A method and system for monitoring blade failure of a wind turbine generator set

Abstract

The present invention discloses a method and system for monitoring blade faults of a wind turbine generator set, the method comprising: collecting noise signals through a noise sensor installed inside the blade of the wind turbine generator set; performing data preprocessing on the noise signals; decomposing the preprocessed noise signals into a limited number of intrinsic modal function components through variational modal decomposition; and performing fault diagnosis through the intrinsic modal function components. When a blade cracks, the method of the present invention can monitor the damage of the blade through the change of the noise inside the blade. The wind turbine generator set blade fault monitoring method proposed by the present invention has a comprehensive monitoring range, can realize all-weather online monitoring, and generates less data, is easy to process, and has lower algorithm development difficulty and application cost.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine generator sets, and in particular to a method and system for monitoring blade faults of wind turbine generator sets. Background Art

[0002] Since wind turbine blades are made of composite materials such as wood and fiberglass, they are poured into molds and then assembled by mold closing. As the length and weight of wind turbine blades increase and they move at high speeds in harsh environments for a long time, cracks in wind turbine blades are common. In order to detect faults in wind turbine blades in a timely manner and prevent them from affecting the operation of wind turbines, it is necessary to monitor the blades of wind turbines.

[0003] The existing wind turbine blade monitoring solutions mainly include optical monitoring, drone monitoring and load monitoring. However, the optical monitoring solution cannot accurately identify blade tip faults and internal crack faults of the blades; drone monitoring is only suitable for inspection functions and cannot achieve long-term online monitoring; load monitoring requires the installation of more strain and acceleration sensors, and the calibration of the blade aerodynamic model and static load, which has the problem of high cost and cumbersome data processing. Therefore, there is an urgent need for a comprehensive and simpler solution for long-term online monitoring of wind turbine blades. Summary of the invention

[0004] In order to overcome the problems existing in the related art, the present invention provides a wind turbine blade fault monitoring method and system to solve the problems of wind turbine blade monitoring solutions in the prior art, such as either limited monitoring range and inability to perform long-term online monitoring, or high deployment cost and cumbersome data processing.

[0005] According to a first aspect of the present invention, a method for monitoring blade failures of a wind turbine is provided, the method comprising:

[0006] Collecting noise signals through a noise sensor installed inside the blades of the wind turbine generator set;

[0007] Performing data preprocessing on the noise signal;

[0008] Decomposing the preprocessed noise signal into a finite number of intrinsic mode function components by variational mode decomposition;

[0009] Fault diagnosis is performed using a number of the intrinsic mode function components.

[0010] Preferably, the noise sensor is an optical fiber MEMS noise sensor.

[0011] Preferably, the noise sensor is installed in the following manner:

[0012] Inside the blade of the wind turbine generator set, starting from the blade root, one of the noise sensors is installed at a specified distance; and the number and relative positions of the noise sensors installed on each blade of the wind turbine generator set remain consistent.

[0013] Preferably, performing data preprocessing on the noise signal includes performing data framing on the noise signal; performing data framing on the noise signal includes:

[0014] Taking the wind rotor as the center, the motion trajectory of the wind turbine blade on the wind rotor rotation plane is divided into a number of angle ranges, and the noise signal collected in each angle range is marked with a corresponding angle label; and the angle label marked on each blade of the wind turbine is different;

[0015] The noise signal is framed according to the angle label.

[0016] Preferably, performing data preprocessing on the noise signal includes performing data filtering on the noise signal using a Kalman filter; performing data filtering on the noise signal using a Kalman filter includes:

[0017] The state equation of the wind turbine generator set collecting the noise signal is:

[0018] x t =Ax t-1 +Bu t-1 +ω t-1

[0019] p(ω)~N(0,Q)

[0020] Among them, x t-1 is the state vector at time t-1, u t-1 is the input vector at time t-1, ω t-1 is process noise, A is the state transfer coefficient matrix of the noise signal, B is an optional control input matrix, Q is the covariance matrix of the process noise, and the process noise satisfies Gaussian distribution;

[0021] The measurement equation of the noise signal is:

[0022] z t =Hx t +v t

[0023] p(v)~N(0,R)

[0024] Among them, z t is the measurement vector at time t, v tis the corresponding measurement error, H is the measurement coefficient matrix, R is the covariance matrix of the measurement error, and the measurement error satisfies Gaussian distribution;

[0025] Based on Kalman filtering, the time update equation of the noise signal is:

[0026]

[0027]

[0028] in, is the predicted value of the state vector, and P is the covariance matrix of the error;

[0029] Based on Kalman filtering, the measurement update equation of the noise signal is:

[0030]

[0031]

[0032]

[0033] Where K is the Kalman filter gain.

[0034] Preferably, the fault diagnosis is performed by using a plurality of the intrinsic mode function components, including:

[0035] Inputting a plurality of said intrinsic mode function components into a specified long short-term memory network to obtain a fault prediction result; wherein said long short-term memory network is obtained in advance by training a data set consisting of a plurality of fault signals and healthy signals;

[0036] If the fault prediction result indicates that a fault has occurred, a fault warning is initiated to prompt fault handling.

[0037] Preferably, the long short-term memory network comprises an input layer, a hidden layer and an output layer; wherein the hidden layer comprises a memory unit and a control unit, and the control unit comprises an input gate, a forget gate and an output gate;

[0038] The long short-term memory network performs the following calculations in sequence during the forward propagation process:

[0039] The value of the forget gate is calculated and updated, and the formula is:

[0040] f (t) =σ(W f h (t-1) +W i x (t) +b f )

[0041] Among them, σ is the sigmoid activation function, W f is the forget gate weight, b f is the forget gate threshold;

[0042] The value of the input gate is calculated and updated as follows:

[0043] i (t) =σ(W i h (t-1) +W i x (t) +b i )

[0044] c ,(t) =tanh(W c h (t-1) +W c x (t) +b c )

[0045] Among them, W i , W c , b i and b c are the weights and thresholds corresponding to the sigmoid activation function and tanh activation function of the input gate respectively;

[0046] The value of the memory cell is calculated and updated, and the formula is:

[0047]

[0048] in, is the dot product;

[0049] The value of the output gate is calculated and updated as follows:

[0050] o (t) =σ(W o h (t-1) +W o x (t) +b o )

[0051]

[0052] Among them, W o is the output gate weight, b o is the output gate threshold, h (t) is the output vector of the hidden layer;

[0053] Update the prediction output at the current moment. The formula is:

[0054]

[0055] Wherein, V and c are the weight and threshold of the connection from the hidden layer to the output layer respectively.

[0056] Preferably, in the process of decomposing the preprocessed noise signal into a finite number of intrinsic mode function components by variational mode decomposition, the variational mode decomposition formula is:

[0057]

[0058]

[0059] Where, f is the input noise signal, u k are the K intrinsic mode function components obtained after decomposition, K is the number of the intrinsic mode function components, ω k is the center frequency of the natural mode function component, δ(t) is the impulse function;

[0060] The Lagrange multiplication operator is introduced into the variational mode decomposition formula for solution, and its expression is:

[0061]

[0062] Among them, λ is the Lagrange multiplication operator, α is the penalty factor, f(t) is the noise signal to be decomposed, and u k (t) is the kth intrinsic mode function component;

[0063] After introducing the Lagrange multiplication operator, the solution method of the variational mode decomposition formula is:

[0064] Initialize the intrinsic mode function components The center frequency of the natural mode function component and the Lagrange multiplication operator

[0065] Initialize the number of iterations n = 0, perform several iterations with n = n + 1, and in each iteration, calculate the intrinsic mode function component by the following formula The center frequency of the natural mode function component and the Lagrange multiplication operator To update:

[0066]

[0067]

[0068]

[0069] Where, τ is the noise tolerance;

[0070] According to the preset precision value ε, the iterative calculation stops when the following conditions are met:

[0071]

[0072] Output several intrinsic mode function components obtained by decomposition.

[0073] Preferably, performing data preprocessing on the noise signal includes performing data screening on the noise signal; performing data screening on the noise signal includes at least one of the following:

[0074] Determine the operating state of the wind turbine according to the wind turbine grid-connected signal, and filter the noise signal collected by the wind turbine in the non-grid-connected state;

[0075] It is determined whether the wind turbine set is in a yaw state according to the wind turbine yaw signal, and noise signals collected by the wind turbine set in the yaw state are filtered.

[0076] According to a second aspect of the present invention, there is provided a wind turbine blade fault monitoring system, the system comprising a plurality of noise sensors installed inside the wind turbine blade, a signal modem, a signal receiving station and a monitoring center;

[0077] The noise sensor is used to collect the noise signal generated by the blades of the wind turbine;

[0078] The signal modem is used to receive the noise signal collected by the noise sensor, and transmit the noise signal to the signal receiving station via a wireless connection;

[0079] The signal receiving station is used to transmit the noise signal to the monitoring center through the wind turbine ring network;

[0080] The monitoring center is used to store the noise signal and perform fault monitoring on the wind turbine blades by using the wind turbine blade fault detection method as described in any embodiment of the present invention.

[0081] The present invention discloses a method and system for monitoring blade faults of a wind turbine generator set. The method and system monitor the noise signals generated by various parts of the blade by installing a noise sensor inside the blade to identify the fault condition of the blade, and the monitoring range is comprehensive. Moreover, since the noise sensor is installed inside the blade, it is not affected by the external environment, and all-weather online monitoring can be achieved. Moreover, the present invention only performs fault diagnosis through the noise signals collected by the noise sensor. Compared with the existing load monitoring scheme, the amount of data generated is less, the data type is single, the processing is more convenient, and the algorithm development difficulty and application cost are lower.

[0082] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 The present invention is a flow chart of a method for monitoring blade failure of a wind turbine according to an embodiment of the present invention.

[0084] Figure 2 It is a schematic diagram of the structure of a long short-term memory network according to an embodiment of the present invention.

[0085] Figure 3 The figure is a schematic structural diagram of a wind turbine blade fault monitoring system according to an embodiment of the present invention.

[0086] Figure 4 The present invention is a flowchart of an operation of a wind turbine blade fault detection system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0087] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0088] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0089] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0090] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0091] like Figure 1 As shown, Figure 1The present invention is a flow chart of a method for monitoring blade failure of a wind turbine according to an embodiment of the present invention, comprising the following steps:

[0092] Step S101, collecting noise signals through a noise sensor installed inside a wind turbine blade;

[0093] Step S102, performing data preprocessing on the noise signal;

[0094] Step S103, decomposing the preprocessed noise signal into a finite number of intrinsic mode function components by variational mode decomposition;

[0095] Step S104: performing fault diagnosis through a plurality of intrinsic mode function components.

[0096] In step S101, a noise sensor can be installed inside the blade of a wind turbine to collect the noise information emitted by the blade when the wind turbine is running in real time, generate a noise signal, and the noise signal can be used to determine whether the blade of the wind turbine is faulty. Specifically, in order to avoid the influence of the state or difference of the blade itself on the judgment of the noise signal, before applying the wind turbine blade fault monitoring method of the present invention, it is necessary to perform a blade status check on all blades of the wind turbine to ensure that each blade using the same noise identification method is of the same model and in good condition.

[0097] Specifically, the noise sensor used in the present invention can be a fiber optic MEMS (Micro Electromechanical System) noise sensor. The fiber optic MEMS noise sensor refers to a noise sensor made by using a MEMS process to manufacture a sensor chip and packaging it with a dual fiber collimator. Among them, the MEMS chip has the characteristics of small size, stable performance and good parameter consistency; and combined with the fiber optic sensing technology, it overcomes the problem of mutual restriction between the "broadband" and "high precision" of existing sensors. Compared with traditional sensors, the fiber optic MEMS noise sensor has the advantages of electrical insulation of the sensing end, no need for power supply, anti-electromagnetic interference, small size, long transmission distance, etc., and is more suitable for collecting noise signals inside the blades of wind turbines. Specifically, in some embodiments, the noise sensor used can also be other MEMS noise sensors, other fiber optic noise sensors, or other types of noise sensors, as long as it can be installed inside the blades of wind turbines and can realize the transmission of noise signals between wind turbines, and the present invention does not limit this.

[0098] Specifically, the noise sensor can be installed in the following manner: inside the wind turbine blade, starting from the blade root, one noise sensor is installed at a specified distance; and the number and relative positions of the noise sensors installed on each blade of the wind turbine are kept consistent.

[0099] Specifically, when installing noise sensors, one noise sensor can be installed at a specified distance inside the wind turbine blade, starting from the blade root, so as to reduce the number of noise sensors installed as much as possible while ensuring that a sufficient relative distance is maintained between each noise sensor so that the noise signals collected by all noise sensors are more comprehensive. Specifically, the above-mentioned specified distance can be 10m, that is, one noise sensor can be installed at a distance of 10m starting from the blade root; specifically, the distance between noise sensors can also be set to other values, which is not limited by the present invention.

[0100] Specifically, when installing noise sensors, the number and relative positions of noise sensors installed on each blade of the wind turbine can be kept consistent, so as to avoid the installation differences of noise sensors on different blades affecting the analysis of noise signals and the judgment of fault conditions. Specifically, the installation positions of all noise sensors can be marked to determine and ensure whether the number and relative positions of noise sensors installed on each blade are consistent based on the position marks.

[0101] Specifically, the noise sensor may be installed inside the blade in other installation ways, as long as the installed noise sensor can collect all noise information generated inside the blade, and the present invention is not limited to this.

[0102] Specifically, a signal modem can be set at the bottom of each blade of the wind turbine set, and a signal receiving station can be set under each wind turbine set, such as at the bottom of the tower or the cabin. The signal modem on each blade of the wind turbine set is connected to the signal receiving station under the wind turbine set through wireless communication, and each signal receiving station is connected to the monitoring center through the wind turbine set ring network; therefore, all noise sensors can be connected to the signal modem of the corresponding blade, and then the noise signals collected by them can be transmitted to the signal receiving station under the wind turbine set through wireless transmission. The signal receiving station then transmits the noise signal data to the monitoring center through the wind turbine set ring network. The monitoring center stores the noise signal data and determines whether the blade has a fault based on the noise signal.

[0103] In step S102, before analyzing the noise signal collected by the noise sensor to determine the fault condition, the noise signal may be first processed by data screening, data filtering, and data framing to reduce the impact of errors generated by the noise signal during the collection and transmission process, thereby improving the accuracy of the noise signal analysis results and the accuracy of fault condition judgment.

[0104] Specifically, since the wind turbine will generate additional noise when it is not in operation or in a yaw state, and the noise signals collected from these additional noises cannot be used to analyze the fault state, before analyzing the noise signals, data screening can be performed to retain available data, such as unit data in a grid-connected state, and exclude non-available data, such as unit data in a non-grid-connected state and unit data in a yaw state.

[0105] Specifically, performing data preprocessing on the noise signal may include performing data screening on the noise signal; performing data screening on the noise signal may include at least one of the following:

[0106] Determine the operating status of the wind turbine according to the wind turbine grid-connected signal, and filter the noise signals collected by the wind turbine in the non-grid-connected state;

[0107] It is determined whether the wind turbine is in a yaw state according to the wind turbine yaw signal, and the noise signal collected by the wind turbine in the yaw state is filtered.

[0108] Specifically, when performing data screening on noise signals, other non-usable data not mentioned in the specification of the present invention may also be excluded, and the present invention does not limit the type of data screened.

[0109] Specifically, since wind turbines generate aerodynamic noise during operation and are mixed with other background environmental noise, in order to improve the quality of the noise signal, data filtering can be used to remove the interfering noise characteristics in the noise signal and improve the signal-to-noise ratio of the noise signal.

[0110] Specifically, since Kalman filtering is more suitable for non-stationary random processes, data preprocessing of the noise signal may include data filtering of the noise signal, for example, using Kalman filtering to perform data filtering on the noise signal; using Kalman filtering to perform data filtering on the noise signal may include:

[0111] The state equation of the wind turbine generator set collecting the noise signal is:

[0112] x t =Ax t-1 +Bu t-1 +ω t-1

[0113] p(ω)~N(0,Q)

[0114] Among them, x t-1 is the state vector at time t-1, u t-1 is the input vector at time t-1, ω t-1is the process noise, A is the state transfer coefficient matrix of the noise signal, B is the optional control input matrix, Q is the covariance matrix of the process noise, and the process noise satisfies the Gaussian distribution;

[0115] The measurement equation of the noise signal is:

[0116] z t =Hx t +v t

[0117] p(v)~N(0,R)

[0118] Among them, z t is the measurement vector at time t, v t is the corresponding measurement error, H is the measurement coefficient matrix, R is the covariance matrix of the measurement error, and the measurement error satisfies the Gaussian distribution;

[0119] In the present invention, the process of Kalman filtering is mainly divided into two parts: time update and measurement update. Based on Kalman filtering, the time update equation of the noise signal is:

[0120]

[0121]

[0122] in, is the predicted value of the state vector, and P is the covariance matrix of the error; the above formula indicates that the predicted value of the state vector at time t is updated using the estimated value of the covariance matrix of the state vector and the error at time t-1, that is, the state predicted value is corrected using the measured data;

[0123] Based on Kalman filtering, the measurement update equation of the noise signal is:

[0124]

[0125]

[0126]

[0127] Where K is the Kalman filter gain. Where the difference between the measurement vector and the predicted value of the measurement vector predicted based on the predicted value of the state vector in the previous step is It is called the new information, and its product with the Kalman gain K is used to correct the error of the predicted value of the state vector. The smaller the Kalman gain K is, the more the result of the filter estimation trusts the predicted value of the state vector. Conversely, it means that the result of the filter estimation is more dependent on the measured value. It can be seen that the Kalman filter uses the predicted value of the state vector of the current estimation system. Estimate the next system state This estimate is then corrected using the actual measured output.

[0128] Specifically, when performing data filtering on the noise signal, other filters may also be used for data filtering, and the present invention does not limit this.

[0129] Specifically, since the rotation of the wind turbine blades is a non-stationary process, the noise signal collected by the noise sensor is also a non-stationary signal, but it can be considered that its frequency is relatively fixed in a short time range and has a short-time stable characteristic. Therefore, before decomposing the noise signal, the noise signal can be framed according to the operating law of the wind turbine, so as to perform fault analysis based on the noise signal of each frame.

[0130] Specifically, due to the periodic rotation of the fan impeller, the gravity load factor and the uneven wind load caused by wind shear, the fan blades are subjected to non-steady impacts; and because the rotation speed of the wind wheel is time-varying, the noise signal collected by the noise sensor still fluctuates in a short period of time. The motion trajectory of the blades on the wind wheel rotation plane of the fan is a circle with the wind wheel as the center. Therefore, a starting point can be selected on the wind wheel rotation plane, such as marking the top of the wind wheel rotation plane as 0°, and then dividing it into multiple angle ranges clockwise with the wind wheel as the center, and marking each angle range as a label for the noise signal for data framing. In addition, due to the process differences of the three blades of the fan, the three blades need to be marked separately when performing data framing.

[0131] Specifically, data preprocessing of the noise signal may include data framing of the noise signal; data framing of the noise signal may include: taking the wind rotor as the center, dividing the angles of the motion trajectory of the wind turbine blades on the wind rotor rotation plane into a number of angle ranges, and marking the corresponding angle label for the noise signal collected in each angle range; and the angle label marked on the noise signal on each blade of the wind turbine is different; and then data framing the noise signal according to the angle label.

[0132] Specifically, each angle range may include only one angle, or may include a plurality of consecutive angles. The specific number of angles included may be set according to the granularity of dividing the data as required, and the present invention does not impose any limitation on this.

[0133] Specifically, in order to ensure that the sound characteristic parameters of the noise signal are not lost and are smooth, an overlapping frame method can be used when data is framed.

[0134] Specifically, when performing data framing on the noise signal, the data may also be divided according to other rules, which is not limited in the present invention.

[0135] In step S103, the preprocessed noise signal can be decomposed into a limited number of intrinsic mode function (IMF) components through variational mode decomposition (VMD), so as to determine the fault condition of the wind turbine according to the characteristics of each IMF component.

[0136] VMD is an adaptive, completely non-recursive method for modal variation and signal processing. VMD has the advantage of being able to determine the number of modal decompositions. Its adaptability is reflected in the determination of the number of modal decompositions of a given sequence based on actual conditions. In the subsequent search and solution process, it can adaptively match the optimal center frequency and limited bandwidth of each mode, and can achieve effective separation of intrinsic modal components (IMFs), frequency domain division of signals, and then obtain effective decomposition components of a given signal, and finally obtain the optimal solution to the variational problem. VMD overcomes the problems of endpoint effects and modal component aliasing in the EMD (Empirical Mode Decomposition) method, and has a more solid mathematical theoretical foundation. It can reduce the non-stationarity of time series with high complexity and strong nonlinearity, and decompose to obtain relatively stable subsequences containing multiple different frequency scales, which is suitable for non-stationary sequences.

[0137] Specifically, in the process of decomposing the preprocessed noise signal into a limited number of IFM components through VMD, the formula of its variational mode decomposition can be:

[0138]

[0139]

[0140] Where, f is the input noise signal, u k are the K intrinsic mode function components obtained after decomposition, K is the number of intrinsic mode function components, ω k is the center frequency of the natural mode function component, δ(t) is the impulse function;

[0141] The Lagrange multiplication operator (Lagrange) is introduced into the variational mode decomposition formula for solution, and its expression is:

[0142]

[0143] Among them, λ is the Lagrange multiplication operator, α is the penalty factor, f(t) is the noise signal to be decomposed, and u k (t) is the kth intrinsic mode function component;

[0144] After introducing the Lagrange multiplication operator, the solution to the variational mode decomposition formula is:

[0145] Initialize the intrinsic mode function components The center frequency of the natural mode function component and the Lagrange multiplication operator

[0146] Initialize the number of iterations n = 0, and perform several iterations with n = n + 1. In each iteration, the intrinsic mode function components are calculated by the following formula Center frequency of the intrinsic mode function component and the Lagrange multiplication operator To update:

[0147]

[0148]

[0149]

[0150] Where τ is the noise tolerance; in each update, the decomposition is stopped when the number of IMF components obtained by decomposing the noise signal reaches K;

[0151] According to the preset precision value ε, the iterative calculation stops when the following conditions are met:

[0152]

[0153] Output several intrinsic mode function components obtained by decomposition. Specifically, the precision value ε can be set according to actual needs, and the present invention does not limit this.

[0154] Specifically, in step S103, the noise signal in a healthy state and the noise signal obtained in a known fault state can be decomposed into a finite number of IMF components such as IMF1 and IMF2 through VMD. Specifically, for different fault types, different IMF component numbers K and penalty factors α can be set to maximize the difference between the fault signal and the normal signal with 1 to 2 or more IMF components, thereby distinguishing different fault signals from normal signals.

[0155] In step S104, the characteristics of several IMF components decomposed by VMD can be used to determine whether a fault exists and the corresponding fault type. Specifically, fault diagnosis can be performed by a neural network model method, that is, a training set is formed by obtaining a healthy signal in a healthy state and a fault signal obtained under several known faults, and input into a preset neural network for training, so as to obtain a neural network that can be used to determine wind turbine blade fault monitoring, and use the neural network to analyze several IMF components obtained by VMD decomposition of the real-time collected noise signal, so as to predict whether a fault occurs and the type of fault.

[0156] Specifically, the neural network used in the present invention can be a long short-term memory network. The long short-term memory network (Long Short-Term Memory, LSTM) is a special recurrent neural network (Recurrent Neural Network, RNN) structure used to process sequence data, such as speech recognition, natural language processing, video analysis and other tasks. The LSTM network introduces a memory cell (memory cell) for storing and updating information in the sequence, and introduces three gates (gate) to control the flow of information in the memory cell: input gate, forget gate and output gate. Among them, the input gate controls the inflow of new inputs, the forget gate controls the forgetting of historical information, and the output gate controls the output of information in the memory cell. The function of the input gate is to selectively record new information into the cell state, the forget gate selectively forgets the information in the cell, and the output gate brings the stored information to the next neuron. The switch state of the three gates is controlled by the sigmoid function, so that the flow of information can be adaptively controlled. By introducing memory cells and gate structures, the LSTM network can better process long sequence data and alleviate the gradient disappearance and gradient explosion problems to a certain extent.

[0157] Specifically, the process of fault diagnosis through several intrinsic mode function components may include:

[0158] Inputting a plurality of intrinsic mode function components into a specified long short-term memory network to obtain a fault prediction result; wherein the long short-term memory network is previously trained by a data set consisting of a plurality of fault signals and healthy signals;

[0159] If the fault prediction result indicates that a fault has occurred, a fault warning is initiated to prompt fault handling.

[0160] Specifically, Figure 2 As shown, Figure 2It is a structural diagram of a long short-term memory network according to an embodiment of the present invention, and the long short-term memory network may include an input layer, a hidden layer and an output layer; wherein the hidden layer may include a memory unit and a control unit, and the control unit may include an input gate, a forget gate and an output gate.

[0161] Among them, the input gate is used to control the inflow of new inputs and update the information in the memory unit; the input gate allows the LSTM network to control the inflow of new inputs and selectively update the information in the memory unit, so as to better capture the long-term dependencies in the sequence data. The forget gate is used to control the forgetting of historical information and update the information in the memory unit; the forget gate allows the LSTM network to selectively forget historical information and update the information in the memory unit; this mechanism can better capture the long-term dependencies in the sequence data and alleviate the gradient disappearance and gradient explosion problems. The output gate is used to control the output of information in the memory unit; the output gate allows the LSTM network to control the output of information in the memory unit, so as to better capture the long-term dependencies in the sequence data and improve the prediction accuracy of the model.

[0162] The LSTM network performs the following calculations in the forward propagation process:

[0163] Calculate the updated forget gate value, the formula is:

[0164] f (t) =σ(W f h (t-1) +W i x (t) +b f )

[0165] Among them, σ is the sigmoid activation function, W f is the forget gate weight, b f is the forget gate threshold;

[0166] Calculate the value of the updated input gate, the formula is:

[0167] i (t) =σ(W i h (t-1) +W i x (t) +b i )

[0168] c ’(t) =tanh(W c h (t-1) +W c x (t) +b c )

[0169] Among them, W i , Wc 、b i and b c They are the weights and thresholds corresponding to the sigmoid activation function and tanh activation function of the input gate respectively;

[0170] Calculate the updated memory cell value, the formula is:

[0171]

[0172] in, is the dot product;

[0173] Calculate the updated output gate value, the formula is:

[0174] o (t) =σ(W o h (t-1) +W o x (t) +b o )

[0175]

[0176] Among them, W o is the output gate weight, b o is the output gate threshold, h (t) is the output vector of the hidden layer;

[0177] Update the prediction output at the current moment. The formula is:

[0178]

[0179] Among them, V and c are the weight and threshold of the connection from the hidden layer to the output layer, respectively.

[0180] After each forward propagation process, the LSTM network performs a reverse calculation based on the error between the predicted value and the actual value, thereby updating each weight and threshold until the maximum number of iterations is met.

[0181] Before fault monitoring, the pre-collected fault signal and health signal data can be imported into the LSTM network model for iterative training, and then the trained LSTM network can be used for fault diagnosis. In step S104, it is only necessary to input several IMF components obtained by decomposing the noise signal by VMD into the preset LSTM network model to obtain the fault condition of the blade corresponding to the noise signal, including whether a fault occurs and the type of fault that occurs.

[0182] Specifically, in other embodiments, other neural network models besides the LSTM network may be used for fault diagnosis, and other algorithms besides the neural network may be used to analyze the IMF components to determine the fault condition, as long as it can be used for noise identification and can be classified according to the IMF components decomposed by VMD, the present invention is not limited to this.

[0183] Specifically, after performing fault diagnosis on several IMF components to obtain fault prediction results, corresponding fault processing can be performed according to the fault prediction results. For example, if the fault prediction result indicates that a fault has occurred, a fault warning can be initiated to prompt the corresponding fault processing to repair the blade fault. Specifically, the method of initiating a fault warning can include any one or more of whistle warning, sending a warning SMS, outputting a fault report or other warning methods, and the present invention is not limited to this.

[0184] The wind turbine blade fault monitoring method disclosed in the present invention installs a noise sensor inside the blade. When the blade cracks, the damage of the blade can be monitored by the change of the noise inside the blade. For example, when a blade cracks, the natural frequency of the blade changes, and the noise it emits will also change; and when the blade cracks more seriously and is accompanied by air leakage, a whistle will appear inside the blade; therefore, by monitoring the difference in noise generated when there is no fault and after the fault, and comparing the noise difference of the three blades of the wind turbine, the fault conditions such as blade cracking can be identified through methods such as noise recognition models. Compared with the existing optical monitoring scheme, the wind turbine blade fault monitoring method proposed in the present invention can identify various fault conditions according to noise changes, has a comprehensive monitoring range, generates less data, is easy to process, and has lower technical difficulty and application cost; compared with the UAV monitoring scheme, the wind turbine blade fault monitoring method proposed in the present invention can realize all-weather online monitoring because the noise sensor is installed inside the blade and is not affected by the external environment; compared with the load monitoring scheme, the wind turbine blade fault monitoring method proposed in the present invention only installs a certain number of noise sensors, so the number of sensors is small, the construction period is short, and the cost is low. At the same time, the amount of data generated in the later stage is also small, the processing is convenient, and the algorithm development difficulty is also lower.

[0185] Corresponding to the above-mentioned embodiment of the method for monitoring blade faults of a wind turbine generator set, the present invention further provides a system for monitoring blade faults of a wind turbine generator set.

[0186] like Figure 3 As shown, Figure 3 The present invention is a schematic diagram of a wind turbine blade fault monitoring system according to an embodiment of the present invention, wherein the system includes a plurality of noise sensors installed inside the wind turbine blade, a signal modulator, a signal receiving station and a monitoring center;

[0187] Among them, the noise sensor is used to collect the noise signal generated by the blades of the wind turbine;

[0188] The signal modem is used to receive the noise signal collected by the noise sensor and transmit the noise signal to the signal receiving station through a wireless connection;

[0189] The signal receiving station is used to transmit the noise signal to the monitoring center through the wind turbine ring network;

[0190] The monitoring center is used to store the noise signal and to perform fault monitoring on the wind turbine blades by using the wind turbine blade fault detection method as described in any embodiment of the present invention.

[0191] Specifically, the process of the monitoring center performing fault monitoring on the wind turbine blades may include:

[0192] The noise signal is collected by the noise sensor installed inside the wind turbine blade;

[0193] Perform data preprocessing on the noise signal;

[0194] Decompose the preprocessed noise signal into a limited number of intrinsic mode function components through variational mode decomposition;

[0195] Fault diagnosis is performed through several intrinsic mode function components.

[0196] Specifically, the noise sensor may be an optical fiber MEMS noise sensor.

[0197] Specifically, the noise sensor can be installed in the following manner: inside the wind turbine blade, starting from the blade root, one noise sensor is installed at a specified distance; and the number and relative positions of the noise sensors installed on each blade of the wind turbine are kept consistent.

[0198] Specifically, performing data preprocessing on the noise signal may include performing data framing on the noise signal; performing data framing on the noise signal may include:

[0199] Taking the wind rotor as the center, the motion trajectory of the wind turbine blades on the wind rotor rotation plane is divided into several angle ranges, and the noise signal collected in each angle range is marked with a corresponding angle label; and the angle label marked on the noise signal on each blade of the wind turbine is different;

[0200] The noise signal is framed according to the angle labels.

[0201] Specifically, performing data preprocessing on the noise signal may include performing data filtering on the noise signal, for example, performing data filtering on the noise signal using a Kalman filter; performing data filtering on the noise signal using a Kalman filter may include:

[0202] The state equation of the wind turbine generator set collecting the noise signal is:

[0203] x t =Ax t-1 +Bu t-1 +ω t-1

[0204] p(ω)~N(0,Q)

[0205] Among them, x t-1 is the state vector at time t-1, u t-1 is the input vector at time t-1, ω t-1 is the process noise, A is the state transfer coefficient matrix of the noise signal, B is the optional control input matrix, Q is the covariance matrix of the process noise, and the process noise satisfies the Gaussian distribution;

[0206] The measurement equation of the noise signal is:

[0207] z t =Hx t +v t

[0208] p(v)~N(0,R)

[0209] Among them, z t is the measurement vector at time t, v t is the corresponding measurement error, H is the measurement coefficient matrix, R is the covariance matrix of the measurement error, and the measurement error satisfies the Gaussian distribution;

[0210] Based on Kalman filtering, the time update equation of the noise signal is:

[0211]

[0212]

[0213] in, is the predicted value of the state vector, and P is the covariance matrix of the error;

[0214] Based on Kalman filtering, the measurement update equation of the noise signal is:

[0215]

[0216]

[0217]

[0218] Where K is the Kalman filter gain.

[0219] Specifically, the process of fault diagnosis through several intrinsic mode function components may include:

[0220] Inputting a plurality of intrinsic mode function components into a specified long short-term memory network to obtain a fault prediction result; wherein the long short-term memory network is previously trained by a data set consisting of a plurality of fault signals and healthy signals;

[0221] If the fault prediction result indicates that a fault has occurred, a fault warning is initiated to prompt fault handling.

[0222] Specifically, the long short-term memory network may include an input layer, a hidden layer and an output layer; wherein the hidden layer may include a memory unit and a control unit, and the control unit may include an input gate, a forget gate and an output gate;

[0223] The LSTM network performs the following calculations in the forward propagation process:

[0224] Calculate the updated forget gate value, the formula is:

[0225] f (t) =σ(W f h (t-1) +W i x (t) +b f )

[0226] Among them, σ is the sigmoid activation function, W f is the forget gate weight, b f is the forget gate threshold;

[0227] Calculate the value of the updated input gate, the formula is:

[0228] i (t) =σ(W i h (t-1) +W i x (t) +b i )

[0229] c ’(t) =tanh(W c h (t-1) +W c x (t) +b c )

[0230] Among them, W i , W c 、b iand b c They are the weights and thresholds corresponding to the sigmoid activation function and tanh activation function of the input gate respectively;

[0231] Calculate the updated memory cell value, the formula is:

[0232]

[0233] in, is the dot product;

[0234] Calculate the updated output gate value, the formula is:

[0235] o (t) =σ(W o h (t-1) +W o x (t) +b o )

[0236]

[0237] Among them, W o is the output gate weight, b o is the output gate threshold, h (t) is the output vector of the hidden layer;

[0238] Update the prediction output at the current moment. The formula is:

[0239]

[0240] Among them, V and c are the weight and threshold of the connection from the hidden layer to the output layer, respectively.

[0241] Specifically, in the process of decomposing the preprocessed noise signal into a finite number of intrinsic mode function components through variational mode decomposition, the variational mode decomposition formula can be:

[0242]

[0243]

[0244] Where, f is the input noise signal, u k are the K intrinsic mode function components obtained after decomposition, K is the number of intrinsic mode function components, ω k is the center frequency of the natural mode function component, δ(t) is the impulse function;

[0245] The Lagrange multiplication operator is introduced into the variational mode decomposition formula for solution, and its expression is:

[0246]

[0247] Among them, λ is the Lagrange multiplication operator, α is the penalty factor, f(t) is the noise signal to be decomposed, and u k (t) is the kth intrinsic mode function component;

[0248] After introducing the Lagrange multiplication operator, the solution to the variational mode decomposition formula is:

[0249] Initialize the intrinsic mode function components The center frequency of the natural mode function component and the Lagrange multiplication operator

[0250] Initialize the number of iterations n = 0, and perform several iterations with n = n + 1. In each iteration, the intrinsic mode function components are calculated by the following formula Center frequency of the intrinsic mode function component and the Lagrange multiplication operator To update:

[0251]

[0252]

[0253]

[0254] Where, τ is the noise tolerance;

[0255] According to the preset precision value ε, the iterative calculation stops when the following conditions are met:

[0256]

[0257] Output several intrinsic mode function components obtained by decomposition.

[0258] Specifically, performing data preprocessing on the noise signal may include performing data screening on the noise signal; performing data screening on the noise signal may include at least one of the following:

[0259] Determine the operating status of the wind turbine according to the wind turbine grid-connected signal, and filter the noise signals collected by the wind turbine in the non-grid-connected state;

[0260] It is determined whether the wind turbine is in a yaw state according to the wind turbine yaw signal, and the noise signal collected by the wind turbine in the yaw state is filtered.

[0261] Specifically, the implementation process of the functions and effects of each device in the above system is specifically described in the implementation process of the corresponding steps in the above method, which will not be repeated here.

[0262] Next, the present invention further illustrates the method and system of the present invention through an operation example of a wind turbine blade fault monitoring system. Figure 4 As shown, Figure 4 This is an operation flow chart of a wind turbine blade fault detection system according to an embodiment of the present invention. In the system, first, a noise signal is collected by a noise sensor installed inside the blade, and the noise signal is transmitted to the monitoring center for analysis through a signal modem and a signal receiving station in turn; then, the noise signal is removed from the interference noise characteristics in the noise signal by Kalman filtering; then, the processed noise signal is decomposed into several IMF components by VMD; then, the several IMF components obtained by decomposition are input into the LSTM network for training, or input into the trained LSTM network for identification; then, the model is deployed through the trained LSTM network, or a fault alarm is performed based on the fault prediction result obtained by the LSTM network identification; and in the event of a fault, the relevant personnel are notified to perform the corresponding fault processing; and if no fault occurs, the next frame of noise signal data continues to be VMD decomposed and fault diagnosed.

[0263] The above embodiments of the present invention are merely examples for clearly explaining the present invention, and are not intended to limit the implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for monitoring blade faults of a wind turbine generator set, characterized in that: The method comprises: Collecting noise signals through a noise sensor installed inside the blades of the wind turbine generator set; Performing data preprocessing on the noise signal; Decomposing the preprocessed noise signal into a finite number of intrinsic mode function components by variational mode decomposition; Performing fault diagnosis through a number of said inherent mode function components; The noise sensor is installed as follows: Inside the blade of the wind turbine generator set, starting from the blade root, one of the noise sensors is installed at a specified distance; and the number and relative position of the noise sensors installed on each blade of the wind turbine generator set remain consistent; A signal modulator is set at the bottom of each blade of the wind turbine, and a signal receiving station is set under each wind turbine. The signal modulator on each blade of the wind turbine is connected to the signal receiving station under the wind turbine through wireless communication, and each signal receiving station is connected to the monitoring center through the wind turbine ring network; In the process of decomposing the preprocessed noise signal into a finite number of intrinsic mode function components through variational mode decomposition, the variational mode decomposition formula is: Where, f is the input noise signal, u k are the K intrinsic mode function components obtained after decomposition, K is the number of the intrinsic mode function components, ω k is the center frequency of the natural mode function component, δ(t) is the impulse function; The Lagrange multiplication operator is introduced into the variational mode decomposition formula for solution, and its expression is: Among them, λ is the Lagrange multiplication operator, α is the penalty factor, f(t) is the noise signal to be decomposed, and u k (t) is the kth intrinsic mode function component; After introducing the Lagrange multiplication operator, the solution method of the variational mode decomposition formula is: Initialize the intrinsic mode function components The center frequency of the natural mode function component and the Lagrange multiplication operator Initialize the number of iterations n = 0, perform several iterations with n = n + 1, and in each iteration, calculate the intrinsic mode function component by the following formula The center frequency of the natural mode function component and the Lagrange multiplication operator To update: Where, τ is the noise tolerance; According to the preset precision value ε, the iterative calculation stops when the following conditions are met: Output several intrinsic mode function components obtained by decomposition.

2. The method according to claim 1, characterized in that The noise sensor is an optical fiber MEMS noise sensor.

3. The method according to claim 1, characterized in that The data preprocessing of the noise signal includes data framing of the noise signal; Performing data framing on the noise signal includes: Taking the wind rotor as the center, the motion trajectory of the wind turbine blade on the wind rotor rotation plane is divided into a number of angle ranges, and the noise signal collected in each angle range is marked with a corresponding angle label; and the angle label marked on each blade of the wind turbine is different; The noise signal is framed according to the angle label.

4. The method according to claim 1, characterized in that: The data preprocessing of the noise signal includes performing data filtering on the noise signal using a Kalman filter; performing data filtering on the noise signal using a Kalman filter includes: The state equation of the wind turbine generator set collecting the noise signal is: x t =Ax t-1 +Bu t-1 +ω t-1 p(ω)~N(0,Q) Among them, x t-1 is the state vector at time t-1, u t-1 is the input vector at time t-1, ω t-1 is process noise, A is the state transfer coefficient matrix of the noise signal, B is an optional control input matrix, Q is the covariance matrix of the process noise, and the process noise satisfies Gaussian distribution; The measurement equation of the noise signal is: z t =Hx t +v t p(v)~N(0,R) Among them, z t is the measurement vector at time t, v t is the corresponding measurement error, H is the measurement coefficient matrix, R is the covariance matrix of the measurement error, and the measurement error satisfies Gaussian distribution; Based on Kalman filtering, the time update equation of the noise signal is: in, is the predicted value of the state vector, and P is the covariance matrix of the error; Based on Kalman filtering, the measurement update equation of the noise signal is: Where K is the Kalman filter gain.

5. The method according to claim 1, characterized in that The fault diagnosis is performed by using a plurality of the intrinsic mode function components, including: Inputting a plurality of said intrinsic mode function components into a specified long short-term memory network to obtain a fault prediction result; wherein said long short-term memory network is obtained in advance by training a data set consisting of a plurality of fault signals and healthy signals; If the fault prediction result indicates that a fault has occurred, a fault warning is initiated to prompt fault handling.

6. The method according to claim 5, characterized in that The long short-term memory network includes an input layer, a hidden layer and an output layer; wherein the hidden layer includes a memory unit and a control unit, and the control unit includes an input gate, a forget gate and an output gate; The long short-term memory network performs the following calculations in sequence during the forward propagation process: The value of the forget gate is calculated and updated, and the formula is: f (t) =σ(W f h (t-1) +W i x (t) +b f ) Among them, σ is the sigmoid activation function, W f is the forget gate weight, b f is the forget gate threshold; The value of the input gate is calculated and updated as follows: i (t) =σ(W i h (t-1) +W i x (t) +b i ) c' (t) =tanh(W c h (t-1) +W c x (t) +b c ) Among them, W i , W c 、b i and b c are the weights and thresholds corresponding to the sigmoid activation function and tanh activation function of the input gate respectively; The value of the memory cell is calculated and updated, and the formula is: C (t) =C (t-1) ⊙f (t) +i (t) ⊙c' (t) Among them, ⊙ is the dot product; The value of the output gate is calculated and updated as follows: o (t) =σ(W o h (t-1) +W o x (t) +b o ) h (t) =o (t) ⊙tanh(C (t) ) Among them, W o is the output gate weight, b o is the output gate threshold, h (t) is the output vector of the hidden layer; Update the prediction output at the current moment. The formula is: Wherein, V and c are the weight and threshold of the connection from the hidden layer to the output layer respectively.

7. The method according to claim 1, characterized in that The data preprocessing of the noise signal includes data screening of the noise signal; the data screening of the noise signal includes at least one of the following: Determine the operating state of the wind turbine according to the wind turbine grid-connected signal, and filter the noise signal collected by the wind turbine in the non-grid-connected state; It is determined whether the wind turbine set is in a yaw state according to the wind turbine yaw signal, and noise signals collected by the wind turbine set in the yaw state are filtered.

8. A wind turbine blade fault monitoring system, characterized in that: The system includes a plurality of noise sensors installed inside the blades of the wind turbine generator set, a signal modem, a signal receiving station and a monitoring center; The noise sensor is used to collect the noise signal generated by the blades of the wind turbine; The signal modem is used to receive the noise signal collected by the noise sensor, and transmit the noise signal to the signal receiving station via a wireless connection; The signal receiving station is used to transmit the noise signal to the monitoring center through the wind turbine ring network; The monitoring center is used to store the noise signal and perform fault monitoring on the wind turbine blade by using the wind turbine blade fault detection method according to any one of claims 1 to 7.

Images