阅读说明：本技术 一种大型水轮发电机组故障诊断方法 (Fault diagnosis method for large hydroelectric generating set ) 是由 马建峰 潘骏 周叶 陈文华 缪熙熙 曹登峰 刘胜柱 成德明 张炜 邵保安 于 2021-01-20 设计创作，主要内容包括：本发明公开了一种大型水轮发电机组故障诊断方法,先采集n种类型的水轮发电机故障振动信号和水轮发电机正常运行的振动信号,信号数据分为训练集和测试集,每个(n+1)种振动状况需采集用于训练的100个数据点和用于测试的50个数据点；然后利用采样频率信息,使用快速谱峭度将水轮发电机振动时域信号转换为频域信号；用堆叠稀疏自编码器对频域信号进行分类；用粒子群优化算法对堆叠稀疏自编码器进行超参数选择,进而选择适用于水轮发电机组故障诊断的最优超参数；采用已训练合格的堆叠自编码网络对测试样本进行测试,以识别水轮发电机组的振动信号的故障类型。本发明无需大量故障样本数据即可对水轮机故障进行诊断,并且故障诊断精度高。(The invention discloses a fault diagnosis method for a large-scale water turbine generator set, which comprises the steps of firstly collecting fault vibration signals of n types of water turbine generators and vibration signals of normal operation of the water turbine generators, dividing signal data into a training set and a testing set, wherein each (n +1) vibration conditions need to collect 100 data points for training and 50 data points for testing; then, converting the vibration time domain signal of the hydraulic generator into a frequency domain signal by using the sampling frequency information and the rapid spectral kurtosis; classifying the frequency domain signals by using a stacked sparse self-encoder; carrying out super-parameter selection on the stacked sparse self-encoder by using a particle swarm optimization algorithm, and further selecting an optimal super-parameter suitable for fault diagnosis of the hydroelectric generating set; and testing the test sample by adopting the trained and qualified stacked self-coding network so as to identify the fault type of the vibration signal of the water-turbine generator set. The method can diagnose the fault of the water turbine without a large amount of fault sample data, and has high fault diagnosis precision.)

1. A fault diagnosis method for a large hydroelectric generating set is characterized by comprising the following steps:

a. acquiring fault vibration signals of n types of hydraulic generators and vibration signals of normal operation of the hydraulic generators, wherein signal data are divided into a training set and a testing set, and each (n +1) vibration conditions need to acquire 100 data points for training and 50 data points for testing;

b. converting the hydro-generator vibration time domain signal into a frequency domain signal by using the sampling frequency information and the rapid spectral kurtosis,

K_x(f)＝S₄(f)/(S₂(f))²-2 (equation 1)

S_n(f)＝E<|L(f,t)|ⁿ>(formula 2)

Wherein f is not equal to 0, S_n(f) Being the n-th moment of the signal, E<·>For averaging, | · | is modulo, L (f, t) is the complex envelope of the signal x (t) at f;

c. classifying the frequency domain signals by using a stacked sparse self-encoder;

d. the sparse automatic encoder applies constraint to the hidden unit, so as to activate the inactive hidden unit;

e. carrying out super-parameter selection on the stacked sparse self-encoder by using a particle swarm optimization algorithm, and further selecting an optimal super-parameter suitable for fault diagnosis of the hydroelectric generating set;

f. and testing the test sample by adopting the trained and qualified stacked self-coding network so as to identify the fault type of the vibration signal of the water-turbine generator set.

2. The method for fault diagnosis of a large hydroelectric generating set according to claim 1, wherein the stacked sparse self-encoder maps the input data into the hidden layer using formula 3, the hidden layer is reconstructed using formula 4,

equation 3 is:equation 4 is:

whereinB is a bias vector between the input layer and the hidden layer.

3. The method for diagnosing the fault of the large hydroelectric generating set according to claim 1, wherein in the step d, the reconstruction error of the sparse automatic encoder is as follows:

wherein the content of the first and second substances,representing the Kullberg-Leibler divergence, beta is the weight of the sparsity penalty term, p is the sparsity parameter,is the average amount of activation of the hidden nodes.

4. The fault diagnosis method for the large hydroelectric generating set according to claim 1, wherein the particle swarm optimization algorithm comprises three parameters of individual experience, overall experience and current movement of particles, and the position and the speed of each particle can be obtained through a formula 6 and a formula 7:

where d is the dimension of the kth iteration of the particle (1. ltoreq. d. ltoreq.n), v is the velocity of the ith particle in this range, w is the weight of the inertia, c₁And c₂Representing individual and global learning factors, respectively.

5. The method for diagnosing the fault of the large hydroelectric generating set according to claim 1, wherein the last layer of the stacked sparse self-encoder is a softmax classifier, and a softmax equation is defined as:

wherein: theta₁，θ₂，……，θ_k∈Rⁿ⁺¹Are the model parameters.

6. The method for diagnosing the fault of the large hydroelectric generating set according to claim 4, wherein the global exploration and the local exploration are balanced by the formula 9,

wherein: w is a_minIs the minimum inertia weight, w_maxAnd u is the current iteration for the maximum inertia weight.

Technical Field

The invention relates to the technical field of hydraulic generators, in particular to a fault diagnosis method for a large hydraulic generator set.

Background

With the enhancement of energy conservation and environmental protection awareness of people, hydroelectric power generation as a green energy source is being vigorously developed. The large hydropower station in China has the characteristics of high water head, high altitude, poor cavitation performance, strong mechanical vibration, complex arrangement, multiple sets of hydraulic units, long water conduit, huge water flow inertia, close hydraulic power coupling and the like, so that the running environment of the water-turbine generator set is severe day by day, more and more exciting factors causing the fault of the water-turbine generator set are generated, and a series of international academic frontier problems and engineering technical problems to be solved urgently are brought to the safe and stable running of the water-turbine generator set. In order to ensure safe and reliable operation of the hydroelectric generating set in the whole life cycle, accurate fault diagnosis is necessary to be carried out on the hydroelectric generating set, so that the dynamic performance and the operation efficiency of the hydroelectric generating set are improved, and disastrous accidents of a hydropower station are prevented.

In the prior art, theories of carrying out fault diagnosis around the hydroelectric generating set are formed, wherein expert system technology and neural network technology are hot spots of research and application. The fault diagnosis method based on the expert system utilizes rich practical experience accumulated by experts, can explain reasoning process by simulating ideas of expert analysis and problem solving, and becomes a fault diagnosis method. However, the expert system has poor knowledge acquisition ability, fault tolerance ability and self-learning ability, and is not suitable for application and success in water turbine fault diagnosis. The appearance of the neural network provides a new solution for diagnosing faults of the hydroelectric generating set. The fault diagnosis system involves five main steps: data/signal acquisition, data/signal pre-processing, feature extraction, feature reduction/selection, and fault diagnosis. In recent years, with the continuous development of technology, the fault diagnosis system has developed a deep learning model, which simplifies the fault diagnosis process into three main steps: signal acquisition, signal preprocessing and fault diagnosis. Because the deep learning model has multiple hidden layers, the feature extraction and selection task is automated. However, deep learning models require large data sets to enable the learning process to be effective to produce accurate fault diagnosis results. The hydroelectric generating set has the characteristics of less abnormal sample data and no calibration of faults, so that the deep learning model is easy to over-fit.

Disclosure of Invention

The invention aims to provide a fault diagnosis method for a large hydroelectric generating set, which does not need a large amount of fault sample data and has high fault diagnosis precision.

In order to achieve the purpose, the invention adopts the following technical scheme:

a fault diagnosis method for a large hydroelectric generating set comprises the following steps:

a. acquiring fault vibration signals of n types of hydraulic generators and vibration signals of normal operation of the hydraulic generators, wherein signal data are divided into a training set and a testing set, and each (n +1) vibration conditions need to acquire 100 data points for training and 50 data points for testing;

b. converting the hydro-generator vibration time domain signal into a frequency domain signal by using the sampling frequency information and the rapid spectral kurtosis,

K_x(f)＝S₄(f)/(S₂(f))²-2 (equation 1)

S_n(f)＝E〈|L(f,t)|ⁿEquation 2

Wherein f is not equal to 0, S_n(f) Taking the n-order spectral moment of the signal, E < is the average value, | is the modulus, and L (f, t) is the complex envelope of the signal x (t) at f;

c. classifying the frequency domain signals by using a stacked sparse self-encoder;

d. the sparse automatic encoder applies constraint to the hidden unit, so as to activate the inactive hidden unit;

e. carrying out super-parameter selection on the stacked sparse self-encoder by using a particle swarm optimization algorithm, and further selecting an optimal super-parameter suitable for fault diagnosis of the hydroelectric generating set;

f. and testing the test sample by adopting the trained and qualified stacked self-coding network so as to identify the fault type of the vibration signal of the water-turbine generator set.

Preferably, the stacked sparse self-encoder maps the input data into the hidden layer using equation 3, the hidden layer is reconstructed by equation 4,

equation 3 is:equation 4 is:

whereinB is a bias vector between the input layer and the hidden layer.

Preferably, in step d, the reconstruction error of the sparse automatic encoder is:

wherein the content of the first and second substances,representing the Kullberg-Leibler divergence, beta is the weight of the sparsity penalty term, p is the sparsity parameter,is the average amount of activation of the hidden nodes.

Preferably, the particle swarm optimization algorithm comprises three parameters of individual experience, overall experience and current motion of the particles, and the position and the speed of each particle can be obtained through a formula 6 and a formula 7:

where d is the dimension of the kth iteration of the particle (1. ltoreq. d. ltoreq.n), v is the velocity of the ith particle in this range, w is the weight of the inertia, c₁And c₂Representing individual and global learning factors, respectively.

Preferably, the last layer of the stacked sparse autoencoder is a softmax classifier, and the softmax equation is defined as:

wherein: theta₁，θ₂，……，θ_k∈Rⁿ⁺¹Are the model parameters.

Preferably, the ability to survey globally and locally is balanced by equation 9,

wherein: w is a_minIs the minimum inertia weight, w_maxAnd u is the current iteration for the maximum inertia weight.

Therefore, the invention has the following beneficial effects: the fault of the hydraulic generator can be accurately diagnosed and identified by adopting limited hydraulic generator fault sample data; the invention does not need a large amount of fault sample data and has high fault diagnosis precision.

Detailed Description

A fault diagnosis method for a large hydroelectric generating set comprises the following steps:

a. acquiring fault vibration signals of n types of hydraulic generators and vibration signals of normal operation of the hydraulic generators, wherein signal data are divided into a training set and a testing set, and each (n +1) vibration conditions need to acquire 100 data points for training and 50 data points for testing;

b. converting the hydro-generator vibration time domain signal into a frequency domain signal by using the sampling frequency information and the rapid spectral kurtosis,

K_x(f)＝S₄(f)/(S₂(f))²-2 (equation 1)

S_n(f)＝E<|L(f,t)|ⁿ>(formula 2)

Wherein f is not equal to 0, S_n(f) Being the n-th moment of the signal, E<·>For averaging, | · | is modulo, L (f, t) is the complex envelope of the signal x (t) at f;

c. classifying the frequency domain signals by using a stacked sparse self-encoder;

d. the sparse automatic encoder applies constraint to the hidden unit, so as to activate the inactive hidden unit;

e. carrying out super-parameter selection on the stacked sparse self-encoder by using a particle swarm optimization algorithm, and further selecting an optimal super-parameter suitable for fault diagnosis of the hydroelectric generating set;

f. and testing the test sample by adopting the trained and qualified stacked self-coding network so as to identify the fault type of the vibration signal of the water-turbine generator set.

The invention introduces a stacked sparse self-encoder to classify frequency domain signals. Firstly, the parameter setting of the stacking sparse self-encoder is carried out, and the stacking sparse self-encoder is formed by overlapping a plurality of sparse self-encoders. A sparse autoencoder includes encoder, hidden layer, decoder functions.

Preferably, the stacked sparse self-encoder maps the input data into the hidden layer using equation 3, the hidden layer is reconstructed by equation 4, equation 3 is:equation 4 is:whereinB is a bias vector between the input layer and the hidden layer.

Sparse autoencoders impose constraints on the hidden units of the autoencoder, thereby activating inactive hidden units.

Further, in step d, the reconstruction error of the sparse automatic encoder is:

wherein the content of the first and second substances,representing the Kullberg-Leibler divergence, beta is the weight of the sparsity penalty term, p is the sparsity parameter,is the average amount of activation of the hidden nodes.

Further, the particle swarm optimization algorithm comprises three parameters of individual experience, overall experience and current movement of the particles, and the position and the speed of each particle can be obtained through formula 6 and formula 7:

where d is the dimension of the kth iteration of the particle (1. ltoreq. d. ltoreq.n), v is the velocity of the ith particle in this range, w is the weight of the inertia, c₁And c₂Representing individual and global learning factors, respectively.

The method is used for carrying out super-parameter selection on the stacked sparse self-encoder based on the particle swarm optimization algorithm. Since limited data samples may cause the network to easily over-fit the training data, an important hyper-parameter of the stacked sparse auto-encoder is the adjuster, and since the adjuster can reduce the over-fitting problem, the invention optimally adjusts the over-fitting. Specifically, the sparsity parameter, the average activation amount of the hidden nodes and the weight of the sparsity penalty term are selected for simultaneous optimization.

Preferably, the last layer of the stacked sparse autoencoder is a softmax classifier, and the softmax equation is defined as:

wherein: theta₁，θ₂，……，θ_k∈Rⁿ⁺¹Are the model parameters.

Further, the ability to survey globally and locally is balanced by equation 9,

wherein: w is a_minIs the minimum inertia weight, w_maxAnd u is the current iteration for the maximum inertia weight.