Intelligent operation and maintenance management platform for wind turbine generator

Abstract

The invention belongs to the technical field of equipment management, and discloses an intelligent operation and maintenance management platform for a wind turbine generator. Comprising a data acquisition module used for acquiring operation data of a wind turbine generator; the data preprocessing module is used for preprocessing the operation data of the wind turbine generator to obtain preprocessed operation data; the fault detection module is used for carrying out fault detection on the wind turbine generator based on the preprocessed operation data and maintaining the wind turbine generator to obtain perfect operation data; the unit optimization module is used for optimizing the perfect operation data to obtain an optimal operation parameter combination; performing control optimization on the wind turbine generator based on the optimal operation parameter combination; the operation efficiency of the wind turbine generator is improved, and meanwhile the fault probability of the wind turbine generator is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment management, and more specifically, to an intelligent operation and maintenance management platform for wind turbines. Background Art

[0002] The patent with application publication number CN108960688A discloses an integrated management system for wind turbines, which relates to the field of wind power generation technology. It can establish an evaluation model for wind turbines, and score and rate the performance of wind turbines in three dimensions: day, month, and year, so as to facilitate timely discovery and resolution of problems, realize scientific management of wind turbines, and make the operation of wind farms more refined, standardized, and intelligent; the management system includes a quantitative evaluation page, a grade evaluation page, and a background database, and both the quantitative evaluation page and the grade evaluation page call on the data of the background database to perform their own calculations and finally display them; the quantitative evaluation page and the grade evaluation page respectively include a daily, monthly, and annual information quantitative rating management module and a grade evaluation management module.

[0003] In the field of existing technologies, the processing of wind turbine operating data is not accurate enough, and the preprocessing method is not efficient enough, making it difficult to process large-scale, multi-type wind turbine operating data; the existing wind turbine fault detection system may still rely on traditional threshold judgment methods, which may cause the fault detection system to report false alarms or miss reports, and may not be sensitive enough to slight changes in the wind turbine, resulting in poor fault detection effects; using traditional optimization algorithms to optimize wind turbines may ignore the influence of the environment or other factors, resulting in poor optimization effects; there is a lack of efficient wind turbine control methods, and it is impossible to quickly process data and perform fault detection based on the data. Manual intervention may be required at certain times, resulting in low efficiency in the operation and management of wind turbines.

[0004] In view of this, the present invention proposes an intelligent operation and maintenance management platform for wind turbines to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: an intelligent operation and maintenance management platform for wind turbines, comprising:

[0006] Data acquisition module, used to collect the operation data of wind turbines;

[0007] A data preprocessing module is used to preprocess the operation data of the wind turbine generator set to obtain preprocessed operation data;

[0008] The fault detection module detects faults of wind turbines based on pre-processed operation data, and performs maintenance on wind turbines to obtain complete operation data;

[0009] The unit optimization module optimizes the perfect operation data to obtain the optimal operation parameter combination; based on the optimal operation parameter combination, the wind turbine is controlled and optimized; each module is connected by wire and / or wireless means.

[0010] Furthermore, the operating data of the wind turbine set is collected by sensors installed on the wind turbine set; the operating data of the wind turbine set includes time series data and spectrum data.

[0011] Furthermore, the method of preprocessing the operation data of the wind turbine generator set includes:

[0012] Preprocessing the time series data to obtain complete time series data; preprocessing the spectrum data to obtain enhanced spectrum data; combining the complete time series data and the enhanced spectrum data to obtain preprocessed running data;

[0013] The method of preprocessing the time series data includes:

[0014] Construct a prediction model to process missing values ​​of time series data and obtain complete time series data;

[0015] The method of constructing a prediction model to process missing values ​​of time series data includes:

[0016] Collect historical time series data of wind turbines, and perform standardization on the historical time series data of wind turbines to obtain a training set of historical time series data; initially complete the training set of historical time series data as training labels for the prediction model;

[0017] Initialize the hidden sequence h t , where t represents the time index and t≤max, and max represents the maximum time corresponding to the data in the historical time series data training set;

[0018] The prediction model includes a loop update structure, a loop reset structure and an output layer, and the historical time series data training set is input into the loop update structure and the loop reset structure to calculate the loop update output and the loop reset output respectively;

[0019] Calculate candidate hidden outputs based on the loop reset output; calculate final hidden outputs based on the candidate hidden outputs and the loop update output; map the final hidden output to a predicted value through the output layer;

[0020] Use the predicted value to process the missing values ​​of the historical time series data training set to obtain a complete historical time series data training set; compare the complete historical time series data training set with the training label. If the absolute value of the difference between the numerical values ​​of the data at the corresponding time is less than or equal to the preset difference threshold, fix the parameters of the prediction model at this time to obtain a trained prediction model;

[0021] The loop updates the output Among them, W Z Represents updating the input weight matrix; U Z represents the update of hidden weight matrix; X t represents the input data at time t, that is, the data at time t in the historical time series data training set; b Z represents the bias term of the loop update structure; h t-1 represents the hidden sequence at time t-1; σ represents an activation function;

[0022] Cycle reset output Among them, W R Represents resetting the input weight matrix; U R Represents resetting the hidden weight matrix; b R A bias term representing the loop reset structure;

[0023] Candidate hidden output Where tanh represents the hyperbolic tangent function; W represents the input weight matrix of the candidate hidden output; U represents the hidden weight matrix of the candidate hidden output; ⊙ represents element-by-element multiplication;

[0024] Hide Sequence When t = max, h max For the final hidden output.

[0025] Furthermore, the method of preprocessing the spectrum data includes:

[0026] Constructing a complex wavelet basis function, and decomposing the spectrum data based on the constructed complex wavelet basis function to obtain wavelet coefficients, and decomposing the spectrum data into n sub-bands; selecting a target frequency band from the n sub-bands;

[0027] The spectrum data is divided into K intrinsic mode functions through variable mode decomposition, and the intrinsic mode functions corresponding to the target frequency band are reconstructed using inverse wavelet transform to obtain enhanced spectrum data.

[0028] Furthermore, the method of decomposing the spectrum data based on the constructed complex wavelet basis function includes:

[0029] Complex wavelet basis function B(q)=P(n1(q)+I×n2(q)); where q represents the time variable corresponding to the spectrum data; n1(q) represents the real wavelet basis function; n2(q) represents the imaginary wavelet basis function; I is the imaginary unit; P represents the scale factor;

[0030] The preset symmetric wavelet basis function is used as the real wavelet basis function n1(q); the Hilbert transform of the real wavelet basis function n1(q) is calculated, and the calculation result is used as the imaginary wavelet basis function n2(q);

[0031] Based on the complex wavelet basis function, the spectrum data F(q) is decomposed into sub-bands at JI scales, and the wavelet coefficients at any scale among the JI scales are calculated;

[0032] The calculation formula of wavelet coefficients is:

[0033] Among them, DW j,s Represents the wavelet coefficient of the spectrum data at the jth scale and the sth translation position; B j,s (q) represents the complex wavelet basis function of the jth scale and sth translation position;

[0034] The wavelet coefficients at each scale are reconstructed by inverse wavelet transform to obtain n sub-bands, that is, the spectrum data is decomposed into n sub-bands.

[0035] Furthermore, the method of screening the target frequency band includes:

[0036] Preset Entropy Threshold Where f(i) represents the average frequency of the i-th frequency band; μ represents the noise variance;

[0037] Noise variance Where EJ represents the sum of spectral entropy of all frequency bands;

[0038] Calculate the spectrum entropy of each sub-band based on the wavelet coefficients, and take the sub-band whose spectrum entropy is less than the entropy threshold TR as the target band;

[0039] The calculation formula of spectral entropy is:

[0040] Among them, E j,i represents the spectral entropy of the i-th sub-band at the j-th scale; H j,i represents the energy proportion of the i-th sub-band at the j-th scale;

[0041] The energy proportion of the i-th sub-band at the j-th scale Among them, DW j,i represents the wavelet coefficient of the i-th subband at the j-th scale;

[0042] The method of dividing the spectrum data into K intrinsic mode functions includes:

[0043] Constructing the cost function Among them, u m (q) represents the mth intrinsic mode function decomposed from the spectrum data by variable mode decomposition; δ(q) represents the Dirac function; ω mrepresents the central frequency of the mth intrinsic mode function; λ represents the regularization coefficient; ‖...‖2 represents the L2 norm;

[0044] The intrinsic mode function is optimized to minimize the value of the cost function. When the value of the cost function is minimized, the parameters of the cost function are fixed. At this time, K in the parameter is the result of variable mode decomposition, that is, it is divided into K intrinsic mode functions;

[0045] The calculation formula for optimizing the intrinsic mode function is:

[0046] in, It refers to the mth eigenmode function of the a+1th iteration; represents inverse Fourier transform; represents the Fourier transform of the mth eigenmode function at the ath iteration; represents the center frequency of the mth eigenmode function at the ath iteration; represents the mth intrinsic mode function of the ath iteration; α represents the penalty factor;

[0047] The center frequency is updated during the optimization of the eigenmode function;

[0048] The calculation formula for updating the center frequency is:

[0049] in, represents the center frequency of the mth eigenmode function at the a+1th iteration; represents the Fourier transform of the mth intrinsic mode function at the a+1th iteration.

[0050] Furthermore, the method of performing fault detection on the wind turbine generator set based on the pre-processed operating data includes:

[0051] Construct a fault identification model, use the fault identification model to process the pre-processed operation data to obtain fault classification data; perform maintenance on the wind turbines based on the fault classification data; collect the operation data of the wind turbines after maintenance, and update the fault classification data based on the operation data of the wind turbines after maintenance to obtain complete operation data;

[0052] The method of constructing the fault identification model includes:

[0053] Collect historical fault data, use clustering algorithm to detect historical fault data, obtain abnormal data, determine the fault type of abnormal data by querying the preset wind turbine database and add corresponding fault code for each fault type; use the fault code corresponding to each fault type as training label;

[0054] An adaptive algorithm is used to perform feature extraction on historical fault data to obtain optimal feature data; the optimal feature data is used to train a fault recognition model; a fault recognition loss function of the fault recognition model is defined, and the function value of the fault recognition loss function is calculated until the function value of the fault recognition loss function no longer decreases, and the parameters of the fault recognition model are fixed at this time to obtain a trained fault recognition model.

[0055] Furthermore, the method of using the adaptive algorithm to perform feature extraction processing on historical fault data includes:

[0056] Initialize population parameters, which include the number of iterations, population size, population position, flexible parameter control factor, and convergence factor; each individual in the population corresponds to a feature subset, and each feature subset corresponds to a piece of data in the historical fault data;

[0057] Define the error rate fitness function and complexity fitness function;

[0058] Error rate fitness function Among them, FP represents the total number of normal feature subsets that are misjudged as fault feature subsets; FN represents the total number of fault feature subsets that are misjudged as normal feature subsets; TP represents the total number of feature subsets that are correctly judged as fault feature subsets; TN represents the total number of feature subsets that are correctly judged as normal feature subsets; η1 and η2 represent the first-class weight factor and the second-class weight factor, respectively, and η1+η2=1;

[0059] Complexity fitness function Among them, AA represents the proportion of the number of correctly judged feature subsets to all feature subsets; AS represents the number of feature subsets of the optimal population; η3 and η4 represent three-category weight factors and four-category weight factors respectively, and η3+η4=1;

[0060] Calculate the first-class flexible parameter and the second-class flexible parameter based on the number of iterations; iterate the population position based on the first-class flexible parameter and the second-class flexible parameter to obtain the optimal population position, and minimize the error rate fitness function and the complexity fitness function based on the optimal population position; when the error rate fitness function and the complexity fitness function obtain the minimum value, the corresponding population at this time is the optimal population, that is, the optimal feature data;

[0061] The calculation formulas of the first type of flexibility parameters and the second type of flexibility parameters are:

[0062] A type of flexible parameter

[0063] Second type of flexibility parameters

[0064] Among them, ρ1 and ρ2 represent the flexible parameter control factors; Indicates the number of iterations; N max Indicates the maximum number of iterations.

[0065] Furthermore, the method of iterating the population position includes:

[0066] The population position is iterated three times, and the optimal population position is obtained at the end of the third round of iteration;

[0067] When the number of iterations In range hour, (g=1,2,...,total); where Loc g represents the position of the g-th individual in the population; ran1 represents a round of iterative random vector; total represents the population size; Represents vector multiplication operation;

[0068] Among them, step1 g represents the moving step length of the g-th individual in the first iteration; RB represents the Brownian motion strategy; Best1 g represents the global optimal individual in the first round of iteration;

[0069] When the number of iterations In range hour, Among them, ran2 represents the second round of iterative random vector; among them, CF represents the convergence factor;

[0070] Among them, step2 g Indicates the moving step length of the g-th individual in the second iteration; Best2 g represents the global optimal individual in the second round of iteration; RL represents the Levy flight strategy;

[0071] When the number of iterations In range hour, Among them, ran3 represents the random vector of three rounds of iterations;

[0072] Among them, step3 g Indicates the moving step length of the g-th individual in the population in the third iteration; Best3 g Represents the global optimal individual in the third round of iteration.

[0073] Furthermore, the method of optimizing and improving the operation data includes:

[0074] The feature extraction algorithm is used to extract features from the perfect operation data to obtain the application operation parameter combination; the optimization algorithm is used to optimize the application operation parameter combination to obtain the optimal operation parameter combination.

[0075] The technical effects and advantages of the intelligent operation and maintenance management platform for wind turbines of the present invention are as follows:

[0076] By collecting the operation data of wind turbines and accurately preprocessing the data, higher quality operation data of wind turbines is obtained. At the same time, a model is built to detect faults in the operation data of wind turbines, and wind turbines are maintained and the operation data is updated. Finally, the operation data after maintenance is optimized and the control of wind turbines is optimized to form an efficient intelligent operation and maintenance management platform for wind turbines. Compared with the existing experience, in the data preprocessing module, the time series data and spectrum data in the operation data are processed by building a prediction model and wavelet decomposition respectively, so that the data is more accurate and convenient for subsequent operations. In the fault detection module, a fault identification model is built to process the data and accurately identify the fault types of wind turbines. At the same time, an adaptive algorithm with flexible parameters is introduced in the fault identification model to obtain a higher quality training set, improve the model training efficiency, and indirectly improve the efficiency and accuracy of the model for fault identification. In the unit optimization module, the control optimization of wind turbines is completed by using feature extraction algorithms and optimization algorithms, which improves the operation efficiency of wind turbines and reduces the probability of wind turbine failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a schematic diagram of an intelligent operation and maintenance management platform for wind turbines according to the present invention;

[0078] Figure 2 This is a schematic diagram of an intelligent operation and maintenance management method for a wind turbine generator system according to the present invention. DETAILED DESCRIPTION

[0079] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0080] Example 1

[0081] See also Figure 1 As shown, the wind turbine intelligent operation and maintenance management platform described in this embodiment includes:

[0082] Data acquisition module, used to collect the operation data of wind turbines;

[0083] A data preprocessing module is used to preprocess the operation data of the wind turbine generator set to obtain preprocessed operation data;

[0084] The fault detection module detects faults of wind turbines based on pre-processed operation data, and performs maintenance on wind turbines to obtain complete operation data;

[0085] The unit optimization module optimizes the perfect operation data to obtain the optimal operation parameter combination; based on the optimal operation parameter combination, the wind turbine is controlled and optimized; each module is connected by wire and / or wireless means.

[0086] The operating data of the wind turbine is collected through the sensors installed on the wind turbine; the system records the changes in data measured by sensors installed on the wind turbine itself, such as temperature sensors, pressure sensors, vibration sensors and wind force sensors, to obtain the corresponding operating data of the wind turbine; the operating data of the wind turbine includes time series data and spectrum data; the time series data represents a sequence of data points sorted by time, and any data point in the sequence has a timestamp corresponding to it, and the time series data includes data such as wind speed, temperature and air pressure; the spectrum data represents the electrical signals generated by measuring periodic physical phenomena in the wind turbine, and converting these periodic signals into spectrum signals to generate spectrum data; for example, vibration signals measured by vibration sensors, acoustic signals measured by acoustic sensors, and current and voltage signals measured by electrical sensors.

[0087] The method of preprocessing the operation data of the wind turbine generator set includes:

[0088] Preprocess the time series data to obtain complete time series data; preprocess the spectrum data to obtain enhanced spectrum data; combine the complete time series data and the enhanced spectrum data to obtain preprocessed running data.

[0089] The method of preprocessing the time series data includes:

[0090] Construct a prediction model to process missing values ​​of time series data and obtain complete time series data.

[0091] The method of constructing a prediction model to process missing values ​​of time series data includes:

[0092] The historical time series data of the wind turbines are collected and standardized to obtain the historical time series data training set; the historical time series data training set is initially supplemented (the missing value processing methods such as linear interpolation method, polynomial interpolation method or spline interpolation method are used to preliminarily process the data set containing missing values) as the training label of the prediction model.

[0093] Initialize the hidden sequence h t , where t represents the time index and t≤max, max represents the maximum time corresponding to the data in the historical time series data training set; at this time, the hidden sequence h t =h0, t=0; h0 is an initialized zero vector.

[0094] The prediction model includes a loop update structure, a loop reset structure and an output layer. The historical time series data training set is input into the loop update structure and the loop reset structure, and the loop update output and the loop reset output are calculated respectively.

[0095] Candidate hidden outputs are calculated based on the loop reset output; the final hidden output is calculated based on the candidate hidden output and the loop update output; the final hidden output is mapped to a predicted value through the output layer.

[0096] The predicted values ​​are used to process the missing values ​​of the historical time series data training set to obtain a complete historical time series data training set; the complete historical time series data training set is compared with the training labels. If the absolute value of the difference between the numerical values ​​of the data at the corresponding moments is less than or equal to the preset difference threshold (for example, the absolute value of the difference between the numerical values ​​of the data at the corresponding moments is less than or equal to 0.5), the parameters are fixed at this time to obtain a trained prediction model; the trained prediction model is used to process the missing values ​​of the time series data to obtain complete time series data.

[0097] The loop updates the output Among them, W Z represents the update input weight matrix (used to convert the current input data into a signal that can be recognized by the loop update structure, which reflects the influence of the current input data on the loop update output); U Z represents the updated hidden weight matrix (used to convert the hidden state of the previous moment into a signal that can be recognized by the cyclic update structure, which reflects the influence of the past hidden state on the cyclic update output); X t represents the input data at time t, that is, the data at time t in the historical time series data training set; b Z h represents the bias term of the cyclic update structure (used to adjust the result of the linear combination); t-1 represents the hidden sequence at time t-1; σ represents an activation function (used to update the output Z of the loop t Controlled in the interval [0,1], if Z t If Z is close to 0, it means that the information of the previous moment is not helpful for the model to predict the missing value and can be ignored. t If it is close to 1, it means that the information of the previous moment is useful for the model to predict the missing values ​​and can be retained).

[0098] Cycle reset output Among them, W R Represents the reset input weight matrix (used to convert the current input data into a signal that can be recognized by the loop reset structure, which reflects the influence of the current input data on the loop reset output); U R represents the reset hidden weight matrix (used to convert the hidden state of the previous moment into a signal that can be recognized by the loop reset structure, which reflects the influence of the past hidden state on the loop reset output); b R Represents the bias term for the loop reset structure.

[0099] Candidate hidden output Among them, tanh represents the hyperbolic tangent function (used to ensure that the candidate hidden output The range is [-1,1]); W represents the input weight matrix of the candidate hidden output (used to convert the current input data into a signal that can represent the candidate hidden state, which reflects the influence of the current input data on the candidate hidden output); U represents the hidden weight matrix of the candidate hidden output (used to convert the hidden state at the previous moment into a signal that can represent the candidate hidden state, which reflects the influence of the past hidden state on the candidate hidden output); ⊙ represents element-by-element multiplication (specifically manifested as the Hadamard product in the candidate hidden output calculation).

[0100] Hide Sequence When t = max, h max For the final hidden output;

[0101] The method of preprocessing the spectrum data includes:

[0102] A complex wavelet basis function is constructed, and the spectrum data is decomposed based on the constructed complex wavelet basis function to obtain wavelet coefficients, and the spectrum data is decomposed into n sub-bands; and a target frequency band is selected from the n sub-bands.

[0103] Through variable mode decomposition (a signal processing method used to decompose complex signals into multiple modal components with different center frequencies), the spectral data is divided into K intrinsic mode functions (IMF functions, which are used to describe the basic vibration modes of nonlinear and non-stationary signals and are often used in the signal processing field), and the intrinsic mode functions corresponding to the target frequency band are reconstructed using inverse wavelet transform to obtain enhanced spectral data.

[0104] The method of decomposing the spectrum data based on the constructed complex wavelet basis function includes:

[0105] Complex wavelet basis function B(q)=P(n1(q)+I×n2(q)); wherein q represents the time variable corresponding to the spectrum data; n1(q) represents the real wavelet basis function; m2(q) represents the imaginary wavelet basis function; I is the imaginary unit; P represents the scale factor (used to control the scaling degree of the wavelet basis function, thereby affecting the resolution and frequency band distribution of the signal decomposition).

[0106] A preset symmetric wavelet basis function (such as the Symlets wavelet basis function) is used as the real wavelet basis function n1(q); the Hilbert transform (a linear operator used to convert a real signal into a complex signal and extract attribute information from the signal) of the real wavelet basis function n1(q) is calculated, and the calculation result is used as the imaginary wavelet basis function n2(q).

[0107] The spectrum data F(q) is decomposed into sub-bands at JI scales based on the complex wavelet basis function, and the wavelet coefficients at any one of the JI scales are calculated.

[0108] The calculation formula of wavelet coefficients is:

[0109] Among them, DW j,s The wavelet coefficient representing the spectral data at the jth scale and the sth translation position (referring to the position change of the wavelet basis function on the time axis, which is used to reflect the distance moved on the time axis when the wavelet basis function performs the convolution operation); B j,s (q) represents the complex wavelet basis function of the jth scale and sth translation position;

[0110] The wavelet coefficients at each scale are reconstructed by inverse wavelet transform to obtain n sub-bands, that is, the spectrum data is decomposed into n sub-bands.

[0111] The method of screening the target frequency band includes:

[0112] Preset Entropy Threshold Where f(i) represents the average frequency of the i-th frequency band; μ represents the noise variance;

[0113] Noise variance Where EJ represents the sum of spectral entropy of all frequency bands;

[0114] The spectral entropy of each sub-band is calculated based on the wavelet coefficients (used to describe the randomness and complexity of the spectral signal. The larger the value, the more complex and disordered the change of the spectral signal is). The sub-band whose spectral entropy is less than the entropy threshold TR is taken as the target band.

[0115] The calculation formula of spectral entropy is:

[0116] Among them, Ej,i represents the spectral entropy of the i-th sub-band at the j-th scale; H j,i represents the energy proportion of the i-th sub-band at the j-th scale;

[0117] The energy proportion of the i-th sub-band at the j-th scale Among them, DW j,i represents the wavelet coefficient of the i-th subband at the j-th scale;

[0118] The method of dividing the spectrum data into K intrinsic mode functions includes:

[0119] Constructing the cost function Among them, u m (q) represents the mth intrinsic mode function decomposed from the spectrum data by variable mode decomposition; δ(q) represents the Dirac function (a distribution used to represent and analyze signals in the signal processing field); ω m represents the center frequency of the mth intrinsic mode function (refers to the center point of the frequency range of a frequency band. Each intrinsic mode function corresponds to a center frequency, which is used to describe the frequency characteristics of the intrinsic mode function); λ represents the regularization coefficient (a hyperparameter in machine learning, used to control the strength of the regularization term to make the model parameters smoother); ‖...‖2 represents the L2 norm (used to optimize the regularization term of the objective function to prevent overfitting of the model and improve the generalization ability of the model).

[0120] The intrinsic mode function is optimized to minimize the value of the cost function. When the value of the cost function is minimized, the parameters of the cost function are fixed. At this time, K in the parameters is the result of variable mode decomposition, that is, it is divided into K intrinsic mode functions.

[0121] The calculation formula for optimizing the intrinsic mode function is:

[0122] in, It refers to the mth eigenmode function of the a+1th iteration; Represents inverse Fourier transform (converting frequency domain signal into time domain signal); The Fourier transform of the mth intrinsic mode function of the ath iteration (converting the time domain signal into the frequency domain signal to represent the frequency component and phase information of the signal); represents the center frequency of the mth eigenmode function at the ath iteration; represents the mth intrinsic mode function of the ath iteration; α represents the penalty factor (used to control the mutual interference between each intrinsic mode function); this formula ensures that each intrinsic mode function is an independent narrowband signal by minimizing the mutual interference between the intrinsic mode functions.

[0123] The center frequency is updated during the optimization of the eigenmode function.

[0124] The calculation formula for updating the center frequency is:

[0125] in, represents the center frequency of the mth eigenmode function at the a+1th iteration; represents the Fourier transform of the mth intrinsic mode function of the a+1th iteration; this formula ensures that each intrinsic mode function can accurately reflect its main frequency part by measuring the weighted average frequency of the intrinsic mode function in the frequency domain.

[0126] Only the intrinsic mode function corresponding to the target frequency band is reconstructed in order to obtain a higher quality spectrum signal. Frequency bands other than the target frequency band usually contain high-frequency noise, so these frequency bands are directly excluded.

[0127] The method of performing fault detection on the wind turbine generator set based on the pre-processed operating data includes:

[0128] Construct a fault identification model (such as an SVM support vector machine model), and use the fault identification model to process the pre-processed operating data to obtain fault classification data; maintain the wind turbine based on the fault classification data (such as preventive maintenance, timely maintenance of minor faults to prevent subsequent faults from becoming more serious; such as corrective maintenance, timely shutdown or adjustment of the wind turbine after a fault occurs to restore the normal operation of the wind turbine); collect the operating data of the wind turbine after maintenance, and update the fault classification data based on the operating data of the wind turbine after maintenance to obtain complete operating data.

[0129] The method of constructing the fault identification model includes:

[0130] Collect historical fault data (including wind turbine voltage instability, gear speed abnormality, and unit temperature abnormality) and use clustering algorithms (such as DBSCAN clustering algorithm) to detect historical fault data to obtain abnormal data. Determine the fault type of the abnormal data by querying a preset wind turbine database and add a corresponding fault code for each fault type (for example, when the fault type of a certain part of the data is identified as abnormal gear speed, set the fault code of the fault to 001A); use the fault code corresponding to each fault type as a training label.

[0131] An adaptive algorithm is used to perform feature extraction on historical fault data to obtain optimal feature data; the optimal feature data is used to train a fault recognition model; a fault recognition loss function (such as a binary cross entropy loss function) of the fault recognition model is defined, and the function value of the fault recognition loss function is calculated until the function value of the fault recognition loss function no longer decreases, and the parameters of the fault recognition model at this time are fixed to obtain a trained fault recognition model.

[0132] The method of using the adaptive algorithm to perform feature extraction processing on historical fault data includes:

[0133] Initialize population parameters, including the number of iterations, population size, population position, flexible parameter control factor (used to calculate flexible parameters. By adjusting the flexible parameters through the flexible parameter control factor, the changes in population position can be dynamically adjusted) and convergence factor (used to control the convergence speed of the algorithm so that the algorithm can more accurately adjust the population position in the later stage and improve the optimization accuracy); each individual in the population corresponds to a feature subset, and each feature subset corresponds to a piece of data in the historical fault data.

[0134] Define the error rate fitness function and the complexity fitness function; the error rate represents the probability of misjudgment during fault detection. The larger the value, the higher the probability of error, resulting in worse fault detection effect. Conversely, the higher the judgment accuracy, the better the fault detection effect. Complexity is used to evaluate the complexity of the feature subset selected during the feature extraction process. The larger the value, the higher the complexity of the extracted feature subset, and the more likely it is to cause misjudgment of fault detection. Conversely, the smaller the complexity of the feature subset, the easier it is to make a successful fault detection judgment.

[0135] Error rate fitness function Among them, FP represents the total number of normal feature subsets (representing feature subsets that are not abnormal data in historical fault data) that are misjudged as fault feature subsets (representing feature subsets that are abnormal data in historical fault data); FN represents the total number of fault feature subsets that are misjudged as normal feature subsets; TP represents the total number of feature subsets that are correctly judged as fault feature subsets; TN represents the total number of feature subsets that are correctly judged as normal feature subsets; η1 and η2 represent the first-class weight factor and the second-class weight factor, respectively (the first-class weight factor and the second-class weight factor are both constants belonging to the interval [0,1]), and η1+η2=1.

[0136] Complexity fitness function Among them, AA represents the proportion of the number of correctly judged feature subsets to all feature subsets; AS represents the number of feature subsets of the optimal population; η3 and η4 represent the three-category weight factor and the four-category weight factor respectively (the three-category weight factor and the four-category weight factor are both constants belonging to the interval [0,1]), and η3+η4=1.

[0137] The first-class flexible parameters and the second-class flexible parameters are calculated based on the number of iterations; the population position is iterated based on the first-class flexible parameters and the second-class flexible parameters to obtain the optimal population position, and the error rate fitness function and the complexity fitness function are minimized based on the optimal population position; when the error rate fitness function and the complexity fitness function reach the minimum value, the corresponding population at this time is the optimal population, that is, the optimal feature data.

[0138] The calculation formulas of the first type of flexibility parameters and the second type of flexibility parameters are:

[0139] A type of flexible parameter

[0140] Second type of flexibility parameters

[0141] Among them, ρ1 and ρ2 represent the flexible parameter control factors (ρ1, ρ2∈(0,+∞)); Indicates the number of iterations; N max Indicates the maximum number of iterations.

[0142] The method of iterating the population position includes:

[0143] The population position is iterated three times, and the optimal population position is obtained at the end of the third round of iteration.

[0144] When the number of iterations In range hour, (g=1,2,...,total); where Loc g Indicates the position of the g-th individual in the population; ran1 represents a round of iterative random vector (composed of a random number of integers in the interval (0, +∞)); total represents the population size; Represents a vector multiplication operation.

[0145] Among them, step1 g represents the moving step of the g-th individual in the population in the first iteration (indicates the distance the individual moves in the search space); RB represents the Brownian motion strategy (indicates a set of vectors based on standard random distribution, used for Brownian motion in a small range); Best1 g Represents the global optimal individual in the first round of iteration (the individual in the population that minimizes the fitness function value).

[0146] When the number of iterations In range hour, Among them, ran2 represents the second-round iterative random vector (composed of a random number of constants belonging to the interval (0,1)); among them, CF stands for convergence factor.

[0147] Among them, step2 g Indicates the moving step length of the g-th individual in the second iteration; Best2 g Represents the global optimal individual in the second round of iteration; RL represents the Levy flight strategy (represents a set of vectors based on Levy distribution, used for a large range of Levy flight movements).

[0148] When the number of iterations In range hour, Among them, ran3 represents a three-round iterative random vector (composed of a random number of integers in the interval (1, +∞)).

[0149] Among them, step3 g Indicates the moving step length of the g-th individual in the population in the third iteration; Best3 g Represents the global optimal individual in the third round of iteration.

[0150] The method for optimizing and improving the operation data includes:

[0151] A feature extraction algorithm (such as a principal component analysis algorithm) is used to extract features from the perfect operation data to obtain an application operation parameter combination; an optimization algorithm (such as a particle swarm algorithm or an ant colony algorithm) is used to optimize the application operation parameter combination to obtain the optimal operation parameter combination.

[0152] This embodiment acquires the operation data of the wind turbine set and accurately pre-processes the data to obtain higher quality operation data of the wind turbine set, and at the same time builds a model to detect faults in the operation data of the wind turbine set, maintains the wind turbine set, and updates the operation data; finally, the operation data after maintenance is optimized and the control of the wind turbine set is optimized to form an efficient intelligent operation and maintenance management platform for the wind turbine set; compared with the existing experience, in the data pre-processing module, the time series data and the spectrum data in the operation data are processed by respectively building a prediction model and wavelet decomposition, so that the data is more accurate and convenient for subsequent operations; in the fault detection module, a fault identification model is built to process the data and accurately identify the fault category of the wind turbine set; at the same time, an adaptive algorithm with flexible parameters is introduced into the fault identification model to obtain a higher quality training set, improve the model training efficiency, and indirectly improve the efficiency and accuracy of the model for fault identification; in the unit optimization module, the control optimization of the wind turbine set is completed by using the feature extraction algorithm and the optimization algorithm, so as to improve the operation efficiency of the wind turbine set and reduce the probability of failure of the wind turbine set.

[0153] Example 2

[0154] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an intelligent operation and maintenance management method for a wind turbine generator set, including:

[0155] S1. Collecting the operation data of wind turbines;

[0156] S2. Preprocessing the operation data of the wind turbine to obtain preprocessed operation data;

[0157] S3. Perform fault detection on the wind turbine generator set based on the pre-processed operation data, and perform maintenance on the wind turbine generator set to obtain complete operation data;

[0158] S4. Optimize the perfect operation data to obtain the optimal operation parameter combination; and optimize the control of the wind turbine based on the optimal operation parameter combination;

[0159] Example 3

[0160] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the intelligent operation and maintenance management method of a wind turbine set provided above is implemented.

[0161] Since the electronic device introduced in this embodiment is an electronic device used to implement a wind turbine intelligent operation and maintenance management method in the embodiment of this application, based on the wind turbine intelligent operation and maintenance management method introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not described in detail here. As long as the technical personnel of this field implement the electronic device used in the wind turbine intelligent operation and maintenance management method in the embodiment of this application, it belongs to the scope of protection of this application.

[0162] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0163] 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 technical users 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. An intelligent operation and maintenance management platform for wind turbines, characterized in that: include: Data acquisition module, used to collect the operation data of wind turbines; A data preprocessing module is used to preprocess the operation data of the wind turbine generator set to obtain preprocessed operation data; The fault detection module detects faults of wind turbines based on pre-processed operation data, and performs maintenance on wind turbines to obtain complete operation data; The unit optimization module optimizes the perfect operation data to obtain the optimal operation parameter combination; based on the optimal operation parameter combination, the wind turbine is controlled and optimized; each module is connected by wire and / or wireless means.

2. The intelligent operation and maintenance management platform for wind turbines according to claim 1 is characterized in that: The operating data of the wind turbine generator set is collected by sensors installed on the wind turbine generator set; the operating data of the wind turbine generator set includes time series data and spectrum data.

3. The intelligent operation and maintenance management platform for wind turbines according to claim 2 is characterized in that: The method of preprocessing the operation data of the wind turbine generator set includes: Preprocessing the time series data to obtain complete time series data; preprocessing the spectrum data to obtain enhanced spectrum data; combining the complete time series data and the enhanced spectrum data to obtain preprocessed running data; The method of preprocessing the time series data includes: Construct a prediction model to process missing values ​​of time series data and obtain complete time series data; The method of constructing a prediction model to process missing values ​​of time series data includes: Collect historical time series data of wind turbines, and perform standardization on the historical time series data of wind turbines to obtain a training set of historical time series data; initially complete the training set of historical time series data as training labels for the prediction model; Initialize the hidden sequence h t , where t represents the time index and t≤max, and max represents the maximum time corresponding to the data in the historical time series data training set; The prediction model includes a loop update structure, a loop reset structure and an output layer, and the historical time series data training set is input into the loop update structure and the loop reset structure to calculate the loop update output and the loop reset output respectively; Calculate candidate hidden outputs based on the loop reset output; calculate final hidden outputs based on the candidate hidden outputs and the loop update output; map the final hidden output to a predicted value through the output layer; Use the predicted value to process the missing values ​​of the historical time series data training set to obtain a complete historical time series data training set; compare the complete historical time series data training set with the training label. If the absolute value of the difference between the numerical values ​​of the data at the corresponding time is less than or equal to the preset difference threshold, fix the parameters of the prediction model at this time to obtain a trained prediction model; The loop updates the output Among them, W Z Represents updating the input weight matrix: U Z represents the update of hidden weight matrix; X t represents the input data at time t, that is, the data at time t in the historical time series data training set; b Z represents the bias term of the loop update structure; h t-1 represents the hidden sequence at time t-1; σ represents an activation function; Cycle reset output Among them, W R Represents resetting the input weight matrix; U R Represents resetting the hidden weight matrix; b R A bias term representing the loop reset structure; Candidate hidden output Where tanh represents the hyperbolic tangent function; W represents the input weight matrix of the candidate hidden output; U represents the hidden weight matrix of the candidate hidden output; ⊙ represents element-by-element multiplication; Hide Sequence When t = max, h max For the final hidden output.

4. The intelligent operation and maintenance management platform for wind turbines according to claim 3 is characterized in that: The method of preprocessing the spectrum data includes: Constructing a complex wavelet basis function, and decomposing the spectrum data based on the constructed complex wavelet basis function to obtain wavelet coefficients, and decomposing the spectrum data into n sub-bands; selecting a target frequency band from the n sub-bands; The spectrum data is divided into K intrinsic mode functions through variable mode decomposition, and the intrinsic mode functions corresponding to the target frequency band are reconstructed using inverse wavelet transform to obtain enhanced spectrum data.

5. The intelligent operation and maintenance management platform for wind turbines according to claim 4 is characterized in that: The method of decomposing the spectrum data based on the constructed complex wavelet basis function includes: Complex wavelet basis function B(q)=P(n1(q)+I×n2(q)); where q represents the time variable corresponding to the spectrum data; n1(q) represents the real wavelet basis function; n2(q) represents the imaginary wavelet basis function; I is the imaginary unit; P represents the scale factor; The preset symmetric wavelet basis function is used as the real wavelet basis function n1(q); the Hilbert transform of the real wavelet basis function n1(q) is calculated, and the calculation result is used as the imaginary wavelet basis function n2(q); Based on the complex wavelet basis function, the spectrum data F(q) is decomposed into sub-bands at JI scales, and the wavelet coefficients at any scale among the JI scales are calculated; The calculation formula of wavelet coefficients is: Among them, DW j，s Represents the wavelet coefficient of the spectrum data at the jth scale and the sth translation position; B j，s (q) represents the complex wavelet basis function of the jth scale and sth translation position; The wavelet coefficients at each scale are reconstructed by inverse wavelet transform to obtain n sub-bands, that is, the spectrum data is decomposed into n sub-bands.

6. The intelligent operation and maintenance management platform for wind turbines according to claim 5, characterized in that: The method of screening the target frequency band includes: Preset Entropy Threshold Where f(i) represents the average frequency of the i-th frequency band; μ represents the noise variance; Noise variance Where EJ represents the sum of spectral entropy of all frequency bands; Calculate the spectrum entropy of each sub-band based on the wavelet coefficients, and take the sub-band whose spectrum entropy is less than the entropy threshold TR as the target band; The calculation formula of spectral entropy is: Among them, E j，i represents the spectral entropy of the i-th sub-band at the j-th scale; H j，i represents the energy proportion of the i-th sub-band at the j-th scale; The energy proportion of the i-th sub-band at the j-th scale Among them, DW j，i represents the wavelet coefficient of the i-th subband at the j-th scale; The method of dividing the spectrum data into K intrinsic mode functions includes: Constructing the cost function Among them, u m (q) represents the mth eigenmode function decomposed from the spectrum data by variable mode decomposition; δ(q) represents the Dirac function; ω m represents the central frequency of the mth intrinsic mode function; λ represents the regularization coefficient; ||...||2 represents the L2 norm; The intrinsic mode function is optimized to minimize the value of the cost function. When the value of the cost function is minimized, the parameters of the cost function are fixed. At this time, K in the parameter is the result of variable mode decomposition, that is, it is divided into K intrinsic mode functions; The calculation formula for optimizing the intrinsic mode function is: in, It refers to the mth eigenmode function of the a+1th iteration; represents inverse Fourier transform; represents the Fourier transform of the mth eigenmode function at the ath iteration; represents the center frequency of the mth eigenmode function at the ath iteration; represents the mth intrinsic mode function of the ath iteration; α represents the penalty factor; The center frequency is updated during the optimization of the eigenmode function; The calculation formula for updating the center frequency is: in, represents the center frequency of the mth eigenmode function at the a+1th iteration; represents the Fourier transform of the mth intrinsic mode function at the a+1th iteration.

7. The intelligent operation and maintenance management platform for wind turbines according to claim 6, characterized in that: The method of performing fault detection on the wind turbine generator set based on the pre-processed operating data includes: Construct a fault identification model, use the fault identification model to process the pre-processed operation data to obtain fault classification data; perform maintenance on the wind turbines based on the fault classification data; collect the operation data of the wind turbines after maintenance, and update the fault classification data based on the operation data of the wind turbines after maintenance to obtain complete operation data; The method of constructing the fault identification model includes: Collect historical fault data, use clustering algorithm to detect historical fault data, obtain abnormal data, determine the fault type of abnormal data by querying the preset wind turbine database and add corresponding fault code for each fault type; use the fault code corresponding to each fault type as training label; An adaptive algorithm is used to perform feature extraction on historical fault data to obtain optimal feature data; the optimal feature data is used to train a fault recognition model; a fault recognition loss function of the fault recognition model is defined, and the function value of the fault recognition loss function is calculated until the function value of the fault recognition loss function no longer decreases, and the parameters of the fault recognition model are fixed at this time to obtain a trained fault recognition model.

8. The intelligent operation and maintenance management platform for wind turbines according to claim 7, characterized in that: The method of using the adaptive algorithm to perform feature extraction processing on historical fault data includes: Initialize population parameters, which include the number of iterations, population size, population position, flexible parameter control factor, and convergence factor; each individual in the population corresponds to a feature subset, and each feature subset corresponds to a piece of data in the historical fault data; Define the error rate fitness function and complexity fitness function; Error rate fitness function Among them, FP represents the total number of normal feature subsets that are misjudged as fault feature subsets; FN represents the total number of fault feature subsets that are misjudged as normal feature subsets; TP represents the total number of feature subsets that are correctly judged as fault feature subsets; TN represents the total number of feature subsets that are correctly judged as normal feature subsets; η1 and η2 represent the first-class weight factor and the second-class weight factor, respectively, and η1+η2=1; Complexity fitness function Among them, AA represents the proportion of the number of correctly judged feature subsets to all feature subsets; AS represents the number of feature subsets of the optimal population; η3 and η4 represent three-category weight factors and four-category weight factors respectively, and η3+η4=1; Calculate the first-class flexible parameter and the second-class flexible parameter based on the number of iterations; iterate the population position based on the first-class flexible parameter and the second-class flexible parameter to obtain the optimal population position, and minimize the error rate fitness function and the complexity fitness function based on the optimal population position; when the error rate fitness function and the complexity fitness function obtain the minimum value, the corresponding population at this time is the optimal population, that is, the optimal feature data; The calculation formulas of the first type of flexibility parameters and the second type of flexibility parameters are: A type of flexible parameter Second type of flexibility parameters Among them, ρ1 and ρ2 represent the flexible parameter control factors; Indicates the number of iterations; N max Indicates the maximum number of iterations.

9. The intelligent operation and maintenance management platform for wind turbines according to claim 8, characterized in that: The method of iterating the population position includes: The population position is iterated three times, and the optimal population position is obtained at the end of the third round of iteration; When the number of iterations In range hour, (g = 1, 2, ..., total); where Loc g represents the position of the g-th individual in the population; ran1 represents a round of iterative random vector; total represents the population size; Represents vector multiplication operation; Among them, step1 g represents the moving step length of the g-th individual in the first iteration; RB represents the Brownian motion strategy; Best1 g represents the global optimal individual in the first round of iteration; When the number of iterations In range hour, Among them, ran2 represents the second round of iterative random vector; among them, CF represents the convergence factor; Among them, step2 g Indicates the moving step length of the g-th individual in the second iteration; Best2 g represents the global optimal individual in the second round of iteration; RL represents the Levy flight strategy; When the number of iterations In range hour, Among them, ran3 represents the random vector of three rounds of iterations; Among them, step3 g Indicates the moving step length of the g-th individual in the population in the third iteration; Best3 g Represents the global optimal individual in the third round of iteration.

10. The intelligent operation and maintenance management platform for wind turbines according to claim 9, characterized in that: The method for optimizing and improving the operation data includes: The feature extraction algorithm is used to extract features from the perfect operation data to obtain the application operation parameter combination; the optimization algorithm is used to optimize the application operation parameter combination to obtain the optimal operation parameter combination.

Images