rotation components to obtain the corresponding bearing fault feature vector.

[0063] Preferably, the support vector machine SVM is combined with the neural network model to construct a deep support vector machine DSVM model, including:

[0064] The overall structure of the deep support vector machine DSVM model includes an input layer, a hidden layer and an output layer; the input layer receives the extracted bearing fault feature vector; the hidden layer performs feature mapping on the input data and selects a fully connected neural network structure as the hidden layer; the output layer replaces the output layer of the neural network with a support vector machine SVM, uses the support vector machine SVM to classify the mapped data, and divides the sample data into positive and negative categories by solving the classification hyperplane;

[0065] The objective function of the support vector machine (SVM) is also applicable to the classification process of the output layer of the deep support vector machine (DSVM) model. Assume that the original data is X i , the data after hidden layer mapping is Φ(x i ), then the objective function of the deep support vector machine DSVM model is as follows:

[0066]

[0067] in, represents the normal vector of the hyperplane; b represents the intercept of the hyperplane; C represents the penalty function; ξ i Represents a slack variable, which is mainly introduced to deal with the possible linear inseparability of data; y i In the binary classification problem, y i The value of is +1 or -1, corresponding to two different categories; Φ(x i ) represents the data after the hidden layer mapping of the deep support vector machine DSVM model; l represents the number of samples;

[0068] Rewrite the constraints of the objective function as follows:

[0069]

[0070] The objective function is rewritten as an unconstrained optimization problem, which is called the loss function of the deep support vector machine DSVM model, as shown below:

[0071]

[0072] The loss function of the deep support vector machine DSVM model is mainly divided into two parts. The first part is Represents the margin hyperplane Distance maximization is also the regularization term in the neural network model; Part II is the hinge loss, which allows some sample points to be misclassified; C is the penalty parameter. If C is selected to be a larger value, the number of misclassified sample points will be as small as possible; if C is selected to be a smaller value, the distance between the interval hyperplanes will be as large as possible, which can tolerate some misclassified training sample points.

[0073] Preferably, the feature vector is input into a deep support vector machine DSVM model for training, and the wind turbine bearing fault is judged, and the diagnosis result is output, including:

[0074] The extracted bearing fault feature vector data set is divided into training set, validation set and test set. The feature vector of the training set is input into the deep support vector machine DSVM model. After hidden layer feature mapping, the prediction result is obtained by the support vector machine SVM classification of the output layer.

[0075] According to the prediction results and the true labels, the loss value is calculated according to the loss function of the deep support vector machine DSVM model, and then the back propagation algorithm is used to update the parameters of each layer of the model, and the training is iterated repeatedly until the stopping condition is met;

[0076] During or after iterative training, input the model with the feature vector of the validation set to obtain the prediction result, and calculate the evaluation indicators including accuracy, recall rate, and F1 value in combination with the corresponding true label, and optimize the model accordingly;

[0077] Finally, the test set feature vector is input into the optimized model to obtain the prediction result, and the fault diagnosis result is output in a clear way for bearing fault diagnosis.

[0078] In a second aspect, a wind turbine bearing fault diagnosis system includes a vibration data acquisition module, an enhanced combined differential morphological filter module, an inherent time scale decomposition module, a deep support vector machine module, and a diagnosis result output module;

[0079] The vibration data acquisition module collects the vibration data of the wind turbine bearing through the D / SENL020 and D / SENN020 sensors and performs preprocessing;

[0080] The enhanced combined differential morphological filter module performs denoising processing on the collected wind turbine bearing vibration signal through the enhanced combined differential morphological filter ECGMF;

[0081] The intrinsic time scale decomposition module uses the intrinsic time decomposition ITD algorithm to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing;

[0082] The deep support vector machine module combines the support vector machine SVM with the neural network model to construct a deep support vector machine DSVM model;

[0083] The diagnosis result output module inputs the feature vector into a deep support vector machine (DSVM) model for training, performs fault diagnosis on the wind turbine bearing, and outputs the diagnosis result.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] 1. In the present invention, the enhanced combined differential morphological filter ECGMF is used to detect bearing fault information, and the characteristic frequency intensity coefficient is used to optimize the scale of the structural element, so that the fault characteristic information in the rolling bearing signal can be effectively extracted. Compared with other morphological filters, the enhanced combined differential morphological filter ECGMF is superior in extracting the fault components of the rolling bearings of large wind turbines and suppressing noise. At the same time, from the two measurement indicators of kurtosis and fault energy ratio, the enhanced combined differential morphological filter ECGMF has obvious advantages over other morphological filters; for the fault diagnosis of rolling bearings of large wind turbines, the enhanced combined differential morphological filter EGMF has higher feature extraction efficiency than other morphological filters.

[0086] 2. In the present invention, the vibration signal of the wind turbine bearing has complex nonlinear, time-varying and multi-coupling characteristics. The inherent time decomposition ITD algorithm can decompose the complex signal into the sum of several inherent rotation components and one trend component. The instantaneous amplitude of the inherent rotation component containing obvious periodic impact components is subjected to spectral analysis, and the amplitude at the fault characteristic frequency is extracted as the characteristic vector for bearing fault diagnosis. The inherent time decomposition ITD algorithm can accurately extract the dynamic characteristics of non-stationary signals, has high disassembly efficiency and frequency resolution, is suitable for analyzing non-stationary signals with time-varying spectra, has small edge effects, does not require spline interpolation, has no screening process, and can process large amounts of data in real time.

[0087] 3. In the present invention, the deep support vector machine DSVM model is used to diagnose the fault of the generator bearing. On the one hand, compared with the support vector machine SVM rice noodles, the hidden layer of the deep support vector machine DSVM model can replace the process of manually selecting the kernel function. In addition, the deep support vector machine DSVM model updates the network parameters through back propagation, and does not need to solve the quadratic programming problem. On the other hand, compared with the neural network, the classification process of the output layer of the deep support vector machine DSVM model is more interpretable. It improves the ability to mine deep laws of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 A schematic flow chart of a wind turbine bearing fault diagnosis method is provided for the present invention;

[0089] Figure 2 A schematic diagram of a wind turbine bearing fault diagnosis system flow is proposed for the present invention;

[0090] Figure 3 This is a structural block diagram of the deep support vector machine DSVM model proposed in the present invention. DETAILED DESCRIPTION

[0091] Reference Figures 1 to 3 .

[0092] The embodiment further illustrates a wind turbine bearing fault diagnosis method and system proposed by the present invention.

[0093] A method for diagnosing a bearing fault of a wind turbine generator comprises the following steps:

[0094] The wind turbine bearing vibration data is collected and preprocessed through D / SENL020 and D / SENN020 sensors;

[0095] The collected wind turbine bearing vibration signal is denoised by using enhanced combined differential morphological filter (ECGMF).

[0096] The intrinsic time decomposition (ITD) algorithm is used to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing.

[0097] The support vector machine SVM is combined with the neural network model to construct a deep support vector machine DSVM model;

[0098] The feature vector is input into the deep support vector machine (DSVM) model for training, and the fault of the wind turbine bearing is judged and the diagnosis result is output.

[0099] The vibration data includes main bearing axial vibration data, main bearing horizontal vibration data, main bearing vertical vibration data, rear bearing vertical vibration data, stator axial vibration data, and stator horizontal vibration data; the preprocessing includes data visualization inspection and data normalization.

[0100] The data visualization inspection plots the collected original vibration signal as a time domain waveform diagram and a frequency domain spectrum diagram, and visually observes the basic characteristics of the signal through visualization, such as whether there are obvious abnormal fluctuations, high-frequency noise interference, and the approximate frequency distribution range of the signal.

[0101] The data normalization is to normalize the vibration signal so that the value range of the signal falls within a specific interval. It is common to normalize it to the interval [0,1] or [-1,1]. The minimum-maximum normalization formula can be used: Where x is the original signal, x min and x max are the minimum and maximum values ​​in the original signal; normalization helps subsequent filters better handle signals of different magnitudes, and in some algorithms it can promote convergence and stable operation.

[0102] The D / SENL020 sensor has the characteristics of high sensitivity and can accurately capture small changes in the bearing vibration process, which plays an important role in detecting some initial and relatively weak fault signals.

[0103] The D / SENN020 sensor excels in broadband response and can cover a wide range of vibration frequencies, ensuring that no vibration frequency components that may be related to faults are missed.

[0104] The collected wind turbine bearing vibration signal is denoised by using the enhanced combined differential morphological filter ECGMF, including:

[0105] The enhanced combined differential morphological filter ECGMF is a filter designed based on the principle of mathematical morphology, which can effectively remove noise by analyzing and processing the morphological characteristics of the signal.

[0106] Determine the multi-scale range to be used based on data visualization analysis and understanding of the vibration source characteristics;

[0107] The structural element is mainly composed of three parts: height, length and shape; the shapes of the structural elements are mainly triangle, ellipse and flat; the more complex the shape of the structural element, the longer the calculation time; therefore, in order to improve the calculation efficiency, the enhanced combined differential morphological filter ECGMF selects a flat structural element with a height of zero; if the scale of the structural element is too long, the fault characteristic signal may be regarded as noise filtering and the characteristic information may be lost; on the contrary, if it is too short, a large amount of noise components will remain in the signal; therefore, it is necessary to select a suitable structural element scale.

[0108] For each selected scale value, the corresponding structural element size is set; the combination of various mathematical morphological operations involved in the enhanced combined differential morphological filter ECGMF transformation at the current scale is clarified; and mathematical morphological operations are performed on the preprocessed vibration signal in turn.

[0109] The mathematical morphology is composed of four basic operators: dilation, erosion, opening and closing. Assume that the original one-dimensional signal f(n) is defined as F=(0,1,…,N-1), and the structure element g(m) is defined as G=(0,1,…,M-1), N≥M. Where n is an independent variable representing the position index or serial number of the discrete sampling points of the one-dimensional signal f(n); m is a variable representing the position index or serial number of the element in the domain of the structure element g(m). The four basic operators are defined as follows:

[0110] Dilation operator:

[0111] Corrosion operator: (fΘg)(n)=min[f(n+m)-g(m)]

[0112] Open operator:

[0113] Closing operator:

[0114] in, ——dilation operator; Θ——erosion operator; ——Open operator; ·——Close operator.

[0115] The corrosion operator can reduce the peak value of the signal and expand the trough of the signal, and the closing operator can suppress the negative pulse of the signal; on the contrary, the expansion operator can reduce the valley value of the signal and expand the peak value of the signal, and the opening operator can suppress the positive pulse of the signal.

[0116] According to this property, two cascade operators consisting of dilation operator and closing operator are defined as:

[0117]

[0118] A cascade operator consisting of an erosion operator and an opening operator is defined as:

[0119]

[0120] F DC and F CD The ability to suppress negative pulses and extract positive pulse signals is enhanced; on the contrary, F EO The ability to suppress positive pulses and extract negative pulses has been enhanced, so new difference operators can be defined:

[0121]

[0122] Considering that the early fault information of wind turbine bearings is often interfered by serious background noise, in order to extract weak fault feature information from strong background noise, combined with F CD_EO and F DC_EO Based on the characteristics of the operator, a new enhanced combined differential morphology operator is proposed:

[0123] ECGMF(n)=F CD_EO (n)·F DC_EO (n)

[0124] At each scale, the vibration signal after the current scale enhanced combined differential morphological filter ECGMF transformation operation is subjected to spectral analysis. Usually, the time domain data is converted into frequency domain data by using methods such as fast Fourier transform (FFT) to obtain the spectrum of the vibration data at the current scale. The spectrum shows the energy distribution of the vibration signal at different frequencies, that is, the amplitude of each frequency component.

[0125] Determine the characteristic frequencies associated with equipment failures based on the failure mechanism of the target equipment (e.g. wind turbine bearings), previous experience, and related theoretical research;

[0126] For the spectrum at each scale, calculate the characteristic frequency intensity coefficient, then the characteristic frequency intensity coefficient C CFI The definition is as follows:

[0127]

[0128] in, ——the amplitude of the ith fault frequency; ——the amplitude of each j point; M, N——fault frequency and total spectrum quantity.

[0129] Since the fault characteristic frequency is relatively less affected by interference, the characteristic frequency intensity coefficient C is selected CFI To determine the optimal scale of the structural element; the required characteristic frequency intensity coefficient C CFI The larger the value is, the more obvious the fault characteristic signal extraction is. Therefore, the maximum characteristic frequency intensity coefficient C CFI The scale corresponding to the value is taken as the optimal scale of the structural element;

[0130] According to the determined optimal scale of the structural element, the corresponding structural element parameters are set; the various mathematical morphological operations involved in the enhanced combined differential morphological filter ECGMF transformation under the optimal scale are re-determined; the pre-processed original vibration signal is substituted into the enhanced combined differential morphological filter ECGMF for mathematical morphological operations to complete the denoising process.

[0131] The intrinsic time decomposition (ITD) algorithm is used to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing, including:

[0132] The components decomposed by the intrinsic time decomposition (ITD) algorithm have certain physical meanings, can effectively reflect the characteristics of the original signal, and are suitable for analyzing signals containing amplitude modulation and frequency modulation components.

[0133] The denoised vibration signal is subjected to intrinsic time decomposition (ITD) decomposition. The vibration signal can be decomposed into a series of intrinsic rotation components with different characteristic scales and the sum of a trend component, specifically:

[0134] For signal X t , define ξ as the baseline extraction operator, so that from the signal X t The residual signal remaining after extracting a baseline from the signal X becomes an inherent rotation component; t A decomposition of is:

[0135] X t =ξX t +(1-ξ)X t =L t +H t

[0136] Among them, L t =ξX t is the baseline signal, H t =(1-ξ)X t is the intrinsic rotation component;

[0137] Assume {T K , K=1,2,…} is the signal X t The local pole of K and LK represents X(T K ) and L(T K ); Assume that in [0,T K ]Define L t and H t , and X t At t∈[0,T K+2 ] is defined; in the interval between consecutive poles (T K ,T K+1 ] defines the piecewise linear baseline extraction factor ξ:

[0138]

[0139] Among them, α∈(0,1), take 0.5; a baseline signal L is obtained by one decomposition. t and an intrinsic rotation component H t , where the intrinsic rotation component H t Represents the local relatively high-frequency component in the signal.

[0140] The baseline signal obtained from the first decomposition is used as the new input signal to continue decomposing until a monotonic signal is obtained; this converts the original vibration signal X t Decomposed into the sum of several inherent rotation components of different frequency bands from high to low and a monotonic trend component. The whole process is expressed as:

[0141]

[0142] Among them, ψξ K X t is the intrinsic rotation component of the K+1th layer, ξ P X t It is the monotonic trend component or the baseline signal extracted to the lowest frequency by stopping the decomposition before the monotonic trend component is obtained.

[0143] Perform spectrum analysis on the instantaneous amplitude of the first K inherent rotation components containing obvious periodic impact components to extract the bearing fault feature vector Where i = 1, 2, ..., k, specifically:

[0144] The intrinsic time decomposition (ITD) algorithm takes the full wave as the unit and proposes a new method to define instantaneous amplitude, instantaneous phase and instantaneous frequency.

[0145] The instantaneous phase, full wave refers to the signal between two adjacent upper zero crossing points; half wave refers to the signal between two zero crossing points; monotonic interval refers to the signal between two adjacent extreme value points; based on this, the instantaneous phase θ t It can be defined as:

[0146]

[0147] Among them, A1>0, A2>0, respectively representing the amplitude of the positive half-wave and negative half-wave parts of a full wave, t1 and t5 are two upper zero-crossing moments, t2 is the maximum point (A1) of the positive half-wave, t3 is the lower zero-crossing moment, and t4 is the minimum point (-A2) of the negative half-wave.

[0148] The instantaneous amplitude, instantaneous amplitude A t Defined in half-wave units, that is, the value of the extreme point of the signal between two zero crossings is a constant in each half-wave, that is:

[0149]

[0150] After completing the spectrum analysis, the characteristic information related to the bearing fault is extracted from the spectrum of each inherent rotation component to form the bearing fault feature vector Where i = 1, 2, ..., k, that is, the above feature extraction operation is performed on the first K selected inherent rotation components to obtain the corresponding bearing fault feature vector.

[0151] The support vector machine SVM is combined with the neural network model to build a deep support vector machine DSVM model, including:

[0152] The overall structure of the deep support vector machine (DSVM) model includes an input layer, a hidden layer and an output layer; the input layer receives the extracted bearing fault feature vector; the hidden layer performs feature mapping on the input data and selects a fully connected neural network structure as the hidden layer; the output layer replaces the output layer of the neural network with an SVM model, and uses the SVM model to classify the mapped data. By solving the classification hyperplane, the sample data is divided into positive and negative categories. The classification process is easy to understand, which enhances the interpretability of the DSVM model.

[0153] The objective function of the support vector machine (SVM) is also applicable to the classification process of the output layer of the deep support vector machine (DSVM) model. Assume that the original data is X i , the data after hidden layer mapping is Φ(x i ), then the objective function of the deep support vector machine DSVM model is as follows:

[0154]

[0155] in, represents the normal vector of the hyperplane; b represents the intercept of the hyperplane; C represents the penalty function; ξ i Represents a slack variable, which is mainly introduced to deal with the possible linear inseparability of data; y i In the binary classification problem, y iThe value of is +1 or -1, corresponding to two different categories; Φ(x i ) represents the data after the hidden layer mapping of the deep support vector machine DSVM model; l represents the number of samples;

[0156] Rewrite the constraints of the objective function as follows:

[0157]

[0158] The objective function is rewritten as an unconstrained optimization problem, which is called the loss function of the deep support vector machine DSVM model, as shown below:

[0159]

[0160] The loss function of the deep support vector machine DSVM model is mainly divided into two parts. The first part is Represents the margin hyperplane Distance maximization is also the regularization term in the neural network model; Part II is the hinge loss, which allows some sample points to be misclassified; C is the penalty parameter. If C is selected to be a larger value, the number of misclassified sample points will be as small as possible; if C is selected to be a smaller value, the distance between the interval hyperplanes will be as large as possible, which can tolerate some misclassified training sample points.

[0161] The feature vector is input into the deep support vector machine DSVM model for training, and the fault diagnosis of the wind turbine bearing is performed to output the diagnosis results, including:

[0162] The extracted bearing fault feature vector dataset is divided into training set, validation set and test set according to a certain ratio; the common division ratio is 80% as training set, 10% as validation set and 10% as test set; ensure that the format and labeling of the data are accurate so that the model can correctly identify and process it.

[0163] The feature vector of the training set is input into the input layer of the deep support vector machine DSVM model, and then passes through the hidden layer for feature mapping. The data processed by the hidden layer is then input into the support vector machine SVM of the output layer for classification to obtain the predicted classification result.

[0164] According to the predicted classification results and the actual sample labels, the loss value is calculated according to the loss function of the deep support vector machine DSVM model; using the calculated loss value, the parameters of the model are updated from the output layer to the input layer through the back propagation algorithm, including the connection weights and bias of the hidden layer and the parameters of the support vector machine SVM.

[0165] Repeat the above steps of forward propagation, loss calculation, back propagation and parameter update, and perform multiple iterative training until the model reaches the predetermined stopping condition; through continuous iterative training, the deep support vector machine DSVM model can gradually learn the characteristics and rules of the data, improve the classification accuracy and the generalization ability of the model.

[0166] During each iterative training process or after training is completed, the feature vector of the validation set is input into the trained deep support vector machine DSVM model, and the predicted classification result is obtained according to the above forward propagation steps; then, the evaluation indicators, including accuracy, recall rate, and F1 value, are calculated based on the predicted results and the real fault labels corresponding to the feature vector of the validation set, and optimization is performed based on the calculated evaluation indicator results;

[0167] The accuracy rate refers to the ratio of the number of correctly predicted samples to the total number of samples; the calculation formula is: Among them, TP represents true positive examples, that is, the number of samples that are actually positive examples and predicted to be positive examples; TN represents true negative examples, that is, the number of samples that are actually negative examples and predicted to be negative examples; FP represents false positive examples, that is, the number of samples that are actually negative examples but predicted to be positive examples; FN represents false negative examples, that is, the number of samples that are actually positive examples but predicted to be negative examples;

[0168] The recall rate Recall refers to the ratio of the number of correctly predicted positive samples to the number of actual positive samples, and the calculation formula is:

[0169] The F1 value is the harmonic mean of precision and recall, which takes into account the balance between precision and recall. The calculation formula is:

[0170] The test set feature vector is input into the evaluated and optimized trained deep support vector machine DSVM model, and the predicted classification result is obtained according to the above forward propagation steps; the predicted fault diagnosis result is output in a clear and understandable way.

[0171] A wind turbine bearing fault diagnosis system includes a vibration data acquisition module, an enhanced combined differential morphological filter module, an inherent time scale decomposition module, a deep support vector machine module and a diagnosis result output module;

[0172] The vibration data acquisition module collects the wind turbine bearing vibration data through D / SENL020 and D / SENN020 sensors and performs preprocessing;

[0173] The enhanced combined differential morphological filter module performs denoising processing on the collected wind turbine bearing vibration signal through the enhanced combined differential morphological filter ECGMF;

[0174] The intrinsic time scale decomposition module uses the intrinsic time decomposition ITD algorithm to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing;

[0175] The deep support vector machine module combines the support vector machine SVM with the neural network model to construct a deep support vector machine DSVM model;

[0176] The diagnosis result output module inputs the feature vector into a deep support vector machine (DSVM) model for training, performs fault diagnosis on the wind turbine bearing, and outputs the diagnosis result.

[0177] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A wind turbine bearing fault diagnosis method, characterized in that: The following steps are involved: The wind turbine bearing vibration data is collected and preprocessed through D / SENL020 and D / SENN020 sensors; The collected wind turbine bearing vibration signal is denoised by using enhanced combined differential morphological filter (ECGMF). The intrinsic time decomposition (ITD) algorithm is used to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing. The support vector machine SVM is combined with the neural network model to construct a deep support vector machine DSVM model; The feature vector is input into the deep support vector machine (DSVM) model for training, and the fault of the wind turbine bearing is judged and the diagnosis result is output.

2. A wind turbine bearing fault diagnosis method according to claim 1, characterized in that: The vibration data includes main bearing axial vibration data, main bearing horizontal vibration data, main bearing vertical vibration data, rear bearing vertical vibration data, stator axial vibration data, and stator horizontal vibration data; the preprocessing includes data visualization inspection and data normalization.

3. A wind turbine bearing fault diagnosis method according to claim 2, characterized in that: The collected vibration signal is denoised by using the enhanced combined differential morphological filter (ECGMF), including: Determine the multi-scale range to be adopted based on data visualization analysis and understanding of the characteristics of the vibration source; select a flat structural element with zero height; For each selected scale value, the corresponding structural element size is set; the combination of various mathematical morphological operations involved in the enhanced combined differential morphological filter ECGMF transformation at the current scale is clarified; mathematical morphological operations are performed on the preprocessed vibration signal in turn; The mathematical morphology is composed of four basic operators: dilation, erosion, opening and closing. Assume that the original one-dimensional signal f(n) is defined as F=(0,1,…,N-1), and the structure element g(m) is defined as G=(0,1,…,M-1), N≥M. Where n is an independent variable representing the position index or serial number of the discrete sampling points of the one-dimensional signal f(n); m is a variable representing the position index or serial number of the element in the domain of the structure element g(m). The four basic operators are defined as follows: Dilation operator: Corrosion operator: (fΘg)(n)=min[f(n+m)-g(m)] Open operator: Closing operator: in, ——dilation operator; Θ——erosion operator; ——Open operator; ·——Close operator; The corrosion operator can reduce the peak value of the signal and expand the trough of the signal, and the closing operator can suppress the negative pulse of the signal; on the contrary, the expansion operator can reduce the valley value of the signal and expand the peak value of the signal, and the opening operator can suppress the positive pulse of the signal; According to this property, two cascade operators consisting of dilation operator and closing operator are defined as: A cascade operator consisting of an erosion operator and an opening operator is defined as: F DC and F CD The ability to suppress negative pulses and extract positive pulse signals is enhanced; on the contrary, F EO The ability to suppress positive pulses and extract negative pulses has been enhanced, so new difference operators can be defined: Considering that the early fault information of wind turbine bearings is often interfered by serious background noise, in order to extract weak fault feature information from strong background noise, combined with F CD_EO and F DC_EO Based on the characteristics of the operator, a new enhanced combined differential morphology operator is proposed: ECGMF(n)=F CD_EO (n)·F DC_EO (n) At each scale, the spectrum analysis is performed on the vibration signal after the current scale enhanced combined differential morphological filter ECGMF transformation operation; Determine the characteristic frequencies associated with equipment failures based on the failure mechanism of the target equipment, previous experience, and related theoretical research; For the spectrum at each scale, calculate the characteristic frequency intensity coefficient, then the characteristic frequency intensity coefficient C CFI The definition is as follows: in, ——the amplitude of the ith fault frequency; ——the amplitude of each j point; M, N——fault frequency and total spectrum quantity; Since the fault characteristic frequency is relatively less affected by interference, the characteristic frequency intensity coefficient C is selected CFI To determine the optimal scale of the structural element; the required characteristic frequency intensity coefficient C CFI The larger the value is, the more obvious the fault characteristic signal extraction is. Therefore, the maximum characteristic frequency intensity coefficient C CFI The scale corresponding to the value is taken as the optimal scale of the structural element; According to the determined optimal scale of the structural element, the corresponding structural element parameters are set; the various mathematical morphological operations involved in the enhanced combined differential morphological filter ECGMF transformation under the optimal scale are re-determined; the pre-processed original vibration signal is substituted into the enhanced combined differential morphological filter ECGMF for mathematical morphological operations to complete the denoising process.

4. A wind turbine bearing fault diagnosis method according to claim 3, characterized in that: The intrinsic time decomposition (ITD) algorithm is used to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing, including: The denoised vibration signal is decomposed using intrinsic time decomposition (ITD). The vibration signal can be decomposed into a series of intrinsic rotation components with different characteristic scales and the sum of a trend component, specifically: For signal X t , define ξ as the baseline extraction operator, so that from the signal X t The residual signal remaining after extracting a baseline from the signal X becomes an inherent rotation component; t The first decomposition of is: X t =ξX t +(1-ξ)X t =L t +H t Among them, L t =ξX t is the baseline signal, H t =(1-ξ)X t is the intrinsic rotation component; Assume {T K , K=1,2,…} is the signal X t The local pole of K and L K represents X(T K ) and L(T K ); Assume that in [0,T K ]Define L t and H t , and X t At t∈[0,T K+2 ] is defined; in the interval between consecutive poles (T K ,T K+1 ] defines the piecewise linear baseline extraction factor ξ: Among them, α∈(0,1), take 0.5; a baseline signal L is obtained by one decomposition. t and an intrinsic rotation component H t , where the intrinsic rotation component H t Represents the local relatively high-frequency component in the signal; The baseline signal obtained from the first decomposition is used as the new input signal to continue decomposing until a monotonic signal is obtained; this converts the original vibration signal X t Decomposed into the sum of several inherent rotation components of different frequency bands from high to low and a monotonic trend component. The whole process is expressed as: Among them, ψξ K X t is the intrinsic rotation component of the K+1th layer, ξ P X t It is the monotonic trend component or the baseline signal extracted to the lowest frequency by stopping the decomposition before the monotonic trend component is obtained; Perform spectrum analysis on the instantaneous amplitude of the first K inherent rotation components containing obvious periodic impact components to extract the bearing fault feature vector Where i = 1, 2, ..., k, specifically: The intrinsic time decomposition (ITD) algorithm takes the full wave as the unit and proposes a new method to define the instantaneous amplitude, instantaneous phase and instantaneous frequency. The instantaneous phase, full wave refers to the signal between two adjacent upper zero crossing points; half wave refers to the signal between two zero crossing points; monotonic interval refers to the signal between two adjacent extreme value points; based on this, the instantaneous phase θ t It can be defined as: Among them, A1>0, A2>0, respectively represent the amplitude of the positive half-wave and negative half-wave parts of a full wave, t1 and t5 are two upper zero-crossing moments, t2 is the maximum point (A1) of the positive half-wave, t3 is the lower zero-crossing moment, and t4 is the minimum point (-A2) of the negative half-wave; The instantaneous frequency, instantaneous frequency f t is the instantaneous phase θ t The differential of , that is: The instantaneous amplitude, instantaneous amplitude A t Defined in half-wave units, that is, the value of the extreme point of the signal between two zero crossings is a constant in each half-wave, that is: After completing the spectrum analysis, the characteristic information related to the bearing fault is extracted from the spectrum of each inherent rotation component to form the bearing fault feature vector Where i = 1, 2, ..., k, that is, the above feature extraction operation is performed on the first K selected inherent rotation components to obtain the corresponding bearing fault feature vector.

5. A wind turbine bearing fault diagnosis method according to claim 4, characterized in that: The support vector machine SVM is combined with the neural network model to build a deep support vector machine DSVM model, including: The overall structure of the deep support vector machine DSVM model includes an input layer, a hidden layer and an output layer; the input layer receives the extracted bearing fault feature vector; the hidden layer performs feature mapping on the input data, and selects a fully connected neural network structure as the hidden layer; the output layer replaces the output layer of the neural network with a support vector machine SVM, and uses the support vector machine SVM to classify the mapped data, and divides the sample data into positive and negative categories by solving the classification hyperplane; The objective function of the support vector machine (SVM) is also applicable to the classification process of the output layer of the deep support vector machine (DSVM) model. Assume that the original data is X i , the data after hidden layer mapping is Φ(x i ), then the objective function of the deep support vector machine DSVM model is as follows: x i ≥0,i=1,…,l in, represents the normal vector of the hyperplane; b represents the intercept of the hyperplane; C represents the penalty function; ξ i Represents a slack variable, which is mainly introduced to deal with the possible linear inseparability of data; y i In the binary classification problem, y i The value of is +1 or -1, corresponding to two different categories; Φ(x i ) represents the data after the hidden layer mapping of the deep support vector machine DSVM model; l represents the number of samples; Rewrite the constraints of the objective function as follows: The objective function is rewritten as an unconstrained optimization problem, which is called the loss function of the deep support vector machine DSVM model, as shown below: The loss function of the deep support vector machine DSVM model is mainly divided into two parts. The first part is Represents the margin hyperplane and Distance maximization is also the regularization term in the neural network model; Part II is the hinge loss, which allows some sample points to be misclassified; C is the penalty parameter. If C is selected to be a larger value, the number of misclassified sample points will be as small as possible; if C is selected to be a smaller value, the distance between the interval hyperplanes will be as large as possible, which can tolerate some misclassified training sample points.

6. A wind turbine bearing fault diagnosis method according to claim 5, characterized in that: The feature vector is input into the deep support vector machine DSVM model for training, and the fault diagnosis of the wind turbine bearing is performed to output the diagnosis results, including: The extracted bearing fault feature vector data set is divided into training set, validation set and test set. The feature vector of the training set is input into the deep support vector machine DSVM model. After hidden layer feature mapping, the prediction result is obtained by the support vector machine SVM classification of the output layer. According to the prediction results and the true labels, the loss value is calculated according to the loss function of the deep support vector machine DSVM model, and then the back propagation algorithm is used to update the parameters of each layer of the model, and the training is iterated repeatedly until the stopping condition is met; During or after iterative training, input the model with the feature vector of the validation set to obtain the prediction result, and calculate the evaluation indicators including accuracy, recall rate, and F1 value in combination with the corresponding true label, and optimize the model accordingly; Finally, the test set feature vector is input into the optimized model to obtain the prediction result, and the fault diagnosis result is output in a clear way for bearing fault diagnosis.

7. A wind turbine bearing fault diagnosis system, applied to a wind turbine bearing fault diagnosis method according to claims 1-6, characterized in that: It includes a vibration data acquisition module, an enhanced combined differential morphological filter module, an intrinsic time scale decomposition module, a deep support vector machine module and a diagnosis result output module; The vibration data acquisition module collects the wind turbine bearing vibration data through D / SENL020 and D / SENN020 sensors and performs preprocessing; The enhanced combined differential morphological filter module performs denoising processing on the collected wind turbine bearing vibration signal through the enhanced combined differential morphological filter ECGMF; The intrinsic time scale decomposition module uses the intrinsic time decomposition ITD algorithm to decompose the denoised vibration signal, extract the instantaneous amplitude of the main component and perform spectrum analysis to obtain the fault characteristics of the wind turbine bearing; The deep support vector machine module combines the support vector machine SVM with the neural network model to construct a deep support vector machine DSVM model; The diagnosis result output module inputs the feature vector into a deep support vector machine (DSVM) model for training, performs fault diagnosis on the wind turbine bearing, and outputs the diagnosis result.

Images