WO2022166527A1 - Feedback optimization-based wind turbine blade fault monitoring method - Google Patents

Abstract

A feedback optimization-based wind turbine blade fault monitoring method, comprising the following steps: step I, constructing a data set for wind turbine blade state monitoring; step II, constructing a support vector machine model to segment the feature space of wind turbine blade data; step III, performing optimization solution on parameters of the support vector machine model; and step IV, evaluating the support vector machine model. Said method aims to solve the technical problem that in the existing wind turbine blade fault monitoring method, a single factor is considered, and the single factor does not necessarily have an absolute correlation with a wind turbine blade fault, which causes that a wind turbine blade fault cannot be effectively and accurately monitored; in addition, the problem of wind turbine blade faults is complex, and the existing algorithm has a large number of iterations and a large amount of calculation and is easy to fall into a local optimal solution, not being able to well meet the requirements of wind turbine blade fault monitoring.

Description

A feedback optimization method for fault monitoring of fan blades technical field

The invention belongs to the technical field of wind power generation, and in particular relates to a method for wind turbine blade fault monitoring which integrates technologies such as support vector machine and particle swarm algorithm.

Background technique

Under the background of accelerating the adjustment and optimization of industrial structure, energy structure, and vigorously developing new energy, my country's wind power installed capacity is increasing steadily. Wind power generation will not generate greenhouse gases during the operation, and will not cause damage to the ecological environment. It is a new energy industry vigorously developed in my country. At present, timely and reliable detection of fan blade failures is an important task for power plant operation and maintenance personnel.

Support vector machine SVM (Support Vector Machine) can solve nonlinear segmentation problems in high-dimensional feature space. Particle swarm optimization is a bionic optimization algorithm. By simulating the foraging process of birds, a mathematical model is established. This algorithm Strong global search capability.

At present, the main researches on the fault monitoring of wind turbine blades are as follows:

A fault diagnosis of fan blade cracks based on wavelet packet analysis proposed by Liu Xiaobo et al. This method needs to monitor the vibration signal of the blade through the sensor, decompose the vibration signal by wavelet packet, reconstruct and analyze the characteristic changes in the frequency domain of each signal, Detect fan blade crack faults.

A fault diagnosis method for wind turbine blades based on deep convolutional autoencoders and XGBoost proposed by researchers at the University of Electronic Science and Technology of China Then, the multi-layer deep convolutional auto-encoder is used for unsupervised feature extraction, and the feature vector with fault information is obtained, and then the feature vector is input into XGBoost for feature learning and classification, and finally the fault intelligence of wind turbine blades is realized. detection.

By analyzing the relationship between aerodynamic signals and fan blade failures, Li Shaohui et al. believed that cracks in the blades would cause changes in some frequency bands of the aerodynamic signals. Car M et al. inspected wind turbine blades by a LiDAR-loaded drone. Joshuva A et al. carried out heuristic detection of wind turbine blade faults by establishing a decision tree for the vibration signal of wind turbine blades.

The above research only considers a single factor, and the single factor does not necessarily have an absolute correlation with the failure of the fan blade. At the same time, the problem of fan blade failure is complex. The existing algorithm has a large number of iterations and a large amount of calculation, and the algorithm is easy to fall into the local optimal solution. .

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the problem that the existing fan blade fault monitoring method considers a single factor, and the single factor does not necessarily have absolute correlation with the fan blade fault, which causes the technical problem that the fan blade fault cannot be effectively monitored accurately, Moreover, the problem of fan blade failure is complex, the existing algorithm has many iterations and a large amount of calculation, and the algorithm is easy to fall into the local optimal solution, which cannot well meet the needs of fan blade fault monitoring.

For solving the above-mentioned technical problems, the technical scheme adopted in the present invention is:

A support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm optimization, comprising the following steps:

Step 1: Build a data set for fan blade condition monitoring;

Step 2: Build a support vector machine model to segment the feature space of the fan blade data;

Step 3: Optimize and solve the parameters of the support vector machine model;

Step 4: Evaluate the support vector machine model.

In step 1, the data set consists of D samples (X _i , Y _i ), where X _i ∈ R ⁿ , Y _i ∈ {-1, +1}, wherein the operation data of the fan blades are collected as feature vectors X _i , the fan operation status label Y (normal is +1, fault is -1)) is collected at the same time, and each feature is normalized.

The above normalization process is:

其中μ为X _i特征的样本均值，σ为X _t特征的样本标准差。

where μ is the sample mean of the X _i feature, and σ is the sample standard deviation of the X _t feature.

The above operating data includes wind speed, generator speed, ambient temperature, wind direction angle, yaw position, horizontal acceleration, vertical acceleration, angle, speed, switching temperature of blade 1, blade 2, and blade 3, and the DC current of the switching DC device. , one or more of the cabin temperature.

In the second step, the radial basis function is used to construct a nonlinear segmentation support vector machine model to segment the feature space of the fan blade data.

The classification decision function of the support vector machine model constructed above is:

where K(x,y) is chosen as the radial basis function:

α _i的值域是[0，C]，C是惩罚系数；b为偏移量，sgn为符号函数

The value range of α _i is [0, C], C is the penalty coefficient; b is the offset, and sgn is the sign function

The specific optimization objective function is:

Constraints are: 0≤α _i ≤C, i=1,2,....D

In step 3, the particle swarm algorithm with improved learning factor is used to optimize and solve the α _i , σ parameters (i=1, 2, ....., D) of the support vector machine model.

The velocity and position of the particle in the above particle swarm optimization are vectors in the domain space of the optimization function, that is, v and x are vectors in the parameter space of α _i , σ (i=1,2,....,D), velocity and position The iterative update of is as follows:

x ^(t+1) = x ^(t) + v ^(t)

where C ₁ and C ₂ are learning factors, v is the velocity of the particle, x is the position of the particle, p _best is the optimal solution found by the individual particle in the objective optimization function, and g _best is the optimal solution found by the particle group in the objective optimization function. The optimal solution, t represents the t-th iteration, and rand(0,1) represents any value in the (0,1) interval.

Specifically, the following optimization steps are included:

1) Set the iterative step size m of the particle swarm algorithm, and set the number of particles N;

和初始位置

其中

表示第i个粒子的第j次迭代的速度值，

表示第i个粒子的第j次迭代的位置，径向基函数的σ＝[0.01,1000]，惩罚系数C∈[0.1,500]，学习因子C ₁和C ₂为每一次迭代所有粒子速度和位置向量的L ²范数的平方和的柯西概率的一定倍数； 2) Set the initial velocity of the particles

and initial position

in

represents the velocity value of the jth iteration of the ith particle,

represents the position of the jth iteration of the ith particle, the radial basis function σ=[0.01,1000], the penalty coefficient C∈[0.1,500], and the learning factors _C1 and _C2 are all particle velocities for each iteration and some multiple of the Cauchy probability of the sum of squares of the L2 norm of the position vector ^;

3) According to the initialization conditions, calculate the fitness value of the objective function;

其中第t次(0≤t≤N)迭代的

4) According to the fitness value, cyclically update the speed and position of the particle, the optimal solution p _best of the individual, and the optimal solution g _best of the group, where the jth (0≤j≤m) particle is the tth time (0≤t≤ N) iterative

where the t-th (0≤t≤N) iteration is

min()

represents the minimum value.

5) According to the termination conditions, end the algorithm optimization process: the maximum number of iterations is reached or the L ² norm of the speed and position of the two generations is less than the set value, namely: ||v ^(t+1) -v ^(t) ||≤κ And ||x ^(t+1) -x ^(t) ||≤γ, where κ and γ are precision settings.

In step 4, the optimal SVM model parameter values are obtained, and the SVM model is verified. If the target accuracy rate is not met, the SVM model will be retrained. At this time, the initial velocity and position of the particles are set as the optimal SVM model. The value corresponds to the speed and position multiplied by a certain weight.

Among them, the result of the particle swarm optimization optimization is divided into the feature space by the support vector machine. After evaluation, if the result is not satisfactory, the evaluation result is fed back to the particle swarm optimization algorithm, and the optimization is continued, and the search range is enlarged, so as to Avoid getting stuck in a local optimum.

Compared with the prior art, the present invention has the following technical effects:

1. The present invention uses the Cauchy probability density as a learning factor, and it is not easy to fall into the local optimal solution. Further, the present invention adopts a negative feedback method to avoid the optimization process from falling into the local optimal solution. The result of particle swarm optimization optimization , After the support vector machine is divided in the feature space, after the evaluation, if the result is not satisfactory, the evaluation result is fed back to the particle swarm algorithm, and the optimization is continued, and the search range is enlarged, so as to avoid falling into the local optimal solution;

2. After the cross-validation of the SVM model, if the accuracy rate does not reach the target value, the SVM model needs to be retrained. At this time, the initial speed and position of the particles are set to the current optimal SVM model value corresponding to the training. The particle velocity and position are multiplied by a weight of a certain multiple, in order to let the particle jump out of the local optimal solution when continuing to find the global optimal solution and prevent local oscillation;

3. The present invention uses the characteristics of various fan blades as the input of the model, which greatly ensures the accuracy of the model;

4. The present invention adopts the support vector machine model optimized by the particle swarm algorithm with the improved learning factor to detect the fault of the fan blade, solves the problem that the nonlinear SVM model is difficult to optimize, and is easy to fall into the local optimal area, and at the same time improves the fault diagnosis of electrical equipment. reliability.

Description of drawings

Below in conjunction with accompanying drawing and embodiment, the present invention will be further described:

FIG. 1 is a flow chart of the present invention.

Detailed ways

As shown in Figure 1, a feedback optimization support vector machine fan blade fault monitoring method includes the following steps:

Step 1: Build a data set for fan blade condition monitoring;

Step 2: Build a support vector machine model to segment the feature space of the fan blade data;

Step 3: Optimize and solve the parameters of the support vector machine model;

Step 4: Evaluate the support vector machine model.

In step 1, the data set consists of D samples (X _i , Y _i ), where X _i ∈ R ⁿ , Y _i ∈ {-1, +1}, wherein the operation data of the fan blades are collected as feature vectors X _i , the fan operation status label Y (normal is +1, fault is -1)) is collected at the same time, and each feature is normalized.

The standardization process is:

其中μ为X _i特征的样本均值，σ为X _t特征的样本标准差。

where μ is the sample mean of the X _i feature, and σ is the sample standard deviation of the X _t feature.

Among them, the operating data includes wind speed, generator speed, ambient temperature, wind direction angle, yaw position, horizontal acceleration, vertical acceleration, angle, speed, switching temperature of blade 1, blade 2, blade 3, switching DC DC One or more of current and cabin temperature.

In the second step, the radial basis function is used to construct a nonlinear segmentation support vector machine model to segment the feature space of the fan blade data.

The classification decision function of the constructed support vector machine model is:

where K(x,y) is chosen as the radial basis function:

α _i的值域是[0，C]，C是惩罚系数，b为偏移量，sgn为符号 函数

；

The value range of α _i is [0, C], C is the penalty coefficient, b is the offset, and sgn is the sign function

;

The specific optimization objective function is:

Constraints are: 0≤α _i ≤C, i=1,2,....D

In step 3, the particle swarm algorithm with improved learning factor is used to optimize and solve the α _i , σ parameters (i=1, 2, ....., D) of the support vector machine model.

Among them, the speed and position of the particle in the particle swarm optimization algorithm are the vectors in the domain space of the optimization function, that is, v and x are the vectors in the parameter space of α _i , σ (i=1,2,....,D);

The iterative update of velocity and position is as follows:

x ^(t+1) = x ^(t) + v ^(t) ;

where C ₁ and C ₂ are learning factors, v is the velocity of the particle, x is the position of the particle, p _best is the optimal solution found by the individual particle in the objective optimization function, and g _best is the optimal solution found by the particle group in the objective optimization function. The optimal solution, t represents the t-th iteration, and rand(0,1) represents any value in the (0,1) interval.

When performing the optimization solution, first set the number of particles, the speed and position of each particle are vectors in the domain space of the optimization function, calculate the fitness value of each particle, update the speed v and position x of each particle, and each particle The speed and position of is related to the individual optimal solution p _best and the group optimal solution g _best . In order to smooth the solution process, the Cauchy probability density is used as the learning factor.

The present invention uses the Cauchy probability density as the learning factor, so it is not easy to fall into the local optimal solution. Further, the present invention adopts the negative feedback method to avoid the optimization process from falling into the local optimal solution. The support vector machine is divided in the feature space. After evaluation, if the result is not satisfactory, the evaluation result will be fed back to the particle swarm algorithm, and the optimization will continue to be searched, and the search range will be enlarged, so as to avoid falling into the local optimal solution;

As implementable, the following optimization steps are provided:

1) Set the iterative step size of the particle swarm algorithm m=1500, and set the number of particles N=20;

和初始位置

其中

表示第i个粒子的第j次迭代的速度值，

表示第i个粒子的第j次迭代的位置，径向基函数的σ∈[0.01,1000]，惩罚系数C∈[0.1,500]，学习因子C ₁和C ₂为每一次迭代所有粒子速度和位置向量的L ²范数的平方和的柯西概率的1.5倍和2.5倍， 2) Set the initial velocity of the particles

and initial position

in

represents the velocity value of the jth iteration of the ith particle,

represents the position of the jth iteration of the ith particle, the radial basis function σ∈[0.01,1000], the penalty coefficient C∈[0.1,500], and the learning factors _C1 and _C2 are all particle velocities for each iteration and 1.5 times and ^2.5 times the Cauchy probability of the sum of squares of the L2 norm of the position vector,

The learning factors C ₁ and C ₂ are certain multiples of the Cauchy probability of the sum of the squares of the L ² norm of all particle velocities and position vectors at each iteration;

The Cauchy probability is used as a learning factor. Because the Cauchy probability density function is flat, it is not easy for the particle swarm optimization algorithm to enter the local optimal solution.

3) According to the initialization conditions, calculate the fitness value of the objective function;

其中第t次(0≤t≤N)迭代的

4) According to the fitness value, cyclically update the speed and position of the particle, the optimal solution p _best of the individual, and the optimal solution g _best of the group, where the jth (0≤j≤m) particle is the tth time (0≤t≤ N) iterative

where the t-th (0≤t≤N) iteration is

5) According to the termination conditions, end the algorithm optimization process: the maximum number of iterations is reached or the L ² norm of the speed and position of the two generations is less than the set value, namely: ||v ^(t+1) -v ^(t) ||≤κ And ||x ^(t+1) -x ^(t) ||≤γ, κ=0.4 and γ=0.8 can be selected.

In step 4, the optimal SVM model parameter values are obtained, and 10-fold cross-validation is performed on the SVM model. If the target accuracy rate is not met, the SVM model will be retrained. At this time, the initial velocity and position of the particles are set to be optimal The SVM model values corresponding to velocity and position are multiplied by 1.5 times the weight.

Claims (10)

A feedback-optimized fan blade fault monitoring method, characterized in that it includes the following steps: Step 1: Build a data set for fan blade condition monitoring; Step 2: Build a support vector machine model to segment the feature space of the fan blade data; Step 3: Optimize and solve the parameters of the support vector machine model; Step 4: Evaluate the support vector machine model. The support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm optimization according to claim 1 is characterized in that, in step 1, the data set is composed of D samples (X _i , Y _i ) enough, where X _i ∈ R ⁿ , Y _i ∈ {-1, +1}, in which, the operation data of the fan blades are collected as the feature vector X _i , and the fan operation status label Y (normal is +1, fault is -1) is collected at the same time. ), and normalize each feature. The support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm algorithm according to claim 2, is characterized in that, described standardization process is:

where μ is the sample mean of the X _i feature, and σ is the sample standard deviation of the X _t feature. The support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm algorithm according to claim 2, wherein the operation data includes wind speed, generator speed, ambient temperature, wind direction angle, yaw position, One or more of horizontal acceleration, vertical acceleration, angle, speed, switching temperature of blade 1, blade 2, blade 3, DC current of switching DC, and cabin temperature. The support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm algorithm according to claim 2, characterized in that: in step 2, a radial basis function is used to construct a nonlinear segmentation support vector machine model to segment Feature space for wind turbine blade data. The support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm algorithm according to claim 5, is characterized in that, the classification decision function of the constructed support vector machine model is:

where K(x,y) is chosen as the radial basis function:

The value range of α _i is [0, C], C is the penalty coefficient, and b is the offset; sgn is a symbolic function

The specific optimization objective function is:

Constraints are: 0≤α _i ≤C, i=1,2,....D

The SVM fan blade fault monitoring method based on improved learning factor particle swarm optimization according to claim 6, characterized in that, in step 3, the particle swarm optimization with improved learning factor is used to optimize and solve the α of the support vector machine model. _i , the σ parameter (i=1,2,...,D). The support vector machine fan blade fault monitoring method optimized by the improved learning factor particle swarm algorithm according to claim 7, is characterized in that, the speed and position of the particles in the particle swarm algorithm are vectors in the domain space of the optimization function, namely v and x is a vector of α _i , σ (i=1,2,....,D) parameter space, and the iterative update of velocity and position is as follows:

x ^(t+1) = x ^(t) + v ^(t) where C ₁ and C ₂ are learning factors, v is the velocity of the particle, x is the position of the particle, p _best is the optimal solution found by the individual particle in the objective optimization function, and g _best is the optimal solution found by the particle group in the objective optimization function. The optimal solution, t represents the t-th iteration, and rand(0,1) represents any value in the (0,1) interval. The support vector machine fan blade fault monitoring method optimized based on improved learning factor particle swarm optimization according to claim 8, is characterized in that, specifically comprises the following optimization steps: 1) Set the iterative step size m of the particle swarm algorithm, and set the number of particles N; 2) Set the initial velocity of the particles

and initial position

in

represents the velocity parameter of the jth iteration of the ith particle,

represents the jth iteration position parameter of the ith particle, σ∈[0.01,1000] of the radial basis function, penalty coefficient C∈[0.1,500], learning factors _C1 and _C2 are all particle velocities for each iteration and some multiple of the Cauchy probability of the sum of squares of the L2 norm of the position vector ^;

3) According to the initialization conditions, calculate the fitness value of the objective function; 4) According to the fitness value, cyclically update the speed and position of the particle, the optimal solution p _best of the individual, and the optimal solution g _best of the group, where the jth (0≤j≤m) particle is the tth time (0≤t≤ N) iterative

where the t-th (0≤t≤N) iteration is

min() represents the minimum value; 5) According to the termination conditions, end the algorithm optimization process: the maximum number of iterations is reached or the L ² norm of the speed and position of the two generations is less than the set value, namely: ||v ^(t+1) -v ^(t) ||≤κ And ||x ^(t+1) -x ^(t) ||≤γ, where κ and γ are precision settings. According to one of claims 1 to 9, the support vector machine fan blade fault monitoring method based on improved learning factor particle swarm algorithm optimization is characterized in that, in step 4, the optimal SVM model parameter value is obtained, and the SVM The model is verified. If the target accuracy rate is not met, the SVM model will be retrained. At this time, the initial speed and position of the particle are set to the speed and position corresponding to the optimal SVM model value multiplied by a certain weight.

Images