Multi-source information processing wind driven generator rotor demagnetization fault diagnosis method

Abstract

The invention relates to a method for diagnosing a rotor demagnetization fault of a wind driven generator based on multi-source information processing. The method comprises the following steps: firstly, establishing a digital twinborn model of a health state and a demagnetization fault state of the wind driven generator; the vibration acceleration of the surface of the stator tooth groove of the wind driven generator is obtained through electromagnetic-structural field coupling calculation; the obtained signals are constructed into a database, the signals in different states in the database are decomposed into multiple modal signals, the fitness function of the signals is calculated, and an optimal decomposition mode is found out; converting the multiple modal signals into feature vectors through a time-frequency domain technology; inputting feature vectors of signals measured in different wind driven generator states into the optimized probabilistic neural network for training, and finally obtaining a neural network capable of accurately judging the rotor state of the wind driven generator; vibration data measured by a detection probe on the wind driven generator in field investigation are input into the neural network, and the current operation state of the wind driven generator rotor can be distinguished.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis of wind turbines, and particularly relates to a method for diagnosing the demagnetization fault of the rotor of a wind turbine by processing multi-source information. Background Art

[0002] Permanent magnet wind turbines use permanent magnets instead of electromagnets in traditional generators. Permanent magnets can generate a stable magnetic field, avoiding the energy loss caused by the flow of current in electromagnets. Moreover, since there are no vulnerable parts such as brushes and slip rings, permanent magnet wind turbines have high reliability and usually long operating life, which makes them more suitable for some environments with high reliability requirements. Therefore, permanent magnet synchronous motors are now widely used in fields such as elevators, electric vehicles, and wind power generation.

[0003] However, due to complex environments, material life, manufacturing defects, etc., the performance of permanent magnet synchronous motors will gradually deteriorate as the component performance deteriorates. After that, various fault types such as eccentricity and bearing faults will inevitably occur in permanent magnet synchronous motors. In addition, since permanent magnets replace the excitation windings, permanent magnet synchronous motors will also have unique demagnetization faults. These faults will affect the normal use of the equipment and, in severe cases, will affect the property and personal safety of related industries. Therefore, it is of great significance to detect and diagnose the faults of permanent magnet synchronous motors to ensure the stability of the system and thus reduce accidents.

[0004] In previous research on fault diagnosis methods, most domestic studies only focused on single faults and collected single signals, and rarely used the method of fusing multiple fault signals. Therefore, there is an urgent need to provide a method for diagnosing the demagnetization fault of the rotor of a wind turbine by processing multi-source information to improve the accuracy of fault diagnosis. Summary of the Invention

[0005] Object of the Invention

[0006] To solve the problems in the prior art, the present invention provides a method for diagnosing the demagnetization fault of the rotor of a wind turbine by processing multi-source information. The method combines a digital twin model with an algorithm. By using the variational mode decomposition algorithm to extract feature vectors and classifying the feature vector data through a probabilistic neural network, the fault type of the input signal can be diagnosed. This method can be used for some devices that are difficult to effectively diagnose faults due to scarce fault data. This method uses finite element analysis in digital twin technology to supplement the data that cannot be obtained through traditional experiments, thereby improving the accuracy of fault prediction. Through this method, even in the case of limited data, accurate diagnosis and evaluation of the health status of the equipment can be achieved.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for diagnosing the demagnetization fault of a wind turbine rotor with multi-source information processing, comprising the following steps:

[0009] Step 1: Establish a digital twin model of a permanent magnet motor and simulate the digital twin model of the permanent magnet motor;

[0010] Step 2: Conduct electromagnetic-structural field coupling simulation on the digital twin model of the wind turbine, and obtain the electromagnetic force of the motor and the vibration acceleration on the stator and rotor by calculating the air-gap magnetomotive force and the air-gap permeance coefficient;

[0011] Step 3: Analyze the vibration characteristics obtained at different measuring points under different demagnetization degrees of the wind turbine based on the digital twin model, and establish a characteristic database;

[0012] Step 4: Perform data preprocessing on the data prepared in the database;

[0013] Step 5: Classify the extracted feature vectors through a neural network to obtain a neural network, and then complete the diagnosis and obtain the fault type by inputting the signal into the neural network.

[0014] As a further description of the above solution, step 1 includes the following steps:

[0015] Step 1.1: Based on the structural design parameters of the wind turbine, establish a digital twin model of a permanent magnet wind turbine;

[0016] Step 1.2: Based on the parameters of the wind turbine, establish a two-dimensional model, and then establish a fault model. Modulate the original values of the coercive force of the permanent magnet by 10%, 20%, 30%, 40%, and 50% respectively, corresponding to 10%, 20%, 30%, 40%, and 50% permanent magnet demagnetization faults.

[0017] As a further description of the above solution, step 2 includes the following steps:

[0018] Step 2.1: Based on the electromagnetic parameters in the design parameters of the wind turbine, including: rated voltage, winding winding method, number of permanent magnets, stator and rotor materials, etc., conduct electromagnetic field simulation on the digital twin model of the wind turbine and obtain the electromagnetic force by calculating the air-gap magnetomotive force and the air-gap permeance coefficient;

[0019] b(θ,t) = f(θ,t)λ(θ,t)

[0020] In the formula, f(θ,t) is the air-gap magnetomotive force; λ(θ,t) is the air-gap permeance coefficient;

[0021] Step 2.2: Based on the material properties of the results of the digital twin model of the wind turbine, including the material density, elastic modulus, and Poisson's ratio of each structure as the basic material properties, and using the electromagnetic force applied to the stator obtained from the electromagnetic field simulation as the input for the structural field vibration simulation.

[0022] As a further description of the above solution, the said Step 3 includes the following steps:

[0023] Step 3.1: Obtain the fault characteristics based on the demagnetization state of the digital twin model of the wind turbine, including: the amplitudes of the time-domain vibration signals at multiple measurement points on the stator and rotor surfaces, the conversion of the multi-measurement-point vibration signals to the frequency domain, the vibration acceleration amplitudes and their changes at the fundamental frequency, double fundamental frequency, and their multiple frequencies.

[0024] Step 3.2: Modify the digital twin model of the wind turbine through historical fault data to make its calculation more accurate;

[0025] Step 3.3: Conduct electromagnetic-structural coupling simulations on the digital twin models of the permanent magnet wind turbines in the normal state and different demagnetization conditions, and extract the vibration accelerations at the measurement point positions in the normal operating state and different fault states;

[0026] Step 3.4: Repeat Step 3.3 for a set number of times to obtain the vibration accelerations at different measurement point positions as the analysis data samples. At the same time, conduct vibration tests to obtain the vibration accelerations, and correct the analysis data samples obtained from the simulation calculations. Take the vibration acceleration amplitudes including the fundamental frequency and multiple frequencies in the normal state and different fault states as the frequency domain characteristics of the vibration signals and construct a feature database.

[0027] As a further description of the above solution, the said Step 4 includes the following steps:

[0028] Step 4.1: Decompose the signal to separate the fault characteristics of the system vibration. First, construct the following constrained variational model:

[0029]

[0030] In the formula, {u k} = {u 1 , …, u K} represents the k IMF modal components obtained by VMD decomposition, {w k} = {w 1 , …, w K} is the set of the central frequencies of each modal component, K 模态 is the number of modes obtained by decomposition, represents the partial derivative operation, δ(t) represents the Dirac function, * represents convolution, and f(t) is the original signal;

[0031] After introducing the penalty factor α and the Lagrange multiplier operator λ(t), the alternating direction method of multipliers is used to alternately iterate and update { uk}, { wk} and λ to find the optimal solution of the constrained variational model. The results are as follows:

[0032]

[0033] In the formula are respectively 's Fourier transform, and n is the number of iterations;

[0034] Step 4.2: After decomposing the healthy state, rotor demagnetization of 10%, 20%, 30%, 40%, and 50% fault states into Modes by VMD, extract the energy entropy of each Mode of the VMD decomposition as the feature vector;

[0035] As a further description of the above solution, the step 5 includes the following steps:

[0036] Step 5.1: Integrate the feature vectors of the healthy state, rotor demagnetization of 10%, 20%, 30%, 40%, and 50% fault states, and add corresponding labels to the healthy state, rotor demagnetization of 10%, 20%, 30%, 40%, and 50% fault states;

[0037] Step 5.2: Use a probabilistic neural network to classify the integrated feature vectors;

[0038] Step 5.3: Optimize the probabilistic neural network, that is, optimize the minimization of the loss function, so that the distribution of the probabilistic neural network is as consistent as possible with the true class distribution. By iteratively updating the parameters, an optimal solution is finally found to improve the classification accuracy of the probabilistic neural network;

[0039] Step 5.4: Train the probabilistic neural network;

[0040] Step 5.5: After inputting the wind turbine detection data into VMD for decomposition and inputting the feature vectors extracted from VMD into the probabilistic neural network, the probabilistic neural network will calculate the similarity between the feature vectors and the probability density functions of the data in various states in the training set, and then classify the wind turbine detection data to predict the fault type of each input detection data.

[0041] Advantages and effects of the present invention:

[0042] 1. Step 3 of the present invention adopts a fault detection method combining a digital twin model and an algorithm, which can be used for some devices that are difficult to perform effective fault diagnosis due to scarce fault data. This method uses finite element analysis in digital twin technology to supplement the data that cannot be obtained by traditional experiments, thereby improving the accuracy of fault prediction. Even in the case where it is difficult to diagnose faults in wind turbine equipment due to insufficient fault data, the present invention can construct an accurate mathematical model, effectively reducing the challenges in the field of fault diagnosis. Through this method, even in the case of limited data, it is possible to accurately diagnose and evaluate the health status of the equipment.

[0043] 2. Step 4.2 of the present invention proposes a Subtraction Averaging Optimizer Algorithm (SABO) to optimize the selection of parameters such as the penalty factor and the number of decomposition modes of Variational Mode Decomposition (VMD), ensuring that the vibration signal components are evenly distributed, with obvious features, clear and reasonable frequency band division, and no under-decomposition or mode mixing, etc., so as to obtain a better feature extraction effect.

[0044] 3. Step 5.3 of the present invention proposes an optimization algorithm to optimize the wind turbine demagnetization fault and health status assessment model of the Probabilistic Neural Network (PNN). The PNN is optimized by the gradient descent method to improve the accuracy of fault diagnosis. This model can be used in the case of scarce data and can effectively solve the problem of difficult fault diagnosis in wind turbines due to insufficient fault data. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of an invention for a method of diagnosing the demagnetization fault state of a wind turbine;

[0046] Figure 2 It is a partial simulation experiment model of the wind turbine of the present invention;

[0047] Figure 3 It is a rotor twin model of the wind turbine of the present invention;

[0048] Figure 4 It is a vibration equivalent spring-mass model diagram of a part of the wind turbine of the present invention;

[0049] Figure 5 It is a vibration acceleration-time image of the motor in four states measured at a certain measuring point on the casing in the simulation experiment of the present invention, where: (a) is the healthy state, and (b) is the 10% demagnetization state;

[0050] Figure 6 It is a signal image obtained by decomposition in Step 4 of the present invention, where: (a) is the healthy state, and (b) is the 10% demagnetization state;

[0051] Figure 7The effect diagram obtained by training the neural network in step 5.4 of the present invention;

[0052] Figure 8 The effect diagram of classifying the faults of the wind turbine by the wind turbine in step 5.5 of the present invention.

[0053] In the accompanying drawings, the list of components represented by each reference numeral is as follows:

[0054] 1 - Motor housing, 2 - Stator, 3 - Rotor, 4 - Permanent magnet, 5 - Rotor support, 6 - Rotating shaft. Detailed implementation manners

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0056] A method for diagnosing the demagnetization fault of the rotor of a wind turbine with multi - source information processing includes the following steps:

[0057] Step 1: Establish a digital twin model of the wind turbine under normal and faulty conditions;

[0058] Based on the structural design parameters of the wind turbine, including the inner and outer diameters and lengths of the rotating shaft 6, the lamination length, width, slot size of the stator 2, the thickness and radius of the lamination, the radius of the rotor 3, the rotor thickness, the length, width, height and shape of the permanent magnet 4, the inner and outer diameters of the rotor support 5, the inner and outer diameters, radius and thickness of the motor housing 1, and the sectional area shape size and length, etc., a digital twin model of the permanent - magnet wind turbine is established. Figure 2 What is shown is an exploded view of the overall wind turbine;;

[0059] Step 1.2: Based on the wind turbine parameters, establish a two - dimensional model, and then establish a fault model. The coercivity of the permanent magnet is modulated by 10%, 20%, 30%, 40% and 50% of the original value respectively, corresponding to 10%, 20%, 30%, 40% and 50% permanent magnet demagnetization faults. In this article, the healthy state and 10% demagnetization are taken as examples mainly;

[0060] Step 2: Conduct electromagnetic - structural field coupling simulation on the digital twin model of the wind turbine, and obtain the electromagnetic force of the motor and the vibration acceleration on the stator and rotor by calculating the air - gap magnetomotive force and the air - gap permeance coefficient.

[0061] Step 2.1: Based on the electromagnetic parameters in the design parameters of the wind turbine, including rated voltage, winding winding method, number of permanent magnets, stator and rotor materials, etc., perform electromagnetic field simulation on the digital twin model of the wind turbine to obtain electromagnetic force;

[0062] Based on the established high-precision two-dimensional model, first perform two-dimensional field electromagnetic calculation with a voltage source as the excitation, including radial air-gap flux density and tangential flux density, and then calculate the radial electromagnetic force and tangential electromagnetic force through the radial flux density and tangential flux density. Since the magnetic permeability of ferromagnetic materials is much greater than that of air, when the magnetic induction line enters the iron core of the stator, it is basically perpendicular to the surface of the iron core. Therefore, the tangential component of the air-gap magnetic density and the tangential electromagnetic force are not considered at present. The formula for the radial air-gap magnetic density is as follows:

[0063] b(θ,t) = f(θ,t)λ(θ,t) (1)

[0064] In the formula, f(θ,t) is the air-gap magnetomotive force; λ(θ,t) is the air-gap permeance coefficient.

[0065] The air-gap magnetomotive force can be expressed as:

[0066]

[0067] In the formula, p is the number of pole pairs of the motor, μ 谐波极对数 is the harmonic pole pair number of the rotor air-gap magnetomotive force; v is the harmonic pole pair number of the stator air-gap magnetomotive force; is the amplitude of the rotor air-gap harmonic magnetomotive force, F v is the amplitude of the stator air-gap harmonic magnetomotive force; are the initial angles of the stator and rotor magnetomotive forces respectively; w 0 is the electrical angular frequency of the rotor rotation.

[0068] The air-gap permeance function can be expressed as:

[0069]

[0070] In the formula, λ 0 is the amplitude of the average air-gap permeance; k z is the harmonic order; is the amplitude of the air-gap k z order tooth harmonic permeance; Z 0 is the number of stator slots of the motor.

[0071] The radial electromagnetic force per unit area on the inner surface of the motor can be expressed as:

[0072]

[0073] In the formula, b(θ,t) is the radial air-gap magnetic density; t is the time variable; θ is the circumferential position angle; μ0 is the magnetic permeability in vacuum.

[0074] Substitute the above equations (1), (2), and (3) into equation (4), and the expression for the radial electromagnetic force is obtained as:

[0075]

[0076] When performing finite element analysis on the situation of different degrees of demagnetization of the permanent magnet in the present invention, the coercivity H of the permanent magnet material parameters is changed c to achieve the preset of the demagnetization fault. The demagnetization fault calculation is as follows:

[0077] The coercivity after demagnetization is:

[0078] H' f =(1 - s)H c (6)

[0079] In the formula, H' f is the coercivity after the material is demagnetized, and s is the demagnetization coefficient.

[0080] Another important parameter characterizing the performance of the permanent magnet material is the remanence coercivity H of the permanent magnet c , and the formula at this time is:

[0081]

[0082] In the formula, H c is the material coercivity; B r is the remanence intensity; μ 材料 is the magnetic permeability of the material.

[0083] The material magnetic permeability μ 材料 is equal to the vacuum magnetic permeability μ 0 multiplied by the relative magnetic permeability μ r , and the formula is as follows:

[0084] μ 材料 =μ 0 μ r (8)

[0085] Then the air-gap magnetic density after demagnetization is:

[0086]

[0087] In the formula, l is the effective magnetic circuit length.

[0088] Therefore, the expression for the radial electromagnetic force after demagnetization is:

[0089]

[0090] According to the remanence B rAccording to the definition, it is incorrect to change the remanence. Therefore, in this simulation, the coercive force H of the permanent magnet material parameters is changed c to preset the demagnetization fault.

[0091] Step 2.2: Based on the results of the material properties of the digital twin model of the wind turbine, including the material density, elastic modulus, and Poisson's ratio of each structure as the basic material properties, and the electromagnetic force applied to the stator obtained from the electromagnetic field simulation as the input, perform a structural field vibration simulation;

[0092] Couple the electromagnetic force results into the harmonic response analysis module, and calculate the vibration acceleration images of the synchronous motor itself during mechanical faults and normal operation through harmonic response analysis.

[0093]

[0094] In the formula, [M], [C], and [K] are the mass matrix, damping matrix, and stiffness matrix of the mechanical system respectively; are the vibration acceleration, vibration velocity, and vibration displacement vectors of each node of the system respectively; {F} is the externally applied sinusoidal harmonic load.

[0095] Assume that {F} and {u} are sinusoidal excitation and sinusoidal response that vary harmonically at frequency Ω, and the formula is as follows:

[0096] {F} = {F max e iψ}e iΩt (12)

[0097] {u} = {u max e iψ}e iΩt (13)

[0098] At this time, F max is the maximum amplitude of the force, u max is the maximum amplitude of the displacement, Ω is the input circular frequency, w is the output circular frequency, and ψ is the phase angle.

[0099] Substitute equations (7) and (8) into equation (6) to obtain the actual calculation equation for harmonic response analysis as:

[0100] (-Ω 2 [M] + IΩ[C] + [K])({u 1} + i{u 2}) = ({F 1} + i{F 2}) (14)

[0101] In the formula, F 1 and F 2 represent the external force vector, u1 and u 2 represents the displacement vector of the system.

[0102] Step 3: Analyze the vibration characteristics of different measuring points of the wind turbine under different degrees of demagnetization faults based on the digital twin model, and establish a vibration characteristic database containing different fault conditions;

[0103] Step 3.1: Obtain the fault characteristics based on the demagnetization state of the digital twin model of the wind turbine, including: the amplitudes of the time-domain vibration signals of multiple measuring points on the stator and rotor surfaces, the conversion of the vibration signals of multiple measuring points to the frequency domain, the vibration acceleration amplitudes and their changes at the fundamental frequency, double fundamental frequency, and their multiple frequencies;

[0104] Step 3.2: Conduct electromagnetic-structural coupling simulations on the digital twin models of the permanent magnet wind turbines in the normal state and different demagnetization conditions, and extract the vibration accelerations at the measuring point positions in the normal operating state and different fault states.

[0105] Step 3.4: Repeat Step 3.3 for a set number of times to obtain the vibration accelerations at different measuring point positions as analysis data samples. At the same time, conduct vibration tests to obtain the vibration accelerations, and correct the analysis data samples obtained from the simulation calculations. Take the vibration acceleration amplitudes at the vibration fundamental frequency and multiple frequencies in the normal state and different fault states as the frequency domain characteristics of the vibration signals and construct a characteristic database;

[0106] Step 3.4.1: In the vibration test of the permanent magnet synchronous generator, 4 PCB piezoelectric acceleration sensors are used as vibration sensors. The information is converted through the vibration acquisition device port and loaded to the PC side through data interaction for storing and analyzing vibration data. Select 4 positions on the surface of the motor housing of the permanent magnet motor experimental prototype to place measuring points, which are respectively at the front end cover and the rear end cover of the permanent magnet motor housing, at the 1 / 2 height on the left side of the motor housing, and at the 1 / 2 height on the right side of the motor housing.

[0107] Demagnetization fault simulation experiment: To simulate the demagnetization fault, the test motor is equipped with a normal rotor and a demagnetized rotor, and the demagnetization process is carried out on the permanent magnets of the demagnetized rotor. The demagnetization process is to place the permanent magnet in a magnetic field with a gradually decreasing magnetic field intensity, so as to gradually reduce the magnetization degree. The specific operation is to place the permanent magnet in a demagnetizer with an adjustable magnetic field intensity, and then gradually reduce the magnetic field intensity until the permanent magnet is completely demagnetized.

[0108] The experimental acquisition platform selects an 8-channel vibration collector, and 4 channels are used to collect vibration data at 4 measurement points in the experiment. During the experiment, the vibration test is carried out by energization. The operating load and electrical coefficients of the permanent magnet motor are monitored through the test bench, and the vibration acceleration of 4 measurement points is obtained through sensors. The vibration signal data is acquired and converted by the vibration acquisition device and then stored on the PC side. Different operating conditions of the body structure are simulated in the experiment. The number of sampling points for each group of vibration signals is set to 1500, the acquisition channels are set to 4, corresponding to 4 measurement point sensors. The sampling time for each group of signals is 500 ms. When the permanent magnet motor operates stably, the time-frequency domain vibration signals are collected, and repeated data acquisitions are carried out based on different states.

[0109] The vibration acceleration data is transmitted to the PC side by the vibration collector, and the time-frequency domain conversion is automatically completed by the Fourier transform through the PC side processor. The vibration signal waveforms and frequency domain distribution rules collected by 4 channels can be observed on the PC side. The overall experimental process includes vibration signal acquisition experiments under no-load operation, 50%, 70%, 100% load operation conditions, and demagnetization fault conditions, and multiple groups of vibration signals are repeatedly collected.

[0110] When the wind turbine is in the demagnetization fault condition, the complexity of the time-frequency domain characteristics of the vibration signal is relatively high, and the high-frequency components in each frequency band have a greater impact, making it difficult to directly compare and analyze the simulation calculation and experimental results. Therefore, the vibration signals under 50%, 70% and 100% load conditions are used to verify the correctness of the digital calculation model of the wind turbine respectively.

[0111] The time-domain vibration signals of the test show relatively obvious periodic characteristics, and the frequency domain does not match. The overall frequency domain distribution characteristics of the experiment and simulation calculation are almost the same under different load conditions. The amplitudes are basically concentrated at 40 and 80 Hz. There are also low amplitudes at the other even multiples of the frequency for the test and simulation signals. There are certain distribution differences in the overall vibration amplitudes, but the numerical differences are not large.

[0112] Step 4: Preprocess the data prepared in the database. The data preprocessing process is as follows:

[0113] Step 4.1: Decompose the signal to separate the fault characteristics of the system vibration.

[0114] Through the digital model simulation calculation, the vibration acceleration data at the surface measurement point position after the rotor demagnetization fault can be extracted. Combining the experimental acquisition of multi-measurement point vibration signals to form a vibration signal dataset, and then the dataset is subjected to VMD fault feature extraction. The process is as follows:

[0115] Load the original data, input the modal parameter k and the penalty factor α, and mark the input data of length N as X(N), expressed as

[0116] Xp (N) = {x 1 , x 2 , …, x n}, p = 1, 2, 3, …, 10; N = 1, 2, 3, …, N

[0117] 4.1.1: Decompose the original data into K IMF modal components u with specific sparsity through VMD k (k = 1, 2, 3, …, K), and the center frequency w of each component can be determined k . First, construct the constrained variational model as follows:

[0118]

[0119] In the formula, represents the k IMF modal components obtained by VMD decomposition, is the set of the center frequencies of each modal component, K 模态 is the number of modal components obtained by decomposition, represents the partial derivative operation, δ(t) represents the Dirac function, * represents convolution, and f(t) represents the target signal.

[0120] 4.1.2: When solving, introduce the quadratic penalty factor α and the Lagrange multiplier operator λ(t) based on Equation (15), and transform Equation (15) into an unconstrained optimization model. The expression is as follows:

[0121]

[0122] 4.1.3: Based on Equation (17) and Equation (18), find the optimal solution of the constrained variational model by alternately iteratively updating and λ(t). The result is as follows:

[0123]

[0124] In the formula are respectively 's Fourier transform, and n is the number of iterations.

[0125] 4.1.4: First, set the number of iterations n = 0, and then initialize and

[0126] 4.1.5: Update and

[0127] 4.1.6: Update

[0128]

[0129] τ is the Lagrangian multiplier update parameter.

[0130] 4.1.7: Determine whether to converge according to the above formula (19). If the convergence condition is met, the decomposition process ends. Otherwise, the iteration number n = n + 1, and return to steps (5) and (6) to continue the decomposition.

[0131] In the formula, ε represents the convergence condition parameter, and the value is set to 10 -7 .

[0132] Step 4.2: Decompose the signal to separate the fault characteristics of the system vibration. First, construct the constrained variational model as follows:

[0133] After decomposing the healthy state, 10%, 20%, 30%, 40%, and 50% rotor demagnetization fault states into Modes by VMD, extract the energy entropy of each Mode of the VMD decomposition as the feature vector. The calculation process is as follows:

[0134] First, perform algorithm initialization. The set of exploration individuals constitutes the population of the algorithm. From a mathematical model perspective, the population algorithm can be represented by a matrix as shown in the following formula:

[0135]

[0136] x id = l d + k id ·(u d - l d )

[0137] The exploration individual randomly initializes the search space position through the above formula. In the formula, X is the SABO population matrix, x i is the i-th search individual, x id is its d-th dimension (decision variable) in the search space, k id is a random number in the interval [0, 1], u d and l d are the upper and lower bounds of the decision variable respectively, n SABO is the number of search times, and m 决策 is the number of decision variables.

[0138] Each search individual is a candidate optimization result of the target problem. Therefore, the evaluation value of the target function can be obtained based on each search individual, as shown by the following vector To represent, based on the position of each population member specifying the value of the decision variable of the problem, evaluate the target function and store it in the vector. Therefore, the number of elements in the vector is equal to the iteration number n.

[0139]

[0140] Among them, is the objective function vector value, F i is the solution value of the objective function based on the i-th search individual.

[0141] The optimization of the evaluation value of the objective function is to find the optimal solution by analyzing the quality of the solution proposed by the best search individual. The position of the search individual in the search space is updated in each iteration. Therefore, the process of identifying and saving the best search individual continues until the last iteration of the algorithm in the setting.

[0142] SABO is based on a new computational concept "- V ", called the v-subtraction between search individual iterations, as shown in the following formula:

[0143]

[0144] In the formula, is a vector randomly generated in the interval [1, 2] and with dimension m 维度 ; F(A) and F(B) are the objective function values of search individuals A and B respectively; sign is the signum function.

[0145] In the SABO algorithm, the position update of any search individual x i in the search space is calculated through the arithmetic mean of the v-subtraction of each search individual x j , as shown in the following formula:

[0146]

[0147] In the formula, is the new deduced position of the i-th search agent, n 搜索次数 is the total number of searches, is a vector with dimension m, and the values of its components are normally distributed in the interval [0, 1].

[0148] After that, if the new deduced position leads to an improvement in the objective function value, it can be accepted as the new corresponding position, according to the following formula:

[0149]

[0150] In the formula, F i and F i newThey are the objective function value of the search agent and the particle position formula respectively. In step 4.2 of the present invention, a Subtraction Averaging Optimizer algorithm (SABO) is proposed to optimize and select the penalty factor and the number of decomposition modes of the Variational Mode Decomposition (VMD), ensuring that the vibration signal components are evenly distributed, with obvious features, clear and reasonable frequency band division, and no under-decomposition or mode aliasing, etc., so as to obtain better feature extraction effects.

[0151] Step 4.3: Take the vibration signal as the input function, and perform parameter optimization iterations respectively. Set the value range of the modal component to be 3 - 9, which is used as the SABO constraint condition. In the sample parameter optimization results, the modal result can be expressed as:

[0152]

[0153] where Mode represents the modal component.

[0154] Confirm the input parameters in step 4

[0155] Set the numerical range of K to be 3 - 9 and the value range of α to be 100 - 5000 as the SABO constraint condition. Calculate the fitness function for the decomposition modes obtained from each combination of individual input parameters. The calculation method of the fitness function is as follows:

[0156] First, represent the given time series as a vector, and take the number of samples in 1s as x n , for complex time series, usually choose 6 as the embedding dimension, and the embedding vectors extracted from the time series are as follows:

[0157]

[0158] Its permutation pattern is expressed as:

[0159]

[0160] In the permutation, from y 0 ~y 7 and represents x 0 ~x 7 in ascending order.

[0161] According to the permutation pattern, an n 排列数量 ×m 嵌入维数 matrix can be obtained:

[0162]

[0163] It can be seen from the matrix that there are a times the same as the first row, that is, there are a times the same as the first permutation pattern. Then the probability of the first permutation pattern is:

[0164]

[0165] There are b times the same as the second line, and the probability of the second arrangement pattern is:

[0166]

[0167] There are z times the same as the nth line, and the probability of the nth arrangement pattern is:

[0168]

[0169] where n 排列数量 represents the number of arrangements

[0170] The permutation entropy H is calculated by the following formula:

[0171]

[0172] where P i概率 represents the probability of each arrangement.

[0173] Substituting equations (23), (24), and (25) into equation (26) gives the permutation entropy H:

[0174]

[0175] After that, the permutation entropy is normalized to be in the range of 0 to 1, and the specific normalization formula is as follows:

[0176]

[0177] PE represents the permutation entropy after normalization, and m 嵌入维数 is the embedding dimension.

[0178] Substituting equation (27) into equation (28) gives the fitness function value:

[0179]

[0180] Inputting the permutation entropy into the VMD algorithm gives the optimal modal component K corresponding to the fitness function and the value of the penalty factor α.

[0181] For VMD, we choose the VMD energy entropy as the extracted feature vector, and the energy entropy H EN is defined as the following formula:

[0182]

[0183] E i represents the energy of the i-th IMF component, and E sum represents the sum of the energies of all components.

[0184] Step 4.4: Calculate the numerical value of the comprehensive feature fitness function:

[0185] For each combination of modal component K and penalty factor α, establish a fuzzy decision matrix, and calculate the allocation weight based on the fuzzy decision matrix as follows:

[0186]

[0187] In the formula, R is the fuzzy decision matrix, which contains n 特征 feature elements. In the matrix, r ij etc. represent the relative importance degree of comparison between the quantitative expression feature elements r i and r j . It is quantified by the 0.1 - 0.9 scaling method to form an n-order fuzzy decision matrix. The scaling method is shown in Table 4.1. ij i Table 4.1 Principle of Scaling Method Assignment

[0188] Table 4.1 Principle of Scaling Method Assignment

[0189]

[0190] Based on the evaluation elements and index importance degree in the fuzzy decision matrix, calculate the feature weight w i of each influencing factor:

[0191]

[0192] Integrate the weight vector W:

[0193] W = (w 1 , w 2 , …, w n ) T

[0194] After calculating the obtained allocation weight, it is also necessary to check the consistency of the fuzzy decision matrix. First, define the consistency index CI:

[0195]

[0196] In the formula, λ i is the eigenvalue of the fuzzy matrix, n 向量数量 represents the number of vectors used to calculate the permutation entropy in the time series, λ max is the maximum eigenvalue of the fuzzy matrix. Then, based on the set order of the fuzzy matrix, obtain the average random consistency index RI. The relationship between RI and the matrix order n is shown in Table 4.2 below:

[0197] Table 4.2 Average Random Consistency Index

[0198]

[0199] The consistency of the fuzzy judgment matrix is detected by combining the CI and RI calculations, and the calculation formula is as follows:

[0200] CR = CI / RI (33)

[0201] If the calculated judgment matrix CR < 0.1, it is considered that the judgment matrix meets the consistency test and the assigned weight values are reasonable.

[0202] Complete the calculation of the comprehensive fitness function:

[0203]

[0204] where w is the value of the comprehensive fitness function at different points.

[0205] Complete the screening of the comprehensive fitness function:

[0206] ρ = min{w 1 , w 2 , …, w n}

[0207] where ρ is the value of the fitness function under the current combination of K and α.

[0208] Step 4.4.1: Select the modal component K and the penalty factor α corresponding to ρ under the combination of the minimum modal component K and the penalty factor α as the input parameters.

[0209] Step 4.5: Extract the energy entropy as a feature:

[0210] Step 4.5.1: The VMD algorithm decomposes the signal x(t) into multiple modal signals

[0211]

[0212] In the formula, K is the number of decomposed modes, is the k-th modal signal.

[0213] Suppose the decomposed modes are as follows:

[0214]

[0215] The energy E of each IMF mode K can be calculated by the following formula:

[0216]

[0217] where T is the total duration of the time series, is the squared value of the modal signal, representing the energy density.

[0218] Assume the total energy is E total , then the probability P of the k-th modek It can be calculated by the ratio of its energy to the total energy:

[0219]

[0220] where E total is the sum of all modal energies:

[0221]

[0222] In the formula, E k represents the energy of the k-th mode, and K represents the number of modes.

[0223] The calculation formula of energy entropy S is as follows:

[0224]

[0225] In the formula, P k is the probability of the k-th mode.

[0226] Step 5: Classify the extracted feature vectors through a neural network to obtain a neural network. After formation, input the signal into the neural network to complete the diagnosis and obtain the fault type;

[0227] Step 5.1: Integrate the feature vectors of the healthy state, 10%, 20%, 30%, 40%, and 50% rotor demagnetization fault states, and add corresponding labels to the healthy state, 10%, 20%, 30%, 40%, and 50% rotor demagnetization fault states;

[0228] The feature form extracted from each signal is as follows:

[0229] Feature = {S 1 , S 2 , S 3 , …, S K}

[0230] Step 5.2: Use a probabilistic neural network to classify the integrated feature vectors;

[0231] 5.2.1: Each pattern unit forms the dot product Z mn of the input pattern vector Feature and the weight W mn = FeatureiW mn Then perform a non-linear operation on its activation level before outputting to the summing unit δ is the scaling factor. The PNN decision boundary is associated with the smoothing factor. After normalizing Feature and W mn , the non-linear operation of Z mn is converted to:

[0232]

[0233] Among them, represents the nth 类 training vector of the mth 训练向量 class.

[0234] 5.2.2: Neuron calculation in the sample layer:

[0235]

[0236] Among them, P 维度 is the dimension of Feature, i.e., the smoothing factor, m 类别 is the class number, n 模式 is the pattern number.

[0237] 5.2.3: Neuron calculation in the summation layer. The clustering pattern Feature is classified into C m类别 (1, 2, 3... N). Calculate the output of the comprehensive calculation neuron:

[0238]

[0239] Among them, C m类别 represents the class, where m 类别 is the class number, N m is the number of training patterns in class C m类别 , is the normalization constant of the Gaussian distribution, where p 特征 is the dimension of the feature, δ 标准差 is the standard deviation..

[0240] 5.2.4: The decision-making plan of Bayes is d(x) = C m类别 , if:

[0241]

[0242] Among them, is the loss of decision-making error for each class C m类别 , is the class-conditional probability density of Feature, P(C m类别 ) is the probability that the vector conforms to the class. Classify the cluster pattern Feature according to the Bayes decision rule:

[0243]

[0244] In the formula, is the evaluation classification of the feature Feature, N m is the total number of classes of the training samples.

[0245] Step 5.3: Optimize the probabilistic neural network, that is, optimize the minimization of the loss function to make the distribution of the probabilistic neural network as consistent as possible with the true class distribution. By iteratively updating the parameters, an optimal solution is finally found, thereby improving the classification accuracy of the probabilistic neural network;

[0246] 5.3.1: Define the loss function:

[0247] Suppose there is a dataset containing N 数据 training samples, and each sample belongs to one of C 样本 classes. For each sample i, its true label is y i , and the output of the network is the probability vector P i . Then the loss function is:

[0248]

[0249] In the formula, represents that sample i 样本 belongs to class j 类别 , the model predicts the probability that sample i 样本 belongs to class j 类别 .

[0250] 5.3.2: Calculate the gradient:

[0251] For the smoothing parameter Ω 梯度 , the gradient of the loss function is:

[0252]

[0253] In the formula, x i样本 is the feature vector of sample i 样本 , represents the feature vector of the k 样本 th sample in the training set.

[0254] 5.3.3: Update the parameters

[0255] Use the gradient descent method to update Ω 梯度 :

[0256]

[0257] where η is the learning rate.

[0258] 5.3.4: The mathematical model of the finally found optimal solution is as follows:

[0259]

[0260] In the formula, x is the input sample, x k is the training sample belonging to class j, Cj is the set of all training samples of class k.

[0261] In step 5.3 of the present invention, an optimization algorithm is proposed to optimize the probability neural network (PNN) for the demagnetization fault and health state assessment model of wind turbines. The PNN is optimized by the gradient descent method to improve the accuracy of fault diagnosis. This model can be used in cases where data is scarce and can effectively solve the problem of fault diagnosis of wind turbines caused by insufficient fault data.

[0262] Step 5.4: According to the optimal solution mathematical model obtained in step 5.3, 70% of the feature vectors randomly selected from the feature vector database are used as the training set and input into the optimal solution mathematical model, that is, equation (47), to train the probability neural network.

[0263] After the detection data of the wind turbine is input into VMD for decomposition and the feature vectors extracted by VMD are input into the probability neural network, the probability neural network will calculate the similarity between the feature vectors and the probability density functions of the data in various states in the training set, and then classify the detection data of the wind turbine, so as to predict the fault type of each input detection data.

[0264] As Figure 5 shown: The right image represents the fault state of 10% rotor demagnetization, and its amplitude is significantly greater than the healthy state. The image of the healthy state shows a more consistent and stable vibration mode, and the vibration intensity is significantly lower than the 10% demagnetization state.

[0265] As Figure 6 shown: It is the signal image obtained by decomposing in step 4, showing the spectrogram obtained by VMD decomposition of the healthy state and the 10% rotor demagnetization fault state. The Mode obtained by VMD decomposition has different characteristic scales. It can be seen from the two figures that the amplitude increases in multiple frequency bands in the fault state, and these changes can indicate that the rotor demagnetization fault has an impact on the vibration characteristics of the wind turbine. Then, the energy entropy of each Mode can be extracted as the feature vector of the wind turbine state for subsequent use.

[0266] As Figure 7 shown: It is the effect diagram obtained by training the neural network in step 5.4 of the present invention. From the left image, the classification effect of the samples by the probability neural network after training can be known. The relationship between the sample number and the classification result is shown in the image. It can be seen from the figure that the samples are correctly classified, but there are a few classification errors; from the right image, the accuracy after training by the probability neural network can be known, reaching 96.30%, which means that 96.30% of all samples are correctly classified.

[0267] As Figure 8Shown: This is the effect diagram of classifying the faults of the wind turbine in step 5.5 of the present invention. It can be seen from the figure that in the classification results of the prediction samples of six states by the probabilistic neural network, the classification performance of the probabilistic neural network is perfect and there is no misclassification.

[0268] In the practical application of the present invention, based on the actual operating conditions of the wind turbine, the vibration signals of the measuring points on the surface of the motor stator are collected by the PCB acceleration sensor, the vibration signals are decomposed to obtain multiple modal components, the fitness function of the modal components is calculated, the lowest fitness function is selected for decomposition, judgment weights are assigned to the vibration signal multiple frequencies, the comprehensive eigenvalue is obtained through weighted calculation, the comprehensive characteristic calculation value of the vibration signal is compared with the database, the first step is to determine whether the wind turbine has a structural fault. If it is determined that there is a structural fault, then the second step is to determine the type of the wind turbine fault. During the determination, if the calculation result does not meet the characteristic range of the fault operating state, it is regarded as a bad point and this calculation result is determined to be invalid, and the vibration signal is collected again and the above-mentioned characteristic extraction and calculation steps are repeated. If the comparison result meets the characteristic range of the demagnetization fault state in the database, it is determined that it conforms to the corresponding wind turbine demagnetization fault state in the database, and the detection of the demagnetization state of the wind turbine is realized.

[0269] The above are only the embodiments of the present application and do not impose any form of limitation on the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications using the disclosed technical content within the scope of the technical solution of the present application, which are equivalent to equivalent implementation cases and all belong to the scope of the technical solution.

Claims

1. A method for diagnosing wind turbine rotor demagnetization faults using multi-source information processing, characterized in that: The following steps are involved: Step 1: Establish a digital twin model of the permanent magnet motor; Step 2: Perform electromagnetic-structural field coupling simulation on the digital twin model of the wind turbine generator, and obtain the electromagnetic force of the motor and the vibration acceleration on the stator and rotor by calculating the air gap magnetomotive force and air gap permeability coefficient; Step 3: Analyze the vibration characteristics of different measuring points under different demagnetization degrees of the wind turbine based on the digital twin model, and establish a characteristic database; Step 4: Preprocess the data prepared in the database; Step 5: The extracted feature vectors are classified through a neural network to obtain a neural network. After the formation, the signal is input into the neural network to complete the diagnosis and obtain the fault type.

2. The method for wind turbine rotor demagnetization fault diagnosis based on multi-source information processing according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Establish a digital twin model of a permanent magnet wind turbine based on the structural design parameters of the wind turbine; Step 1.2: A two-dimensional model is established based on the wind turbine parameters, and then a fault model is established. The coercive force of the permanent magnet is modulated by 10%, 20%, 30%, 40% and 50% of the original value, corresponding to 10%, 20%, 30%, 40% and 50% permanent magnet demagnetization faults, respectively.

3. The method for wind turbine rotor demagnetization fault diagnosis based on multi-source information processing according to claim 1, characterized in that: The step 2 comprises the following steps: Step 2.1: Based on the electromagnetic parameters in the wind turbine design parameters, including: rated voltage, winding method, number of permanent magnets, stator and rotor materials, etc., the electromagnetic field simulation of the wind turbine digital twin model is performed to obtain the electromagnetic force by calculating the air gap magnetomotive force and air gap permeability coefficient; b(θ,t)=f(θ,t)λ(θ,t) Where f(θ,t) is the air gap magnetomotive force; λ(θ,t) is the air gap permeability coefficient; Step 2.2: Based on the resulting material properties of the digital twin model of the wind turbine, including the material density, elastic modulus and Poisson's ratio of each structure as the basic material properties, the electromagnetic force loaded on the stator obtained by electromagnetic field simulation calculation is used as input for structural field vibration simulation.

4. The method for wind turbine rotor demagnetization fault diagnosis based on multi-source information processing according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Obtain fault characteristics based on the demagnetization state of the wind turbine digital twin model, including: the amplitude of the time-domain vibration signal at multiple measuring points on the stator and rotor surfaces, the conversion of the vibration signal at multiple measuring points to the frequency domain, the amplitude and change of the vibration acceleration at one times the fundamental frequency, two times the fundamental frequency, and multiple frequencies thereof; Step 3.2: Modify the wind turbine digital twin model through historical fault data to make its calculation more accurate; Step 3.3: Perform electromagnetic-structural coupling simulation on the digital twin model of the permanent magnet wind turbine in normal state and under different demagnetization conditions to extract the vibration acceleration of the measuring point under normal operation and different fault conditions. Step 3.4: Repeat step 3.3 to the set number of times to obtain the vibration acceleration at different measuring points as analysis data samples. At the same time, perform vibration tests and obtain vibration acceleration, and correct the analysis data samples obtained by simulation calculations. The vibration acceleration amplitude including the vibration fundamental frequency and multiple frequencies under normal conditions and different fault conditions is taken as the frequency domain characteristics of the vibration signal and a feature database is constructed.

5. The method for wind turbine rotor demagnetization fault diagnosis based on multi-source information processing according to claim 1, characterized in that: The step 4 comprises the following steps: Step 4.1: Decompose the signal and separate the fault characteristics of the system vibration. First, construct the constrained variational model as follows: In the formula, {u k }={u1,…,u K } represents the k IMF modal components obtained by VMD decomposition, {w k }={w1,…,w K } is the set of center frequencies of each modal component, K 模态 is the number of modes obtained by decomposition, represents partial derivative operation, δ(t) represents Dirac function, * represents convolution, and f(t) is the original signal; Then, the penalty factor α and the Lagrange multiplication operator λ(t) are introduced, and the operator alternating direction method is used to calculate { uk }、{ wk } and λ are updated alternately and iteratively to find the optimal solution of the constrained variational model. The results are as follows: In the formula They are f(t), Fourier transform of , n is the number of iterations; Step 4.2: After decomposing the healthy state, the fault states of rotor demagnetization 10%, 20%, 30%, 40% and 50% into Modes through VMD, the energy entropy of each Mode decomposed by VMD is extracted as a feature vector.

6. The method for wind turbine rotor demagnetization fault diagnosis based on multi-source information processing according to claim 1, characterized in that: The step 5 comprises the following steps: Step 5.1: Integrate the feature vectors of the healthy state, the rotor demagnetization 10%, 20%, 30%, 40% and 50% fault states, and add corresponding labels to the healthy state, the rotor demagnetization 10%, 20%, 30%, 40% and 50% fault states; Step 5.2: Use a probabilistic neural network to classify the integrated feature vectors; Step 5.3: Optimize the probabilistic neural network, that is, optimize and minimize the loss function so that the distribution of the probabilistic neural network is as consistent as possible with the true category distribution. By iteratively updating the parameters, an optimal solution is finally found to improve the classification accuracy of the probabilistic neural network. Step 5.4: Train the probabilistic neural network; Step 5.5: After the wind turbine detection data is input into VMD for decomposition and the feature vector extracted from VMD is input into the probabilistic neural network, the probabilistic neural network will calculate the similarity between the feature vector and the probability density function of the data of various states in the training set, and then classify the wind turbine detection data, thereby predicting the fault type of each input detection data.

Images