CN105606914A

CN105606914A - IWO-ELM-based Aviation power converter fault diagnosis method

Info

Publication number: CN105606914A
Application number: CN201510559194.3A
Authority: CN
Inventors: 崔江; 叶纪青; 龚春英
Original assignee: Nanjing University of Aeronautics and Astronautics
Current assignee: Nanjing University of Aeronautics and Astronautics
Priority date: 2015-09-06
Filing date: 2015-09-06
Publication date: 2016-05-25

Abstract

The invention relates to an aviation power converter fault diagnosis method based on a weed algorithm (Invasive? Weed? Optimization, referred to as IWO) and an extreme learning machine (Extreme? Learning? Machine, referred to as ELM), belonging to the field of circuit fault diagnosis. The method includes the following steps: 1) collecting the output signals of the measurable nodes of the aviation power converter in the normal mode and the fault mode; 2) extracting the key features of the collected signals by using the principal component analysis (Principal? Component? Analysis, referred to as PCA) method , to construct a feature sample set; 3) divide the feature sample set into a training sample set and a test sample set, which are used for the training and evaluation of the extreme learning machine respectively; 4) divide the training sample set into training data and test data, and use the training data Train the extreme learning machine, and use the weed algorithm to optimize the number of hidden layer nodes, input weight and hidden layer node bias, so that the extreme learning machine has an optimized classifier structure; 5) The test sample set is used for evaluation Optimized extreme learning machine fault diagnosis performance.

Description

A Fault Diagnosis Method for Aviation Power Converter Based on IWO-ELM

技术领域technical field

本发明涉及一种基于IWO-ELM的航空功率变换器故障诊断方法，属于电路故障诊断领域。The invention relates to an IWO-ELM-based fault diagnosis method for an aviation power converter, belonging to the field of circuit fault diagnosis.

背景技术Background technique

随着航空技术的不断发展，目前飞机上配载的电源系统可分为如下三种类型：低压直流电源、高压直流电源和交流电源。这些电源系统的正常运转，保证了机载设备的正常工作和飞机的安全飞行。航空电源系统中的关键模块为功率变换器，一旦功率变换器发生故障，将直接影响航空电源系统的正常运行，进而对飞机的安全运行构成威胁，甚至会带来巨大的生命财产和安全损失。因此，对航空电源系统中功率变换器进行故障诊断研究，具有十分重大的意义。With the continuous development of aviation technology, the current power supply systems onboard aircraft can be divided into the following three types: low-voltage DC power supply, high-voltage DC power supply and AC power supply. The normal operation of these power systems ensures the normal operation of the airborne equipment and the safe flight of the aircraft. The key module in the aviation power system is the power converter. Once the power converter fails, it will directly affect the normal operation of the aviation power system, which will pose a threat to the safe operation of the aircraft, and even cause huge loss of life, property and safety. Therefore, it is of great significance to study the fault diagnosis of the power converter in the aviation power system.

目前，对航空电源系统中的功率变换器进行故障诊断的方法大致分为基于信号处理的方法、基于模型的方法和基于知识的方法。基于信号处理的方法不需要对对象建模，适用性比较强，形式简单且易于实现，但由于其依赖信号的检测和处理，通常会受信号噪声的影响，局限于特定信号诊断某些特定故障，未能考虑各种故障间的相互影响，当诊断对象变得庞大复杂时，通常需要增加检验手段和计算量，因而该方法不适合对航空电源系统中功率变换器进行故障诊断。基于模型的方法是通过建立对象的模型，将模型预测值与系统观测值进行比较来判别有无故障，但航空电源系统中功率变换器结构千变万化，通用性较差，不适合采用建模的方法进行故障诊断。基于知识的方法不依赖对象的数学模型，从而可以实现对难于建模的对象的故障诊断，在故障诊断领域应用中具有较大优势。因而，本发明考虑通过构建极限学习机模型这一种基于模式识别的方法对航空功率变换器进行故障诊断。At present, the methods for fault diagnosis of power converters in aviation power systems are roughly divided into signal processing-based methods, model-based methods and knowledge-based methods. The method based on signal processing does not need to model the object, has strong applicability, simple form and easy implementation, but because it relies on signal detection and processing, it is usually affected by signal noise and is limited to specific signal diagnosis of certain specific faults , failed to consider the interaction between various faults, and when the diagnostic object becomes large and complex, it usually needs to increase the inspection means and the amount of calculation, so this method is not suitable for fault diagnosis of power converters in aviation power systems. The model-based method is to establish a model of the object and compare the predicted value of the model with the observed value of the system to determine whether there is a fault. However, the structure of the power converter in the aviation power system is ever-changing and the versatility is poor, so it is not suitable for the modeling method. Perform troubleshooting. The knowledge-based method does not rely on the mathematical model of the object, so it can realize the fault diagnosis of the object that is difficult to model, and has great advantages in the field of fault diagnosis. Therefore, the present invention considers the fault diagnosis of the aviation power converter by constructing the extreme learning machine model, a method based on pattern recognition.

极限学习机是一种适用于单隐含层前馈神经网络的学习算法模型，仅需在网络训练前设置合适的隐含层节点数，并随机初始化输入权值以及隐含层节点偏置，即可通过解析运算获得隐含层输出权值。相对于以往的神经网络训练过程，极限学习机模型无需多次迭代运算，整个模型训练一次即可完成。但是极限学习机采用随机初始化输入权值以及隐含层节点偏置的方法，所建立的极限学习机模型会存在隐含层节点数过多、过拟合等现象，从而影响了极限学习机模型的预测速度和精度。杂草算法是一种从自然界杂草进化原理演化而来的随机搜索算法，通过充分利用种群中的优秀个体指导群体的进化，兼顾了种群的选择力度和多样性，能够有效克服不成熟收敛，并且具有算法结构简单、参数少和鲁棒性较好等特点。因而，使用杂草算法优化极限学习机模型，可以改善极限学习机隐含层节点数过多、过拟合等缺陷。Extreme learning machine is a learning algorithm model suitable for single hidden layer feedforward neural network. It only needs to set the appropriate number of hidden layer nodes before network training, and randomly initialize the input weight and hidden layer node bias. The output weight of the hidden layer can be obtained through analytical operation. Compared with the previous neural network training process, the extreme learning machine model does not need multiple iterative operations, and the entire model training can be completed once. However, the extreme learning machine adopts the method of randomly initializing the input weights and the bias of the hidden layer nodes. The established extreme learning machine model will have too many hidden layer nodes and overfitting, which affects the extreme learning machine model. prediction speed and accuracy. The weed algorithm is a random search algorithm evolved from the evolution principle of weeds in nature. By making full use of the excellent individuals in the population to guide the evolution of the population, taking into account the selection strength and diversity of the population, it can effectively overcome immature convergence. And it has the characteristics of simple algorithm structure, few parameters and good robustness. Therefore, using the weed algorithm to optimize the extreme learning machine model can improve the extreme learning machine's defects such as too many nodes in the hidden layer and overfitting.

发明内容Contents of the invention

为了解决上述问题，本发明提出一种基于IWO-ELM的航空功率变换器故障诊断方法，一方面解决了传统神经网络训练速度缓慢的问题，另一方面解决了极限学习机模型隐含层节点数等参数的优化问题，提高了航空功率变换器诊断正确率。In order to solve the above problems, the present invention proposes a fault diagnosis method for aviation power converters based on IWO-ELM. On the one hand, it solves the problem of slow training speed of traditional neural networks; The optimization problem of parameters, etc. improves the diagnostic accuracy of aviation power converters.

本发明为实现上述目的，实现技术方案如下：In order to achieve the above object, the present invention realizes the technical scheme as follows:

首先采集航空功率变换器在正常模式以及故障模式下的可测节点输出信号，并利用主成分分析法提取采集信号的关键特征，构建特征样本集，然后将特征样本集分为训练样本集和测试样本集，分别用于极限学习机的训练和评估。First collect the measurable node output signals of the aviation power converter in normal mode and fault mode, and use the principal component analysis method to extract the key features of the collected signals, construct a feature sample set, and then divide the feature sample set into training sample set and test The sample sets are used for the training and evaluation of the extreme learning machine respectively.

具体的操作步骤如下所示：The specific operation steps are as follows:

1)获取航空功率变换器在正常模式以及故障模式下的可测节点输出信号，利用主成分分析法提取采集信号的关键特征，构建特征样本集，并将特征样本集分为两部分：训练样本集A＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，N}和测试样本集B＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，M}，其中，R代表实数空间，x_i为n维样本特征向量，y_i为m维样本标签，即x_i＝{x_i1，x_i2，...，x_in}，y_i＝{y_i1，y_i2，...，y_im}，训练样本集和测试样本集的样本个数分别为N个和M个。1) Obtain the measurable node output signals of the aeronautical power converter in normal mode and fault mode, use the principal component analysis method to extract the key features of the collected signals, construct a feature sample set, and divide the feature sample set into two parts: training samples Set A={(x _i , y _i )|x _i ∈ R ⁿ , y _i ∈ R ^m , i=1,..., N} and test sample set B={(x _i , y _i )|x _i ∈ R ⁿ , y _i ∈ R ^m , i=1,...,M}, where R represents the real number space, _xi is the n-dimensional sample feature vector, y _i is the m-dimensional sample label, that is, _xi = {x _i1 , x _i2 ,..., x _in }, y _i ={y _i1 , y _i2 ,..., y _im }, the number of samples in the training sample set and the test sample set are N and M respectively indivual.

然后，将训练样本集分为训练数据和测试数据两部分：训练数据A1＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，N₁}和测试样本集A2＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，N₂}，训练数据和测试数据的样本个数分别为N₁个和N₂个，且N＝N₁+N₂；Then, the training sample set is divided into two parts: training data and test data: training data A1={( _xi ,y _i )| _xi ∈R ⁿ , y _i ∈R ^m , i=1,...,N ₁ } and test sample set A2={( _xi , y _i )| _xi ∈ R ⁿ , y _i ∈ R ^m , i=1,..., N ₂ }, the number of samples of training data and test data N ₁ and N ₂ respectively, and N=N ₁ +N ₂ ;

2)随机给定隐含层节点数L，隐含层节点偏置b_j(j＝1，...，L)，输入权重w，激活函数g(x)，其中所选择的激活函数必须满足无穷阶次可微的条件。2) The number L of hidden layer nodes is randomly given, the hidden layer node bias b _j (j=1,...,L), the input weight w, and the activation function g(x), where the selected activation function must be Satisfy the condition of infinite order differentiability.

其中， $w = {[\begin{matrix} w_{1}^{T} & w_{2}^{T} & . . . & w_{L}^{T} \end{matrix}]}_{L \times n}^{T},$ w_j＝[w_j1w_j2...w_jn]_1×n，w_j表示第j个隐含层的输入权重矩阵，w_ji表示输入层第i个神经元与隐含层第j个神经元的连接权值，因而构建的极限学习机模型如图1所示；in, $w = {[\begin{matrix} w_{1}^{T} & w_{2}^{T} & . . . & w_{L}^{T} \end{matrix}]}_{L \times no}^{T},$ w _j ＝[w _j1 w _j2 ...w _jn ] _1×n , w _j represents the input weight matrix of the j-th hidden layer, w _ji represents the i-th neuron of the input layer and the j-th neuron of the hidden layer The connection weights of the elements, so the extreme learning machine model constructed is shown in Figure 1;

3)根据单隐含层神经网络学习机制，即存在β_jk(k＝1，...，m)，w_j，b_j使得如下公式成立：3) According to the learning mechanism of single hidden layer neural network, there exists β _jk (k=1,...,m), w _j , b _j so that the following formula holds true:

$\{\begin{matrix} {Σ Σ}_{j j = = 11}^{L L} {β β}_{j j 11} \cdot \cdot g g (({w w}_{j j} \cdot \cdot {x x}_{i i} + + {b b}_{j j})) = = {y the y}_{i i 11} \\ {Σ Σ}_{j j = = 11}^{L L} {β β}_{j j 22} \cdot \cdot g g (({w w}_{j j} \cdot \cdot {x x}_{i i} + + {b b}_{j j})) = = {y the y}_{i i 22} \\ . . \\ . . \\ . . \\ {Σ Σ}_{j j = = 11}^{L L} {β β}_{jm jm} \cdot \cdot g g (({w w}_{j j} \cdot \cdot {x x}_{i i} + + {b b}_{j j})) = = {y the y}_{im im} \end{matrix} ((i i = = 11,, . . . . . .,, {N N}_{11}))$

其中，β_j＝[β_j1β_j2…β_jm]_1xm，β_j表示第j个隐含层的输出权重矩阵，β_jk表示隐含层第j个神经元与输出层第k个神经元间的连接权值；Among them, β _j =[β _j1 β _j2 …β _jm ] _1xm , β _j represents the output weight matrix of the jth hidden layer, β _jk represents the distance between the jth neuron of the hidden layer and the kth neuron of the output layer The connection weight of ;

4)令4) order

$β β = = {[\begin{matrix} {β β}_{11}^{T T} & {β β}_{22}^{T T} & . . . . . . & {β β}_{L L}^{T T} \end{matrix}]}^{T T}$

$Y Y = = {[\begin{matrix} {y the y}_{11}^{T T} & {y the y}_{22}^{T T} & . . . . . . & {y the y}_{{N N}_{11}}^{T T} \end{matrix}]}^{T T}$

即可将式(1)简化为：The formula (1) can be simplified as:

H·β＝Y(5)H·β=Y(5)

其中H是N₁×L维的隐含层输出矩阵，β是L×m维隐含层输出权重矩阵，Y是N₁×m维期望输出矩阵；Where H is the N ₁ ×L-dimensional hidden layer output matrix, β is the L×m-dimensional hidden layer output weight matrix, and Y is the N ₁ ×m-dimensional expected output matrix;

5)根据Moore-Penrose广义逆矩阵定义，可根据如下公式求解隐含层输出权重矩阵β：5) According to the definition of Moore-Penrose generalized inverse matrix, the hidden layer output weight matrix β can be solved according to the following formula:

β＝(H^T·H)^-1.H^T·Y(6)β＝(H ^T ·H) ^-1 .H ^T ·Y(6)

6)利用杂草算法优化极限学习机模型中隐含层节点数，输入权重以及隐含层节点偏置的具体步骤如下：6) Using the weed algorithm to optimize the number of hidden layer nodes in the extreme learning machine model, the specific steps of input weight and hidden layer node bias are as follows:

6.1)设置杂草算法所需的最大迭代次数iter_max、最小种群数S_min与最大种群数S_max、最小适应度f_min与最大适应度f_max、标准偏差的初始值σ_initial和终值σ_final、非线性模型指数p，并将极限学习机随机产生的隐含层节点数，输入权重以及隐含层节点偏置作为杂草算法的初始种群，均匀分布在搜索空间中；6.1) Set the maximum number of iterations iter _max required by the weed algorithm, the minimum population number S _min and the maximum population number S _max , the minimum fitness f _min and the maximum fitness f _max , the initial value σ _initial and the final value σ of the standard deviation _final , the nonlinear model index p, and the number of hidden layer nodes randomly generated by the extreme learning machine, the input weight and the bias of hidden layer nodes are used as the initial population of the weed algorithm, and are evenly distributed in the search space;

6.2)杂草种群的个体繁殖取决于自身的适应度值以及最小适应度值和最大适应度值，本发明选取图2中极限学习机对测试数据进行测试所得的正确率作为适应度值，此处，正确率定义为：正确分类的样本数目/测试数据总数目×100％，因而繁殖的种子个数S_k可依据如下公式计算：6.2) The individual reproduction of the weed population depends on its own fitness value, minimum fitness value and maximum fitness value. The present invention selects the correctness rate obtained by the extreme learning machine in Fig. 2 to test the test data as the fitness value. , the correct rate is defined as: the number of correctly classified samples/total number of test data × 100%, so the number of seeds S _k to reproduce can be calculated according to the following formula:

${S S}_{k k} = = \frac{{f f}_{k k} - - {f f}_{min min}}{{f f}_{max max} - - {f f}_{min min}} \cdot \cdot (({S S}_{max max} - - {S S}_{min min})) + + {S S}_{min min} - - - - - - ((77))$

其中，k为当前迭代次数(k＝1，...，iter_max)，S_k为第k次迭代的繁殖种子数，f_k为第k次迭代的适应度值；Wherein, k is the current number of iterations (k=1, ..., iter _max ), S _k is the number of breeding seeds of the k iteration, f _k is the fitness value of the k iteration;

6.3)将步骤6.2)中产生的种子根据变方差σ_k以及正态分布规律随机地分布在搜索区域中，并靠近上一代个体，其中变方差计算公式如下：6.3) The seeds generated in step 6.2) are randomly distributed in the search area according to the variable variance σ _k and normal distribution, and close to the previous generation of individuals, where the variable variance calculation formula is as follows:

${σ σ}_{k k} = = \frac{{(({iter iter}_{max max} - - k k))}^{p p}}{{(({iter iter}_{max max}))}^{p p}} \cdot &Center Dot; (({σ σ}_{initial initial} - - {σ σ}_{final final})) + + {σ σ}_{final final} - - - - - - ((88))$

重复进行步骤6.2)和步骤6.3)，直到达到最大种群数；Repeat step 6.2) and step 6.3) until reaching the maximum number of populations;

6.4)将所有个体按照适应度值的大小进行降序排序，剔除适应度值低的个体，并重复进行步骤6.2)、步骤6.3)以及步骤6.4)，直到达到最大迭代次数；6.4) Sort all individuals in descending order according to the size of fitness value, remove individuals with low fitness value, and repeat step 6.2), step 6.3) and step 6.4), until the maximum number of iterations is reached;

7)杂草算法迭代运行结束后，利用步骤6.4)中最大适应度值对应的隐含层节点数、输入权重以及隐含层节点偏置构建极限学习机模型，并用测试样本集B＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，M}检验极限学习机的故障诊断性能，主要借助于故障诊断率判别分析。此处，故障诊断率定义为：正确分类的样本数目/测试样本集总数目×100％。7) After the iterative operation of the weed algorithm is completed, use the hidden layer node number, input weight and hidden layer node bias corresponding to the maximum fitness value in step 6.4) to construct an extreme learning machine model, and use the test sample set B={( x _i , y _i )| _xi ∈ R ⁿ , y _i ∈ R ^m , i=1,..., M} test the fault diagnosis performance of extreme learning machine, mainly by means of fault diagnosis rate discriminant analysis. Here, the fault diagnosis rate is defined as: the number of correctly classified samples/the total number of test sample sets×100%.

本发明有益效果如下：The beneficial effects of the present invention are as follows:

极限学习机采用随机初始化输入权值以及隐含层节点偏置的方法，建立的极限学习机模型会存在隐含层节点数过多、过拟合等现象，导致故障诊断性能较差，本发明使用杂草算法优化极限学习机，可以提高极限学习机的故障诊断正确率。The extreme learning machine adopts the method of randomly initializing the input weight and the bias of hidden layer nodes, and the established extreme learning machine model will have too many hidden layer nodes, overfitting and other phenomena, resulting in poor fault diagnosis performance. The present invention Using the weed algorithm to optimize the extreme learning machine can improve the fault diagnosis accuracy of the extreme learning machine.

附图说明Description of drawings

图1极限学习机模型结构图Figure 1 Structure diagram of extreme learning machine model

图2故障诊断流程图Figure 2 Fault Diagnosis Flowchart

具体实施方式detailed description

1)获取航空功率变换器在正常模式以及故障模式下的可测节点输出信号，利用主成分分析法提取采集信号的关键特征，构建特征样本集，并将特征样本集分为两部分：训练样本集A＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，N}和测试样本集B＝{(x_i，y_i)|x_i∈Rⁿ，y_i∈R^m，i＝1，...，M}，其中，R代表实数空间，x_i为n维样本特征向量，y_i为m维样本标签，即x_i＝{x_i1，_i2，...，x_in}，y_i＝{y_i1y_i2，...，y_im}，训练样本集和测试样本集的样本个数分别为N个和M个。1) Obtain the measurable node output signals of the aeronautical power converter in normal mode and fault mode, use the principal component analysis method to extract the key features of the collected signals, construct a feature sample set, and divide the feature sample set into two parts: training samples Set A={(x _i , y _i )|x _i ∈ R ⁿ , y _i ∈ R ^m , i=1,..., N} and test sample set B={(x _i , y _i )|x _i ∈ R ⁿ , y _i ∈ R ^m , i=1,...,M}, where R represents the real number space, _xi is the n-dimensional sample feature vector, y _i is the m-dimensional sample label, that is, _xi = {x _i1 , _i2 , . . . , x _in }, _y _i ={y _i1 y _i2 , .

2)随机给定隐含层节点数L，隐含层节点偏置b_i(j＝1，...，L)，输入权重w，激活函数g(x)，其中所选择的激活函数必须满足无穷阶次可微的条件。2) The number L of hidden layer nodes is randomly given, the hidden layer node bias b _i (j=1,...,L), the input weight w, and the activation function g(x), where the selected activation function must be Satisfy the condition of infinite order differentiability.

$\{\begin{matrix} {Σ Σ}_{j j = = 11}^{L L} {β β}_{j j 11} \cdot &Center Dot; g g (({w w}_{j j} \cdot &Center Dot; {x x}_{i i} + + {b b}_{j j})) = = {y the y}_{i i 11} \\ {Σ Σ}_{j j = = 11}^{L L} {β β}_{j j 22} \cdot &Center Dot; g g (({w w}_{j j} \cdot &Center Dot; {x x}_{i i} + + {b b}_{j j})) = = {y the y}_{i i 22} \\ . . \\ . . \\ . . \\ {Σ Σ}_{j j = = 11}^{L L} {β β}_{jm jm} \cdot \cdot g g (({w w}_{j j} \cdot \cdot {x x}_{i i} + + {b b}_{j j})) = = {y the y}_{im im} \end{matrix} ((i i = = 11,, . . . . . .,, {N N}_{11}))$

4)令4) order

即可将式(1)简化为：The formula (1) can be simplified as:

H·β＝Y(5)H·β=Y(5)

β＝(H^T·H)^-1·H^T·Y(6)β=(H ^T H) ^-1 H ^T Y (6)

${σ σ}_{k k} = = \frac{{(({iter iter}_{max max} - - k k))}^{p p}}{{(({iter iter}_{max max}))}^{p p}} \cdot \cdot (({σ σ}_{initial initial} - - {σ σ}_{final final})) + + {σ σ}_{final final} - - - - - - ((88))$

Claims

1. An IWO-ELM-based aviation power converter fault diagnosis method is characterized by comprising the following basic steps of:

1) acquiring measurable node output signals of the aviation power converter in a normal mode and a fault mode;

2) extracting key features of the acquired signals by using a principal component analysis method, and constructing a feature sample set;

3) dividing the characteristic sample set into a training sample set and a testing sample set, and respectively using the training sample set and the testing sample set for training and evaluating the extreme learning machine;

4) dividing a training sample set into training data and testing data, training the extreme learning machine by using the training data, and simultaneously realizing the optimization analysis of the parameters of the extreme learning machine by means of an optimization algorithm;

5) and the test sample set is used for evaluating the optimized fault diagnosis performance of the extreme learning machine.

2. The IWO-ELM-based aviation power converter fault diagnosis method according to claim 1, wherein the parameters of the limit learning machine model in step 4) are optimized using a weed algorithm. The weed algorithm optimizes the number of hidden layer nodes, the input weight and the hidden layer node bias, so that the optimized extreme learning machine model has better fault diagnosis performance.