CN113850181A - Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion - Google Patents

Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion Download PDF

Info

Publication number
CN113850181A
CN113850181A CN202111112808.5A CN202111112808A CN113850181A CN 113850181 A CN113850181 A CN 113850181A CN 202111112808 A CN202111112808 A CN 202111112808A CN 113850181 A CN113850181 A CN 113850181A
Authority
CN
China
Prior art keywords
fault
formula
entropy
feature
guide vane
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111112808.5A
Other languages
Chinese (zh)
Inventor
张文广
陆瑶
徐浩博
陈松
牛玉广
王玮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
North China Electric Power University
Original Assignee
North China Electric Power University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by North China Electric Power University filed Critical North China Electric Power University
Priority to CN202111112808.5A priority Critical patent/CN113850181A/en
Publication of CN113850181A publication Critical patent/CN113850181A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F01—MACHINES OR ENGINES IN GENERAL; ENGINE PLANTS IN GENERAL; STEAM ENGINES
    • F01D—NON-POSITIVE DISPLACEMENT MACHINES OR ENGINES, e.g. STEAM TURBINES
    • F01D9/00—Stators
    • F01D9/02—Nozzles; Nozzle boxes; Stator blades; Guide conduits, e.g. individual nozzles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Signal Processing (AREA)
  • Mechanical Engineering (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a fault diagnosis method for a gas turbine inlet guide vane system based on characteristic information fusion; the method comprises the following steps: collecting an original vibration signal; analyzing a fault mechanism; optimizing and decomposing parameters of Variational Modal Decomposition (VMD); extracting fault characteristics; normalizing the state feature vector; encoding a feature vector; impulse neural network (SNN) fault diagnosis. The method adopts a dolphin swarm algorithm to optimize the VMD parameters, and improves the decomposition accuracy; screening IMF components sensitive to fault information by taking kurtosis-mutual information entropy as a basis, and eliminating fault characteristic sensitive mode functions with poor distribution rules and few impact components; fault feature extraction is carried out in a time-frequency domain by adopting a multi-feature entropy algorithm, so that the condition that single feature cannot comprehensively reflect fault feature information is avoided, and a guarantee is provided for accurate diagnosis of faults; the SpikeProp algorithm is adopted to optimize the SNN, and the method has the capability of solving the nonlinear classification problem and enables the training result to be more accurate.

Description

Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion
Technical Field
The invention belongs to the technical field of fault diagnosis of gas turbines, and particularly relates to a fault diagnosis method for a gas turbine inlet guide vane system based on feature information fusion.
Background
The gas turbine is an internal combustion type power machine which takes continuously flowing natural gas as a working medium to drive an impeller to rotate at a high speed, and is widely applied to the fields of aerospace, chemical engineering and the like. A heavy-duty gas turbine used for industrial power generation is an important power system for realizing energy conservation and environmental protection in China. The inlet guide vane is a series of static vanes in front of a first-stage movable vane of the gas turbine compressor, is used as a core component of the gas turbine compressor and plays a role in energy conversion. The angle and the flow of the air flow entering the air compressor are controlled by changing the angle of the guide vane, once the inlet guide vane system fails, the normal operation of the gas turbine is seriously damaged, and even more, the safety accident can be caused.
When the inlet guide vane system fails, the vibration signals received by the sensor become significant non-stationary signals, and a large amount of characteristic information is hidden in the signals. Most of the existing signal analysis methods are Empirical Mode Decomposition (EMD), Local Mean Decomposition (LMD) and the like, and the EMD is easy to have the problem of mode aliasing; LMD is prone to loss of signal singularity. The fault diagnosis method based on feature information fusion adopts Variational Modal Decomposition (VMD) to carry out self-adaptive processing on signals, overcomes the problems of modal aliasing, signal characteristic loss, boundary effect and the like, and has more superiority.
In order to solve the problem that a tiny fault cannot be described, the invention provides a fault diagnosis method based on feature information fusion to analyze multi-scale and multi-features on the basis of introducing an entropy concept, performs feature extraction through four entropy functions, and performs fault diagnosis by adopting a pulse neural network (SNN).
Disclosure of Invention
Aiming at the problems in the background art, the invention provides a fault diagnosis method for a gas turbine inlet guide vane system based on feature information fusion, which is characterized by comprising the following steps of:
step 1, collecting vibration signal data of an inlet guide vane system of a gas turbine and analyzing a fault mechanism of the data to obtain a variation trend of a vibration signal amplitude at a fault moment;
step 2, optimizing parameters of variational modal decomposition by using a group intelligent algorithm;
step 3, carrying out variation mode decomposition on the vibration signal by using the optimized parameters to obtain k intrinsic mode function components, and screening intrinsic mode function components sensitive to fault information by taking kurtosis-mutual information entropy as a basis;
step 4, extracting fault characteristics in a time-frequency domain by adopting a multi-characteristic entropy algorithm, constructing a state characteristic vector, and carrying out normalization processing on the state characteristic vector;
step 5, encoding the normalized feature vectors into a pulse time sequence to obtain a training set;
step 6, inputting the training set into a pulse neural network model for training, constructing the pulse neural network model, and then optimizing the pulse neural network by adopting a SpikeProp algorithm;
step 7, sampling inlet guide vane system fault data and repeating the steps 2 to 5 to obtain a pulse signal which can be identified by the pulse neural network model and used as a test set; and inputting the voltage to the trained pulse neural network model to obtain the membrane voltage of the output neuron, and obtaining the fault category of the vibration signal to be diagnosed when the membrane voltage of a certain neuron reaches a threshold value to send out a pulse and other neurons do not send out the pulse when the membrane voltage of the certain neuron does not reach the threshold value.
The variation trend of the amplitude of the vibration signal at the fault moment in the step 1 is as follows: gas turbineA vibration signal generated by the inlet guide vane system is transmitted to the surface of the casing from the blade root and then transmitted to the vibration sensor; wherein the amplitude of the vibration signal is suddenly changed at the fault moment, and the amplitude r of the vibration signalaComprises the following steps:
Figure BDA0003274367430000021
in the formula, ωcNatural frequency, Hz; ζ is the damping ratio; omega is the rotating speed, r/s;
Figure BDA0003274367430000022
rad, the vibration phase; l is the distance between centers of mass before and after the guide vane is broken, cm; t is time in units of s.
In the step 2, the dolphin swarm algorithm in the swarm intelligence algorithm is used for optimizing the parameters of the variation modal decomposition, and the step 2 comprises the following steps:
setting a parameter range of a modal decomposition number k and a penalty factor alpha, and initializing a population; decomposing the signal through the VMD to obtain a plurality of intrinsic mode function components, and selecting the minimum envelope entropy as a fitness function; then entering a searching stage to find out the optimal fitness value of the individual; entering a calling and receiving stage, and finding out a neighborhood optimal solution; in the predation stage, updating the position of the individual to the optimal solution of the neighborhood, and calculating the optimal fitness value; when the requirement of iteration times is met, outputting the optimal solution of k and alpha, wherein the fitness function FfitThe formula is as follows:
Ffit＝min(Een) (2)
in the formula, EenRepresenting the sparsity of the component signals for the envelope entropy;
Eenthe calculation formula is as follows:
Figure BDA0003274367430000031
wherein τ(s) is an envelope signal of the vibration signal x (t) after hilbert demodulation; e.g. of the typesNormalized value of τ(s); s is 1, 2.
In the step 3, the vibration signal is subjected to variational modal decomposition by using the optimized parameters, the variational modal decomposition decomposes the vibration signal into k intrinsic modal function components, the decomposition sequence is guaranteed to be a modal component with limited bandwidth of intermediate frequency, the sum of the estimated bandwidth of each mode is minimum, the constraint condition is that the sum of all the modes is equal to the original signal, the variational modal decomposition realizes frequency domain division of the signal, further effective decomposition components of the vibration signal are obtained, and finally the optimal solution of the variational problem is obtained;
the vibration signal is decomposed through a variational mode to obtain k intrinsic mode function components, and a constraint variational model formula is established for the vibration signal as follows:
Figure BDA0003274367430000032
in the formula uk(t) is the modal component; omegakIs the intermediate frequency of each modal component; k is the number of IMFs; δ (t) is the shock function; j is an imaginary unit; t is time in units of s;
Figure BDA0003274367430000033
is a harmonic signal; is a convolution calculation; s.t. is a constraint condition;
in order to solve the optimal solution of the constraint problem of the variational model in the formula (4), a quadratic penalty function and a Lagrange multiplier method are utilized to convert the solution into an unconstrained variational model shown in the formula (5):
Figure BDA0003274367430000041
wherein λ (t) is a Lagrangian multiplier;
Figure BDA0003274367430000042
represents a 2-norm;<·>representing the vector inner product;
to solve the unconstrained variational model, lags are sought by the alternating direction multiplier methodSaddle point of the Langerian expression until the iteration stop condition is satisfied
Figure BDA0003274367430000043
Obtaining an optimal solution for the variational model; where ε is the precision.
In the step 3, intrinsic mode function components sensitive to fault information are screened by using kurtosis-mutual information entropy, and the formula is as follows:
Figure BDA0003274367430000044
Emi＝Hmi(x1)+Hmi(X(t))-Hmi(x1,X(t)) (7)
in the formula, KcKurtosis, indicating the severity of the fault;
Figure BDA0003274367430000045
is the average amplitude of signal X (t); k is the number of IMFs; emiThe entropy is mutual information entropy and represents the relevance between two signals; x is the number of1Is any one IMF component; hmi(. cndot.) is the entropy of the signal.
The step 4 comprises the following steps:
step 4.1, the IMF component after the decomposition of the variation mode decomposition is { uk(t)}＝{u1(t),u2(t),...uk(t) }, calculating instantaneous energy entropy Henergy：
Figure BDA0003274367430000046
Figure BDA0003274367430000051
In the formula riIs the percentage of the component energy of the ith intrinsic mode function to the total energy; i ═ {1,2, …, k } is the number of components of the sensitive eigenmode function;
step 4.2 at signalIn the inherent characteristic, singular values have excellent stability, and the energy change of the signal in the sampling time can be expressed by singular value entropy. Forming eigenmode function components into a feature vector matrix u1(t),u2(t),...uk(t)]ΤCalculating singular values V of the feature matrixsingularThereby obtaining the singular entropy, and the formula is:
Figure BDA0003274367430000052
Figure BDA0003274367430000053
in the formula
Figure BDA0003274367430000054
The percentage of singular values of the ith sensitive eigenmode function component in the total singular values is;
step 4.3, expressing the change of signal energy in a frequency domain by the power spectrum entropy, carrying out Fourier transform on the intrinsic mode function component to obtain a power spectrum value chi, and obtaining the power spectrum entropy H according to the energy conservation law of signal time-frequency domain transformpowerThe formula is as follows:
Figure BDA0003274367430000055
Figure BDA0003274367430000056
Figure BDA0003274367430000057
in the formula piThe weight of the power spectrum of the i sensitive intrinsic mode function components in the total amount of the power spectrum is calculated;
step 4.4, the intrinsic mode function component uk(t) Performing a phase space reconstruction of uk(t)' is the eigenmode function component after reconstruction, and a matrix is obtained:
Figure BDA0003274367430000061
Figure BDA0003274367430000062
where m is the embedding dimension; upsilon is a time delay; g is the number of vectors in the reconstructed phase space; a isiThe probability of occurrence of the i reconstructed eigenmode function components;
step 4.5, in order to enable different feature vectors y to be in the same interval before encoding, the feature vector composed of four entropy values is subjected to maximum and minimum normalization processing, so that the feature vector is normalized to be between [0 and 1], and the formula is as follows:
Figure BDA0003274367430000063
in the formula xp *Normalizing the value of the p characteristic; x is the number ofpIs the p-th feature; x is the number ofminIs the minimum of the p-th feature; x is the number ofmaxIs the maximum of the p-th feature.
In step 5, the normalized feature vector features are converted into a pulse time sequence which can be identified by the input neurons by a group coding method, the group coding adopts gamma overlapped Gaussian receptive field coding neurons to map the single features to a high-dimensional feature space, and the activation value formula of the coding neurons is as follows:
Figure BDA0003274367430000064
Figure BDA0003274367430000065
in the formula, AγThe activation value of the gamma encoding neuron under the corresponding characteristic value is obtained; mu.sγThe central value of the Gauss receptive field of the gamma-th coding neuron; sigmahTo encode the width of the neuronal gaussian receptive field.
In step 6, the number of neurons in the input layer of the impulse neural network is the product of the state class number and the feature vector dimension, and the number of neurons in the output layer is the fault class number; synaptic weight initialization between neurons of the input and output layers obeys a uniform distribution between 0-1; the neuron model of the impulse neural network is an impulse response model, and the formula is as follows:
V(t)＝∑fωf∑tfK(t-tf)+Vrest (20)
wherein V (t) is neuron mode voltage in mv; f is the number of synapses; omegafWeight of the f-th synapse; t is tfTime of arrival of the f-th synaptic input pulse in units of s; vrestIs a resting potential; k (-) is the kernel function of the impulse response model; k (t-t)f) A kernel function of the arrival time of the f-th synaptic input pulse;
the kernel function K (t) of the time-dependent behavior of the postsynaptic potential is:
Figure BDA0003274367430000071
in the formula: tau ismAnd τsIs the decay time constant; vnorIs a normalized value;
in the training stage, the spikeProp algorithm is adopted by the impulse neural network to adjust the synaptic connection weight value through the error function feedback, and the error function formula is as follows:
Figure BDA0003274367430000072
in the formula (I), the compound is shown in the specification,
Figure BDA0003274367430000073
the instant at which neuron j actually fires a pulse,
Figure BDA0003274367430000074
the time at which a pulse is expected for neuron j;
optimizing an error function according to a gradient descent method, and updating a synapse weight value based on a chain-type derivation method:
Figure BDA0003274367430000075
in the formula (I), the compound is shown in the specification,
Figure BDA0003274367430000076
is as follows
Figure BDA0003274367430000077
Updating a synaptic weight between the input neuron and the output neuron; eta is the learning rate;
Figure BDA0003274367430000078
is the first to be updated
Figure BDA0003274367430000079
Synaptic weights between the input neurons and the output neurons.
The invention has the beneficial effects that:
1. optimizing the VMD parameters through a dolphin swarm algorithm in a swarm intelligent algorithm, and adaptively selecting the VMD parameters to improve the decomposition accuracy; and screening IMF components sensitive to fault information by taking kurtosis-mutual information entropy as a basis, and removing fault characteristic sensitive mode functions with poor distribution rules and few impact components.
2. The fault characteristics are extracted in a time-frequency domain by adopting a multi-characteristic entropy algorithm, so that the condition that single characteristics cannot comprehensively reflect fault characteristic information is avoided, and the accurate diagnosis of the fault is guaranteed.
3. And the SNN is optimized by combining an SRM neuron model and adopting a SpikeProp algorithm, so that the method has the capability of solving the nonlinear classification problem and ensures that the training result is more accurate.
Drawings
FIG. 1 is a flowchart of an embodiment of a method for diagnosing a fault of a gas turbine inlet guide vane system based on feature information fusion according to the present invention.
FIG. 2 is a flow chart of optimizing VMD parameters by the dolphin swarm algorithm provided by the present invention.
Fig. 3 is an exploded view of the optimized VMD versus vibration signal provided by the present invention.
FIG. 4 is a diagram of the SNN model diagnostic results provided by the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
The embodiment of the present invention shown in fig. 1 comprises the following steps:
step 1, collecting vibration signal data of an inlet guide vane system of a gas turbine and analyzing a fault mechanism of the data to obtain a variation trend of a vibration signal amplitude at a fault moment;
step 2, optimizing parameters of Variational Modal Decomposition (VMD) by using a group intelligent algorithm;
step 3, VMD decomposition is carried out on the vibration signals by using the optimized parameters to obtain k Intrinsic Mode Function (IMF) components, and IMF components sensitive to fault information are screened by taking kurtosis-mutual information entropy as a basis;
step 4, extracting fault characteristics in a time-frequency domain by adopting a multi-characteristic entropy algorithm, constructing a state characteristic vector, and carrying out normalization processing on the state characteristic vector;
step 5, encoding the normalized feature vectors into a pulse time sequence to obtain a training set;
step 6, inputting the training set into an SNN model for training, constructing the SNN model, and then optimizing the SNN by adopting a SpikeProp algorithm;
and 7, sampling fault data of the inlet guide vane system, repeating the steps 2 to 5 to obtain a pulse signal which can be identified by the SNN model and is used as a test set, inputting the pulse signal into the trained SNN model to obtain the membrane voltage of an output neuron, and obtaining the fault category of the vibration signal to be diagnosed when the membrane voltage of a certain neuron reaches a threshold value and the membrane voltage of other neurons do not reach the threshold value and do not send the pulse.
The variation trend of the amplitude of the vibration signal at the fault moment in the step 1 is as follows: vibration signals generated by the gas turbine inlet guide vane system are transmitted to the surface of the casing through the blade root and then transmitted to the vibration sensor. The amplitude of the vibration signal changes suddenly at the fault moment, and the amplitude r of the vibration signalaComprises the following steps:
Figure BDA0003274367430000091
in the formula, ωcNatural frequency, Hz; ζ is the damping ratio; omega is the rotating speed, r/s;
Figure BDA0003274367430000092
rad, the vibration phase; l is the distance between centers of mass before and after the guide vane is broken, cm; t is time in units of s.
In step 2, parameters of the VMD are optimized by using a dolphin swarm optimization (DSA) in a swarm intelligence algorithm, the swarm intelligence algorithm has the characteristics of simplicity, good robustness and wide applicability, biological characteristics of dolphin echo positioning, information exchange, division work cooperation and the like are simulated by the DSA algorithm, and optimization of the algorithm is realized through four key stages of searching, calling, receiving and predating. Other group intelligent algorithms only adopt an advanced solution, while the DSA algorithm utilizes echo positioning and adopts different strategies, so that the optimal solution is easier to obtain. FIG. 2 is a flow chart of optimizing VMD parameters by the dolphin swarm algorithm provided by the present invention, which is divided into:
firstly, setting a parameter range of a modal decomposition number k and a penalty factor alpha, specifically, setting a value range of k to be 2-12 and a value range of alpha to be 100-5000; initializing the population; decomposing the signal through the VMD to obtain a plurality of IMF components, and selecting the minimum envelope entropy as a fitness function; then entering a searching stage to find out the optimal fitness value of the individual; entering a calling and receiving stage, and finding out a neighborhood optimal solution; in the predation stage, the position of the individual is updated to the optimal solution of the neighborhood, and the optimal fitness value is calculated; and when the requirement of the iteration number is met, outputting the optimal solution of k and alpha.
According to the method, the optimal parameters k is 6 and alpha is 2000 are obtained through parameter optimization, so that the VMD decomposition accuracy is effectively improved, and the fault diagnosis model is established more accurately.
Fitness function FfitThe formula is as follows:
Ffit＝min(Een) (2)
in the formula, EenRepresenting the sparsity of the component signals for the envelope entropy;
Eenthe calculation formula is as follows:
Figure BDA0003274367430000101
wherein τ(s) is an envelope signal of the vibration signal x (t) after hilbert demodulation; e.g. of the typesNormalized value of τ(s); s is 1, 2.
And 3, VMD decomposition is carried out on the vibration signal by adopting the optimized parameters, the VMD decomposition decomposes the vibration signal into k IMF components, the decomposition sequence is ensured to be a mode component with limited bandwidth of intermediate frequency, the sum of the estimated bandwidths of all the modes is minimum, and the constraint condition is that the sum of all the modes is equal to the original signal. And the VMD decomposition realizes the frequency domain division of the signals, so that effective decomposition components of the vibration signals are obtained, and finally, the optimal solution of the variation problem is obtained. The result is shown in fig. 3, the vibration signal is decomposed by VMD to obtain 6 eigenmode function components, and the formula of the constraint variational model established for the vibration signal is:
Figure BDA0003274367430000102
in the formula uk(t) is the modal component; omegakIs the intermediate frequency of each modal component; k is the number of IMFs; δ (t) is the shock function; j is an imaginary unit; t is time in units of s;
Figure BDA0003274367430000103
is a harmonic signal; is a convolution calculation; s.t. is a constraint.
In order to solve the optimal solution of the constraint problem of the variational model in the formula (4), a quadratic penalty function and a Lagrange multiplier method are utilized to convert the solution into an unconstrained variational model shown in the formula (5):
Figure BDA0003274367430000111
wherein λ (t) is a Lagrangian multiplier;
Figure BDA0003274367430000112
represents a 2-norm;<·>representing the vector inner product.
In order to solve the unconstrained variational model, saddle points of Lagrange expressions are sought through an alternative direction multiplier method until an iteration stop condition is met
Figure BDA0003274367430000113
I.e. obtaining the optimal solution for the variational model. Where ε is the precision, which is set to 1e-7 during the iteration.
In the step 3, an IMF component sensitive to fault information is screened by using kurtosis-mutual information entropy, and the formula is as follows:
Figure BDA0003274367430000114
Emi＝Hmi(x1)+Hmi(X(t))-Hmi(x1,X(t)) (7)
in the formula, KcKurtosis, indicating the severity of the fault;
Figure BDA0003274367430000115
is the average amplitude of signal X (t); k is the number of IMFs; emiThe entropy is mutual information entropy and represents the relevance between two signals; x is the number of1Is any one IMF component; hmi(. cndot.) is the entropy of the signal.
In step 4, in order to avoid that the single entropy value can not reflect the fault characteristics completely, a multi-characteristic entropy value algorithm is adoptedCalculating instantaneous energy entropy HenergySingular value entropy HsingularPower spectrum entropy HpowerArray entropy HarrangeAnd extracting the fault characteristics of the time-frequency domain in a multi-scale mode by using the calculated characteristic vector consisting of the four entropy values, wherein the characteristic vector is specifically divided into the following steps:
and 4.1, expressing the change of the energy contained in the vibration signal by the energy entropy, reflecting the change of the amplitude of the signal in a time domain, and revealing the inherent characteristics of the signal. Instantaneous energy entropy HenergyThe energy entropy is obtained by combining the information entropy, the frequency band integral energy is adopted to reflect the integral energy distribution of the mechanical fault signal, and the energy distribution of different mechanical faults is different, so the instantaneous energy entropy is also different; in step 3, the VMD decomposed IMF component (modal component) is { u }k(t)}＝{u1(t),u2(t),...uk(t) }, calculating instantaneous energy entropy Henergy：
Figure BDA0003274367430000121
Figure BDA0003274367430000122
In the formula riIs the percentage of the ith IMF component energy to the total energy; and i is the number of sensitive IMF components {1,2, …, k }.
And 4.2, in the inherent characteristics of the signals, singular values have excellent stability, and the energy change of the signals in the sampling time can be expressed by singular value entropy. Forming IMF components into a feature vector matrix u1(t),u2(t),...uk(t)]ΤCalculating singular values V of the feature matrixsingularThereby obtaining the singular entropy, and the formula is:
Figure BDA0003274367430000123
Figure BDA0003274367430000124
in the formula
Figure BDA0003274367430000125
The singular values of the ith sensitive IMF component are a percentage of the total number of singular values.
And 4.3, expressing the change of the signal energy in the frequency domain by the power spectrum entropy, carrying out Fourier transform on the IMF component to obtain a power spectrum value chi, and obtaining the power spectrum entropy H according to the energy conservation law of the signal time-frequency domain transformpowerThe formula is as follows:
Figure BDA0003274367430000126
Figure BDA0003274367430000127
Figure BDA0003274367430000131
in the formula piThe power spectrum of the i sensitive IMF components is weighted according to the total amount of the power spectrum.
Step 4.4, permutation entropy HarrangeThe method is an effective method for depicting the complexity and the randomness of the time sequence, and has the characteristics of simple calculation and excellent noise robustness. The IMF component uk(t) performing a phase space reconstruction, uk(t)' is the reconstructed IMF component, resulting in a matrix:
Figure BDA0003274367430000132
Figure BDA0003274367430000133
where m is the embedding dimension; v is the time delay(ii) a g is the number of vectors in the reconstructed phase space; a isiThe probability of occurrence for the i reconstructed IMF components.
Step 4.5, in order to enable different feature vectors y to be in the same interval before encoding, the feature vector composed of four entropy values is subjected to maximum and minimum normalization processing, so that the feature vector is normalized to be between [0 and 1], and the formula is as follows:
Figure BDA0003274367430000134
in the formula xp *Normalizing the value of the p characteristic; x is the number ofpIs the p-th feature; x is the number ofminIs the minimum of the p-th feature; x is the number ofmaxIs the maximum of the p-th feature.
In the step 5, the normalized feature vector features are converted into a pulse time sequence which can be identified by the input neurons by a group coding method, the group coding adopts gamma overlapped gaussian receptive field coding neurons to map the single features to a high-dimensional feature space, the gamma range is 1 to 8, and the activation value formula of the coding neurons is as follows:
Figure BDA0003274367430000141
Figure BDA0003274367430000142
in the formula, AγThe activation value of the gamma encoding neuron under the corresponding characteristic value is obtained; mu.sγThe central value of the Gauss receptive field of the gamma-th coding neuron; sigmahTo encode the width of the neuronal gaussian receptive field.
In the step 6, the number of the SNN input layer neurons is the product of the state type number (including normal and fault) and the feature vector dimension, and the number of the output layer neurons is the fault type number; synaptic weights between neurons of the input layer and the output layer initialize a uniform distribution between 0-1; the SNN neuron model is an impulse response model (SRM) and has the formula:
V(t)＝Σfωf∑tfK(t-tf)+Vrest (20)
wherein V (t) is neuron mode voltage in mv; f is the number of synapses; omegafWeight of the f-th synapse; t is tfTime of arrival of the f-th synaptic input pulse in units of s; vrestIs a resting potential; k (-) is the kernel function of the impulse response model; k (t-t)f) A kernel function of the arrival time of the f-th synaptic input pulse;
the kernel function K (t) of the time-dependent behavior of the postsynaptic potential is:
Figure BDA0003274367430000143
in the formula: tau ismAnd τsIs the decay time constant; vnorIs a normalized value.
In the training stage, the SNN adopts a supervised SpikeProp algorithm, and adjusts the synaptic connection weight value through the feedback of an error function, wherein the error function formula is as follows:
Figure BDA0003274367430000144
in the formula (I), the compound is shown in the specification,
Figure BDA0003274367430000145
the instant at which neuron j actually fires a pulse,
Figure BDA0003274367430000146
the time at which a pulse is expected for neuron j.
Optimizing an error function according to a gradient descent method, and updating a synapse weight value based on a chain-type derivation method:
Figure BDA0003274367430000151
in the formula (I), the compound is shown in the specification,
Figure BDA0003274367430000152
is as follows
Figure BDA0003274367430000153
Updating a synaptic weight between the input neuron and the output neuron; eta is the learning rate, and the learning rate is set to be 0.001 in the training process;
Figure BDA0003274367430000154
is the first to be updated
Figure BDA0003274367430000155
Synaptic weights between the input neurons and the output neurons.
In the step 7, 100s of inlet guide vane system fault data are taken under the condition that the sampling frequency is 100Hz, the inlet guide vane system vibration signals to be diagnosed are input into the trained SNN model after being processed according to the steps 2 to 5, the input neurons continuously send pulse sequences to the output neurons, the output neuron membrane voltage is stimulated to rise, when certain neuron membrane voltage of an output layer reaches a threshold value to send pulses, and other neurons do not reach the threshold value to send pulses, the fault category of the vibration signals to be diagnosed is obtained, wherein the neuron membrane voltage threshold value is set to be 1.
As shown in fig. 4, a sensor is used for collecting vibration signals of a certain gas turbine inlet guide vane system under normal and 3 sudden-change fault conditions, 100s data is sampled under the condition that the frequency is 100Hz, fault mechanism analysis is carried out, the change trend of the vibration signal amplitude is obtained, the vibration signals are converted into pulse signals which can be identified by an SNN model through steps 2 to 5, wherein the output neuron membrane voltage represented by a fault F2 receives pulse stimulation transmitted by an input neuron within 100s and reaches a threshold voltage about 78s, the rest 3 output neurons are regulated by a spikeprep algorithm, the weight of synapses connected with the input neuron is restrained, the threshold voltage cannot be reached, and the fault category is judged.
The above description is only for the specific embodiments of the present invention, and the protection scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention disclosed herein should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (8)

1. A fault diagnosis method for a gas turbine inlet guide vane system based on feature information fusion is characterized by comprising the following steps:
step 1, collecting vibration signal data of an inlet guide vane system of a gas turbine and analyzing a fault mechanism of the data to obtain a variation trend of a vibration signal amplitude at a fault moment;
step 2, optimizing parameters of variational modal decomposition by using a group intelligent algorithm;
step 3, carrying out variation mode decomposition on the vibration signal by using the optimized parameters to obtain k intrinsic mode function components, and screening the intrinsic mode function components sensitive to fault information by taking kurtosis-mutual information entropy as a basis;
step 4, extracting fault characteristics in a time-frequency domain by adopting a multi-characteristic entropy algorithm, constructing a state characteristic vector, and carrying out normalization processing on the state characteristic vector;
step 5, encoding the normalized feature vectors into a pulse time sequence to obtain a training set;
step 6, inputting the training set into a pulse neural network model for training, constructing the pulse neural network model, and then optimizing the pulse neural network by adopting a SpikeProp algorithm;
step 7, sampling inlet guide vane system fault data and repeating the steps 2 to 5 to obtain a pulse signal which can be identified by the pulse neural network model and used as a test set; and inputting the voltage to the trained pulse neural network model to obtain the membrane voltage of the output neuron, and obtaining the fault category of the vibration signal to be diagnosed when the membrane voltage of a certain neuron reaches a threshold value to send out a pulse and other neurons do not send out pulses when the membrane voltage of the certain neuron does not reach the threshold value.
2. The feature-based message as in claim 1The fault diagnosis method for the gas turbine inlet guide vane system is characterized in that the variation trend of the amplitude of the vibration signal at the fault moment in the step 1 is as follows: the vibration signal generated by the gas turbine inlet guide vane system is transmitted to the surface of the casing through the blade root and then transmitted to the vibration sensor; wherein the amplitude of the vibration signal is suddenly changed at the fault moment, and the amplitude r of the vibration signalaComprises the following steps:
Figure FDA0003274367420000011
in the formula, ωcNatural frequency, Hz; ζ is the damping ratio; omega is the rotating speed, r/s;
Figure FDA0003274367420000012
rad, the vibration phase; l is the distance between centers of mass before and after the guide vane is broken, cm; t is time in units of s.
3. The method for diagnosing the fault of the gas turbine inlet guide vane system based on the feature information fusion as claimed in claim 1, wherein in the step 2, the parameter of the variation modal decomposition is optimized by using a dolphin group algorithm in a group intelligent algorithm, and the step 2 comprises:
setting a parameter range of a modal decomposition number k and a penalty factor alpha, and initializing a population; decomposing the signal through the VMD to obtain a plurality of intrinsic mode function components, and selecting the minimum envelope entropy as a fitness function; then entering a searching stage to find out the optimal fitness value of the individual; entering a calling and receiving stage, and finding out a neighborhood optimal solution; in the predation stage, the position of the individual is updated to the optimal solution of the neighborhood, and the optimal fitness value is calculated; when the requirement of iteration times is met, outputting the optimal solution of k and alpha, wherein the fitness function FfitThe formula is as follows:
Ffit＝min(Een) (2)
in the formula, EenRepresenting the sparsity of the component signals for the envelope entropy;
Eenthe calculation formula is as follows:
Figure FDA0003274367420000021
wherein τ(s) is an envelope signal of the vibration signal x (t) after hilbert demodulation; e.g. of the typesNormalized value of τ(s); s is 1, 2.
4. The method for diagnosing the fault of the gas turbine inlet guide vane system based on the fusion of the characteristic information as claimed in claim 1, wherein in the step 3, the vibration signal is subjected to the variational modal decomposition by using the optimized parameters, the variational modal decomposition decomposes the vibration signal into k intrinsic modal function components, the decomposition sequence is guaranteed to be a modal component with a limited bandwidth of a middle frequency, meanwhile, the sum of the estimated bandwidths of all the modes is minimum, the constraint condition is that the sum of all the modes is equal to the original signal, the variational modal decomposition realizes the frequency domain division of the signal, so that the effective decomposition component of the vibration signal is obtained, and finally, the optimal solution of the variational problem is obtained;
the vibration signal is decomposed through a variational mode to obtain k intrinsic mode function components, and a constraint variational model formula is established for the vibration signal as follows:
Figure FDA0003274367420000031
in the formula uk(t) is the modal component; omegakIs the intermediate frequency of each modal component; k is the number of IMFs; δ (t) is the shock function; j is an imaginary unit; t is time in units of s;
Figure FDA0003274367420000032
is a harmonic signal; is a convolution calculation; s.t. is a constraint condition;
in order to solve the optimal solution of the constraint problem of the variational model in the formula (4), a quadratic penalty function and a Lagrange multiplier method are utilized to convert the solution into an unconstrained variational model shown in the formula (5):
Figure FDA0003274367420000033
wherein λ (t) is a Lagrangian multiplier;
Figure FDA0003274367420000034
represents a 2-norm;<·>representing the vector inner product;
in order to solve the unconstrained variational model, a saddle point of a Lagrange expression is sought through an alternative direction multiplier method until an iteration stop condition is met
Figure FDA0003274367420000035
Obtaining an optimal solution for the variational model; where ε is the precision.
5. The method for diagnosing the system fault of the inlet guide vane of the gas turbine based on the feature information fusion as claimed in claim 1, wherein the intrinsic mode function component sensitive to the fault information is screened by using kurtosis-mutual entropy in the step 3, and the formula is as follows:
Figure FDA0003274367420000041
Emi＝Hmi(x1)+Hmi(X(t))-Hmi(x1,X(t)) (7)
in the formula, KcKurtosis, indicating the severity of the fault;
Figure FDA0003274367420000042
is the average amplitude of signal X (t); k is the number of IMFs; emiThe entropy is mutual information entropy and represents the relevance between two signals; x is the number of1Is any one IMF component; hmi(. cndot.) is the entropy of the signal.
6. The method for diagnosing the fault of the gas turbine inlet guide vane system based on the feature information fusion as claimed in claim 1, wherein the step 4 comprises the following steps:
step 4.1, the IMF component after the decomposition of the variation mode decomposition is { uk(t)}＝{u1(t),u2(t),...uk(t) }, calculating instantaneous energy entropy Henergy：
Figure FDA0003274367420000043
Figure FDA0003274367420000044
In the formula riIs the percentage of the component energy of the ith intrinsic mode function to the total energy; i ═ {1,2, …, k } is the number of components of the sensitive eigenmode function;
and 4.2, in the inherent characteristics of the signals, singular values have excellent stability, and the energy change of the signals in the sampling time can be represented by singular value entropy. Forming eigenmode function components into a feature vector matrix u1(t),u2(t),...uk(t)]ΤCalculating singular values V of the feature matrixsingularThereby obtaining the singular entropy, and the formula is:
Figure FDA0003274367420000051
Figure FDA0003274367420000052
in the formula
Figure FDA0003274367420000053
The percentage of singular value of the ith sensitive eigenmode function component to the total singular value；
Step 4.3, expressing the change of signal energy in a frequency domain by the power spectrum entropy, carrying out Fourier transform on the intrinsic mode function component to obtain a power spectrum value chi, and obtaining the power spectrum entropy H according to the energy conservation law of signal time-frequency domain transformpowerThe formula is as follows:
Figure FDA0003274367420000054
Figure FDA0003274367420000055
Figure FDA0003274367420000056
in the formula piThe weight of the power spectrum of the i sensitive intrinsic mode function components in the total amount of the power spectrum is calculated;
step 4.4, the intrinsic mode function component uk(t) performing a phase space reconstruction, uk(t)' is the eigenmode function component after reconstruction, and a matrix is obtained:
Figure FDA0003274367420000057
Figure FDA0003274367420000058
where m is the embedding dimension; upsilon is a time delay; g is the number of vectors in the reconstructed phase space; a isiThe probability of occurrence of the i reconstructed eigenmode function components;
step 4.5, in order to enable different feature vectors y to be in the same interval before encoding, the feature vector composed of four entropy values is subjected to maximum and minimum normalization processing, so that the feature vector is normalized to be between [0 and 1], and the formula is as follows:
Figure FDA0003274367420000061
in the formula xp *Normalizing the value of the p characteristic; x is the number ofpIs the p-th feature; x is the number ofminIs the minimum of the p-th feature; x is the number ofmaxIs the maximum of the p-th feature.
7. The method for diagnosing the system fault of the inlet guide vane of the gas turbine based on the feature information fusion as claimed in claim 6, wherein in the step 5, the normalized feature vector features are converted into the pulse time sequence which can be identified by the input neurons through a group coding method, the group coding uses gamma overlapped Gaussian receptive field coding neurons to map the single features to a high-dimensional feature space, and the activation value formula of the coding neurons is as follows:
Figure FDA0003274367420000062
Figure FDA0003274367420000063
in the formula, AγThe activation value of the gamma encoding neuron under the corresponding characteristic value is obtained; mu.sγThe central value of the Gauss receptive field of the gamma-th coding neuron; sigmahTo encode the width of the neuronal gaussian receptive field.
8. The method for diagnosing the system fault of the inlet guide vane of the gas turbine based on the fusion of the characteristic information as claimed in claim 1, wherein in the step 6, the number of neurons in the input layer of the impulse neural network is the product of the state type number and the dimension of the characteristic vector, and the number of neurons in the output layer is the number of fault types; synaptic weight initialization between neurons of the input layer and the output layer obeys a uniform distribution between 0-1; the neuron model of the impulse neural network is an impulse response model, and the formula is as follows:
V(t)＝∑fωf∑tfK(t-tf)+Vrest (20)
wherein V (t) is neuron mode voltage in mv; f is the number of synapses; omegafWeight of the f-th synapse; t is tfTime of arrival of the f-th synaptic input pulse in units of s; vrestIs a resting potential; k (-) is the kernel function of the impulse response model; k (t-t)f) A kernel function of the arrival time of the f-th synaptic input pulse;
the kernel function K (t) of the time-dependent behavior of the postsynaptic potential is:
Figure FDA0003274367420000071
in the formula: tau ismAnd τsIs the decay time constant; vnorIs a normalized value;
in the training stage, the spikeProp algorithm is adopted by the impulse neural network to adjust the synaptic connection weight value through the error function feedback, and the error function formula is as follows:
Figure FDA0003274367420000072
in the formula (I), the compound is shown in the specification,
Figure FDA0003274367420000073
the instant at which neuron j actually fires a pulse,
Figure FDA0003274367420000074
the time at which a pulse is expected for neuron j;
optimizing an error function according to a gradient descent method, and updating a synapse weight value based on a chain-type derivation method:
Figure FDA0003274367420000075
in the formula (I), the compound is shown in the specification,
Figure FDA0003274367420000079
is as follows
Figure FDA0003274367420000076
Updating a synaptic weight between the input neuron and the output neuron; eta is the learning rate;
Figure FDA0003274367420000077
is the first to be updated
Figure FDA0003274367420000078
Synaptic weights between the input neurons and the output neurons.
CN202111112808.5A 2021-09-23 2021-09-23 Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion Pending CN113850181A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111112808.5A CN113850181A (en) 2021-09-23 2021-09-23 Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111112808.5A CN113850181A (en) 2021-09-23 2021-09-23 Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion

Publications (1)

Publication Number Publication Date
CN113850181A true CN113850181A (en) 2021-12-28

Family

ID=78979230

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111112808.5A Pending CN113850181A (en) 2021-09-23 2021-09-23 Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion

Country Status (1)

Country Link
CN (1) CN113850181A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115060497A (en) * 2022-06-10 2022-09-16 南昌工程学院 Bearing fault diagnosis method based on CEEMD energy entropy and optimized PNN
CN116304648A (en) * 2023-05-23 2023-06-23 北京化工大学 Gear fault identification method based on optimized pulse enhancement and envelope synchronous averaging
CN116756675A (en) * 2023-08-14 2023-09-15 湘江实验室 Bearing fault diagnosis and classification method based on impulse neural network

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115060497A (en) * 2022-06-10 2022-09-16 南昌工程学院 Bearing fault diagnosis method based on CEEMD energy entropy and optimized PNN
CN116304648A (en) * 2023-05-23 2023-06-23 北京化工大学 Gear fault identification method based on optimized pulse enhancement and envelope synchronous averaging
CN116304648B (en) * 2023-05-23 2023-08-29 北京化工大学 Gear fault identification method based on optimized pulse enhancement and envelope synchronous averaging
CN116756675A (en) * 2023-08-14 2023-09-15 湘江实验室 Bearing fault diagnosis and classification method based on impulse neural network
CN116756675B (en) * 2023-08-14 2023-11-07 湘江实验室 Bearing fault diagnosis and classification method based on impulse neural network

Similar Documents

Publication Publication Date Title
CN113850181A (en) Gas turbine inlet guide vane system fault diagnosis method based on feature information fusion
CN110441065B (en) Gas turbine on-line detection method and device based on LSTM
Xia et al. Multi-stage fault diagnosis framework for rolling bearing based on OHF Elman AdaBoost-Bagging algorithm
Liao et al. Hydroelectric generating unit fault diagnosis using 1-D convolutional neural network and gated recurrent unit in small hydro
Vashishtha et al. An amended grey wolf optimization with mutation strategy to diagnose bucket defects in Pelton wheel
CN111815806B (en) Method for preprocessing flight parameter data based on wild value elimination and feature extraction
CN108073158A (en) Based on PCA and KNN density algorithm Wind turbines Method for Bearing Fault Diagnosis
CN105628425A (en) Rotation machinery early stage fault diagnosis method based on heredity annealing optimization multi-core support vector machine
CN111275004B (en) Bearing fault diagnosis method based on LMD and impulse neural network
CN108919116B (en) Ocean current generator unbalanced stator current fault diagnosis method based on MCCKAF-FFT-Softmax
Feng et al. Finite-sensor fault-diagnosis simulation study of gas turbine engine using information entropy and deep belief networks
Yang et al. Gearbox fault diagnosis based on artificial neural network and genetic algorithms
CN109471049B (en) Satellite power supply system anomaly detection method based on improved stacked self-encoder
CN112101227A (en) Mechanical state monitoring method based on CELMDAN and SSKFDA
Huang et al. Deep residual networks-based intelligent fault diagnosis method of planetary gearboxes in cloud environments
CN114492217B (en) Typhoon and wave height prediction method based on mixed time series perceptron model
CN113642508A (en) Bearing fault diagnosis method based on parameter self-adaptive VMD and optimized SVM
CN113705094A (en) Ship fuel oil pipeline fault prediction method based on PSO-GRU
Han et al. Data-enhanced stacked autoencoders for insufficient fault classification of machinery and its understanding via visualization
Lei et al. P-DNN: An effective intrusion detection method based on pruning deep neural network
Loboda et al. A benchmarking analysis of a data-driven gas turbine diagnostic approach
Li et al. Fault diagnosis of aeroengine control system sensor based on optimized and fused multidomain feature
CN111400964B (en) Fault occurrence time prediction method and device
CN110633741B (en) Time sequence classification method based on improved impulse neural network
Li et al. Class imbalanced fault diagnosis via combining K-means clustering algorithm with generative adversarial networks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination