Fault diagnosis method of gas turbine inlet guide vane system based on feature information fusion

Abstract

The present invention discloses a method for fault diagnosis of gas turbine inlet guide vane system based on feature information fusion; including: original vibration signal acquisition; fault mechanism analysis; variational mode decomposition (VMD) parameter optimization and decomposition; fault feature extraction; state feature vector normalization; feature vector encoding; pulse neural network (SNN) fault diagnosis. The present invention clearly adopts the dolphin group algorithm to optimize VMD parameters to improve the decomposition accuracy; based on the kurtosis-mutual information entropy, the IMF components sensitive to fault information are screened, and the fault feature sensitive modal functions with poor distribution rules and few impact components are eliminated; a multi-feature entropy value algorithm is used to extract fault features in the time-frequency domain to avoid a single feature from being unable to fully reflect the fault feature information, thereby providing a guarantee for accurate fault diagnosis; the SpikeProp algorithm is used to optimize the SNN, which has the ability to solve nonlinear classification problems and makes the training results more accurate.

Description

Technical Field

[0001] The invention belongs to the technical field of gas turbine fault diagnosis, and in particular relates to a gas turbine inlet guide vane system fault diagnosis method based on feature information fusion. Background Art

[0002] Gas turbines are internal combustion power machines that use continuous flow of natural gas as a working fluid to drive the impeller to rotate at high speed. They have been widely used in aerospace, chemical and other fields. Heavy-duty gas turbines used for industrial power generation are an important power system for my country to achieve energy conservation and environmental protection. The inlet guide vanes are a series of stationary blades in front of the first-stage moving blades of the gas turbine compressor. As the core component of the gas turbine compressor, they bear the heavy responsibility of energy conversion. By changing the angle of the guide vanes, the angle and flow rate of the airflow entering the compressor are controlled. Once the inlet guide vane system fails, it will seriously endanger the normal operation of the gas turbine, and even worse, it will cause a safety accident.

[0003] When the inlet guide vane system fails, the vibration signal received by the sensor becomes a significant non-stationary signal, and a large amount of characteristic information is hidden in these signals. Most of the current signal analysis methods are empirical mode decomposition (EMD), local mean decomposition (LMD), etc. EMD is prone to modal aliasing problems; LMD is prone to signal singularity characteristics loss. The proposed fault diagnosis method based on feature information fusion uses variational mode decomposition (VMD) to adaptively process the signal, which overcomes the problems of modal aliasing, signal characteristic loss and boundary effect, and is more superior.

[0004] In order to solve the problem that minor faults cannot be described, the present invention, based on the introduction of the concept of entropy, proposes a fault diagnosis method based on feature information fusion to analyze multi-scale and multi-features, extracts features through four entropy functions, and uses a pulse neural network (SNN) for fault diagnosis. Compared with deep neural networks, it has the advantages of high interpretability, high performance and low power consumption. Summary of the invention

[0005] In view of the problems existing in the background technology, the present invention provides a method for diagnosing faults of an inlet guide vane system of a gas turbine based on feature information fusion, which is characterized by comprising:

[0006] Step 1, collecting the vibration signal data of the gas turbine inlet guide vane system and performing fault mechanism analysis on the data to obtain the change trend of the vibration signal amplitude at the time of the fault;

[0007] Step 2: Optimize the parameters of variational mode decomposition using swarm intelligence algorithm;

[0008] Step 3: Use the optimized parameters to perform variational mode decomposition on the vibration signal to obtain k intrinsic mode function components, and select the intrinsic mode function components that are sensitive to fault information based on kurtosis-mutual information entropy;

[0009] Step 4: Use a multi-feature entropy algorithm to extract fault features in the time-frequency domain, construct a state feature vector, and normalize it;

[0010] Step 5: Encode the normalized feature vector into a pulse time series to obtain a training set;

[0011] Step 6: Input the training set into the spiking neural network model for training, construct the spiking neural network model, and then use the SpikeProp algorithm to optimize the spiking neural network;

[0012] Step 7, sampling the inlet guide vane system fault data and repeating steps 2 to 5 to obtain a pulse signal that can be recognized by the pulse neural network model as a test set; inputting it into the trained pulse neural network model to obtain the membrane voltage of the output neuron, when the membrane voltage of a neuron reaches the threshold and emits a pulse, while other neurons have not reached the threshold and do not emit pulses, the fault category of the vibration signal to be diagnosed is obtained.

[0013] The variation trend of the vibration signal amplitude at the time of the fault in step 1 is as follows: the vibration signal generated by the gas turbine inlet guide vane system is transmitted from the blade root to the casing surface and then to the vibration sensor; wherein the vibration signal amplitude will suddenly change at the time of the fault, and the vibration signal amplitude r a for:

[0014]

[0015] In the formula, ω c is the natural frequency, Hz; ζ is the damping ratio; Ω is the speed, r / s; is the vibration phase, rad; l is the distance between the center of mass before and after the guide vane breaks, cm; t is the time, unit is s.

[0016] In step 2, the parameters of variational mode decomposition are optimized using the dolphin group algorithm in the swarm intelligence algorithm. Step 2 is divided into:

[0017] Set the parameter range of the number of mode decompositions k and the penalty factor α, and initialize the population; decompose the signal through VMD to obtain multiple intrinsic mode function components, and select the minimum envelope entropy as the fitness function; then enter the search phase to find the individual's optimal fitness value; enter the call and receive phase to find the optimal solution in the neighborhood; enter the predation phase, the individual updates the position to the optimal solution in the neighborhood, and calculates the optimal fitness value; when the number of iterations is met, output the optimal solution of k and α, where the fitness function Ffit The formula is:

[0018] F fit =min(E en ) (2)

[0019] In the formula, E en is the envelope entropy, which indicates the sparsity of the component signal;

[0020] E en The calculation formula is:

[0021]

[0022] Where τ(s) is the envelope signal of the vibration signal X(t) after Hilbert demodulation; s is the normalized value of τ(s); s=1,2,...,n.

[0023] In the step 3, the vibration signal is subjected to variational modal decomposition using the optimized parameters. The variational modal decomposition decomposes the vibration signal into k intrinsic modal function components, ensuring that the decomposition sequence is a modal component with a limited bandwidth of an intermediate frequency, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original signal. The variational modal decomposition realizes the frequency domain division of the signal, thereby obtaining the effective decomposition components of the vibration signal, and finally obtaining the optimal solution to the variational problem;

[0024] Among them, the vibration signal is decomposed by variational mode decomposition to obtain k intrinsic mode function components, and the constrained variational model formula for the vibration signal is established as follows:

[0025]

[0026] In the formula, u k (t) is the modal component; ω k is the intermediate frequency of each modal component; k is the number of IMFs; δ(t) is the impulse function; j is the imaginary unit; t is the time, in seconds; is the harmonic signal; * is the convolution calculation; st is the constraint condition;

[0027] In order to find the optimal solution to the constraint problem of the variational model in equation (4), the quadratic penalty function and Lagrange multiplier method are used to transform it into the unconstrained variational model shown in equation (5):

[0028]

[0029] Where λ(t) is the Lagrange multiplier; represents the 2-norm; <·> represents the vector inner product;

[0030] In order to solve the unconstrained variational model, the saddle point of the Lagrangian expression is sought by the alternating direction multiplier method until the iteration stop condition is met. That is, the optimal solution is obtained for the variational model; where ε is the accuracy.

[0031] In step 3, the kurtosis-mutual information entropy is used to screen the intrinsic mode function components that are sensitive to fault information. The formula is:

[0032]

[0033] E mi =H mi (x1)+H mi (X(t))-H mi (x1,X(t)) (7)

[0034] In the formula, K c is the kurtosis, indicating the severity of the fault; is the average amplitude of signal X(t); k is the number of IMFs; E mi is the mutual information entropy, indicating the correlation between two signals; x1 is any IMF component; H mi (·) is the entropy value of the signal.

[0035] The step 4 is divided into:

[0036] Step 4.1: The IMF components after variational mode decomposition are {u k (t)}={u1(t),u2(t),...u k (t)}, calculate the instantaneous energy entropy H energy :

[0037]

[0038]

[0039] Where r i is the percentage of the energy of the i-th intrinsic mode function component to the total energy; i = {1, 2, …, k} is the number of sensitive intrinsic mode function components;

[0040] Step 4.2: Among the inherent characteristics of the signal, the singular value has excellent stability, and the energy change of the signal during the sampling time can be expressed by the singular value entropy. The intrinsic mode function components are formed into an eigenvector matrix [u1(t),u2(t),...u k (t)] Τ , calculate the singular values ​​of the characteristic matrix V singular , thus obtaining the singular entropy, the formula is:

[0041]

[0042]

[0043] In the formula is the percentage of the singular value of the i-th sensitive eigenmode function component to the total number of singular values;

[0044] Step 4.3: Power spectrum entropy represents the change of signal energy in the frequency domain. The intrinsic mode function components are Fourier transformed to obtain the power spectrum value χ. According to the energy conservation law of signal time-frequency domain transformation, the power spectrum entropy H is obtained. power , the formula is:

[0045]

[0046]

[0047]

[0048] Where p i is the weight of the power spectrum of the i-th sensitive intrinsic mode function component in the total power spectrum;

[0049] Step 4.4: Substitute the intrinsic mode function component u k (t) Reconstruct the phase space, u k (t)′ is the reconstructed intrinsic mode function component, and the matrix is ​​obtained:

[0050]

[0051]

[0052] Where m is the embedding dimension; υ is the time delay; g is the number of vectors in the reconstructed phase space; a i is the probability of occurrence of i reconstructed eigenmode function components;

[0053] Step 4.5: In order to make different feature vectors y in the same interval before encoding, the feature vectors composed of four entropy values ​​are normalized to the maximum and minimum values ​​so that they are normalized to [0,1]. The formula is:

[0054]

[0055] Where x p * is the normalized value of the pth feature; x p is the pth feature; x min is the minimum value of the pth feature; x max is the maximum value of the pth feature.

[0056] In step 5, the normalized feature vector features are converted into a pulse time series that can be recognized by the input neurons through the group coding method. The group coding uses γ overlapping Gaussian receptive field coding neurons to map a single feature to a high-dimensional feature space. The activation value formula of the coding neuron is:

[0057]

[0058]

[0059] In the formula, A γ is the activation value of the γth encoding neuron corresponding to the eigenvalue; μ γ is the center value of the Gaussian receptive field of the γth encoding neuron; σ h is the width of the Gaussian receptive field of the encoding neuron.

[0060] In step 6, the number of neurons in the input layer of the spiking neural network is the product of the number of state types and the dimension of the feature vector, and the number of neurons in the output layer is the number of fault categories; the synaptic weights between the neurons in the input layer and the output layer are initialized to obey a uniform distribution between 0 and 1; the neuron model of the spiking neural network is an impulse response model, and the formula is:

[0061] V(t)=∑ f ω f ∑t f K(tt f )+V rest (20)

[0062] Where V(t) is the neuron voltage in mV; f is the number of synapses; ω f is the weight of the fth synapse; t f is the arrival time of the fth synaptic input pulse, in seconds; V rest is the resting potential; K(·) is the kernel function of the impulse response model; K(tt f ) is the kernel function of the arrival time of the input pulse of the f-th synapse;

[0063] The kernel function K(t) of the time-varying characteristics of the postsynaptic potential is:

[0064]

[0065] Where: τ m and τ s is the decay time constant; V nor is the normalized value;

[0066] During the training phase, the spiking neural network uses the SpikeProp algorithm to adjust the synaptic connection weight value through error function feedback. The error function formula is:

[0067]

[0068] In the formula, is the time when neuron j actually emits a pulse, is the time when neuron j is expected to emit a pulse;

[0069] The error function is optimized according to the gradient descent method, and the synaptic weight value is updated based on the chain rule:

[0070]

[0071] In the formula, For the The synaptic weight update value between the input neuron and the output neuron; η is the learning rate; To be updated The synaptic weights between the input neurons and the output neurons.

[0072] The beneficial effects of the present invention are:

[0073] 1. The VMD parameters are optimized through the dolphin group algorithm in the swarm intelligence algorithm, and the VMD parameters are adaptively selected to improve the decomposition accuracy; the IMF components sensitive to fault information are screened based on the kurtosis-mutual information entropy, and the fault characteristic sensitive modal functions with poor distribution rules and few impact components are eliminated.

[0074] 2. Use multi-feature entropy algorithm to extract fault features in the time-frequency domain to avoid the situation where a single feature cannot fully reflect the fault feature information, thus providing a guarantee for accurate fault diagnosis.

[0075] 3. Combined with the SRM neuron model, the SpikeProp algorithm is used to optimize SNN, which has the ability to solve nonlinear classification problems and makes the training results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 The present invention is a flow chart of an embodiment of a method for diagnosing faults of a gas turbine inlet guide vane system based on feature information fusion.

[0077] Figure 2 It is a flow chart of optimizing VMD parameters by the dolphin group algorithm provided by the present invention.

[0078] Figure 3 It is a decomposition diagram of the vibration signal of the optimized VMD provided by the present invention.

[0079] Figure 4 This is a diagram of the SNN model diagnosis results provided by the present invention. DETAILED DESCRIPTION

[0080] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0081] like Figure 1 The embodiment of the present invention shown includes the following steps:

[0082] Step 1, collecting the vibration signal data of the gas turbine inlet guide vane system and performing fault mechanism analysis on the data to obtain the change trend of the vibration signal amplitude at the time of the fault;

[0083] Step 2: Optimize the parameters of variational mode decomposition (VMD) using swarm intelligence algorithm;

[0084] Step 3: Use the optimized parameters to perform VMD decomposition on the vibration signal to obtain k intrinsic mode function (IMF) components, and select the IMF components that are sensitive to fault information based on kurtosis-mutual information entropy;

[0085] Step 4: Use a multi-feature entropy algorithm to extract fault features in the time-frequency domain, construct a state feature vector, and normalize it;

[0086] Step 5: Encode the normalized feature vector into a pulse time series to obtain a training set;

[0087] Step 6: Input the training set into the SNN model for training, build the SNN model, and then use the SpikeProp algorithm to optimize the SNN;

[0088] Step 7: Resample the inlet guide vane system fault data, and repeat steps 2 to 5 to obtain a pulse signal that can be recognized by the SNN model as a test set, which is input into the trained SNN model to obtain the membrane voltage of the output neuron. When the membrane voltage of a neuron reaches the threshold and emits a pulse, while other neurons have not reached the threshold and do not emit pulses, the fault category of the vibration signal to be diagnosed is obtained.

[0089] The variation trend of the vibration signal amplitude at the time of the fault in step 1 is as follows: the vibration signal generated by the gas turbine inlet guide vane system is transmitted from the blade root to the casing surface and then to the vibration sensor. The vibration signal amplitude will suddenly change at the time of the fault. a for:

[0090]

[0091] In the formula, ω c is the natural frequency, Hz; ζ is the damping ratio; Ω is the speed, r / s; is the vibration phase, rad; l is the distance between the center of mass before and after the guide vane breaks, cm; t is the time, unit is s.

[0092] In step 2, the dolphin swarm algorithm (DSA) in the swarm intelligence algorithm is used to optimize the parameters of VMD. The swarm intelligence algorithm is simple, robust, and widely applicable. The DSA algorithm simulates the biological characteristics of dolphins such as echolocation, information exchange, division of labor and cooperation, and realizes the optimization of the algorithm through four key stages: searching, calling, receiving, and predation. Other swarm intelligence algorithms only use forward solutions, while the DSA algorithm uses echolocation and adopts different strategies to make it easier to find the optimal solution. Figure 2 It is a flow chart of optimizing VMD parameters by the dolphin group algorithm provided by the present invention, which is divided into:

[0093] First, set the parameter range of the number of modal decompositions k and the penalty factor α. Specifically, set the value range of k to 2-12 and the value range of α to 100-5000; and initialize the population; decompose the signal through VMD to obtain multiple IMF components, and select the minimum envelope entropy as the fitness function; then enter the search phase to find the individual's optimal fitness value; enter the calling and receiving phase to find the optimal solution in the neighborhood; enter the predation phase, the individual updates its position to the optimal solution in the neighborhood and calculates the optimal fitness value; when the number of iterations required is met, output the optimal solution of k and α.

[0094] The present invention obtains the optimal parameters k=6 and α=2000 by parameter optimization, which effectively improves the accuracy of VMD decomposition and thus establishes a fault diagnosis model more accurately.

[0095] Fitness function F fit The formula is:

[0096] F fit =min(E en ) (2)

[0097] In the formula, E en is the envelope entropy, which indicates the sparsity of the component signal;

[0098] E en The calculation formula is:

[0099]

[0100] Where τ(s) is the envelope signal of the vibration signal X(t) after Hilbert demodulation; s is the normalized value of τ(s); s=1,2,...,n.

[0101] In step 3, the optimized parameters are used to perform VMD decomposition on the vibration signal. VMD decomposition decomposes the vibration signal into k IMF components, ensuring that the decomposition sequence is a modal component with a limited bandwidth and an intermediate frequency. At the same time, the sum of the estimated bandwidths of each mode is minimized, and the constraint condition is that the sum of all modes is equal to the original signal. VMD decomposition realizes the frequency domain division of the signal, and then obtains the effective decomposition components of the vibration signal, and finally obtains the optimal solution to the variational problem. The results are shown in Figure 3 As shown in the figure, after the vibration signal is decomposed by VMD, six intrinsic mode function components are obtained. The constrained variational model formula for the vibration signal is:

[0102]

[0103] In the formula, u k (t) is the modal component; ω k is the intermediate frequency of each modal component; k is the number of IMFs; δ(t) is the impulse function; j is the imaginary unit; t is the time, in seconds; is the harmonic signal; * is the convolution calculation; st is the constraint condition.

[0104] In order to find the optimal solution to the constraint problem of the variational model in equation (4), the quadratic penalty function and Lagrange multiplier method are used to transform it into the unconstrained variational model shown in equation (5):

[0105]

[0106] Where λ(t) is the Lagrange multiplier; represents the 2-norm; <·> represents the vector inner product.

[0107] In order to solve the unconstrained variational model, the saddle point of the Lagrangian expression is sought by the alternating direction multiplier method until the iteration stop condition is met. That is, the optimal solution is obtained for the variational model. Where ε is the accuracy, which is set to 1e-7 during the iteration process.

[0108] In step 3, kurtosis-mutual information entropy is used to screen the IMF components sensitive to fault information, and the formula is:

[0109]

[0110] E mi =H mi (x1)+H mi (X(t))-H mi (x1,X(t)) (7)

[0111] In the formula, K c is the kurtosis, indicating the severity of the fault; is the average amplitude of signal X(t); k is the number of IMFs; E mi is the mutual information entropy, indicating the correlation between two signals; x1 is any IMF component; H mi (·) is the entropy value of the signal.

[0112] In step 4, in order to avoid the single entropy value failing to fully reflect the fault characteristics, a multi-feature entropy value algorithm is used to calculate the instantaneous energy entropy H energy 、Singular value entropy H singular , power spectrum entropy H power , Permutation Entropy H arrange The feature vector composed of the four entropy values ​​is used to extract fault features in the time-frequency domain at multiple scales, which are specifically divided into:

[0113] Step 4.1: Energy entropy represents the change in the energy contained in the vibration signal, reflects the change in the amplitude of the signal in the time domain, and reveals the intrinsic characteristics of the signal. energy It is obtained by combining energy entropy with information entropy. The overall energy of the frequency band is used to reflect the overall energy distribution of the mechanical fault signal. Different mechanical faults have different energy distributions, so the instantaneous energy entropy is also different. In step 3, the IMF component (modal component) after VMD decomposition is {u k (t)}={u1(t),u2(t),...u k (t)}, calculate the instantaneous energy entropy H energy :

[0114]

[0115]

[0116] Where r i is the percentage of the energy of the i-th IMF component to the total energy; i = {1, 2, …, k} is the number of sensitive IMF components.

[0117] Step 4.2: Among the inherent characteristics of the signal, the singular value has excellent stability, and the energy change of the signal during the sampling time can be expressed by the singular value entropy. The IMF components are formed into an eigenvector matrix [u1(t),u2(t),...u k (t)] Τ , calculate the singular values ​​of the characteristic matrix V singular , thus obtaining the singular entropy, the formula is:

[0118]

[0119]

[0120] In the formula is the percentage of the singular value of the i-th sensitive IMF component to the total number of singular values.

[0121] Step 4.3: The power spectrum entropy represents the change of signal energy in the frequency domain. The IMF component is Fourier transformed to obtain the power spectrum value χ. According to the energy conservation law of signal time-frequency domain transformation, the power spectrum entropy H is obtained. power , the formula is:

[0122]

[0123]

[0124]

[0125] Where p i is the weight of the power spectrum of the i-th sensitive IMF component in the total power spectrum.

[0126] Step 4.4: Permutation entropy H arrange It is an effective method to characterize the complexity and randomness of time series, and has the characteristics of simple calculation and excellent noise robustness. k (t) Reconstruct the phase space, u k (t)′ is the reconstructed IMF component, and the matrix is ​​obtained:

[0127]

[0128]

[0129] Where m is the embedding dimension; υ is the time delay; g is the number of vectors in the reconstructed phase space; a i is the probability of occurrence of the i-th reconstructed IMF component.

[0130] Step 4.5: In order to make different feature vectors y in the same interval before encoding, the feature vectors composed of four entropy values ​​are normalized to the maximum and minimum values ​​so that they are normalized to [0,1]. The formula is:

[0131]

[0132] Where x p * is the normalized value of the pth feature; x p is the pth feature; x min is the minimum value of the pth feature; x max is the maximum value of the pth feature.

[0133] In step 5, the normalized feature vector feature is converted into a pulse time sequence that can be recognized by the input neuron through the group coding method. The group coding uses γ overlapping Gaussian receptive field coding neurons to map a single feature to a high-dimensional feature space. The range of γ is 1-8, and the activation value formula of the coding neuron is:

[0134]

[0135]

[0136] In the formula, A γ is the activation value of the γth encoding neuron corresponding to the eigenvalue; μ γ is the center value of the Gaussian receptive field of the γth encoding neuron; σ h is the width of the Gaussian receptive field of the encoding neuron.

[0137] In step 6, the number of neurons in the SNN input layer is the product of the number of state types (including normal and faulty) and the dimension of the feature vector, and the number of neurons in the output layer is the number of fault categories; the synaptic weights between the neurons in the input layer and the output layer are initialized to obey a uniform distribution between 0 and 1; the SNN neuron model is a pulse response model (SRM), and the formula is:

[0138] V(t)=Σ f ω f ∑t f K(tt f )+V rest (20)

[0139] Where V(t) is the neuron voltage in mV; f is the number of synapses; ω f is the weight of the fth synapse; t f is the arrival time of the fth synaptic input pulse, in seconds; V rest is the resting potential; K(·) is the kernel function of the impulse response model; K(tt f ) is the kernel function of the arrival time of the input pulse of the f-th synapse;

[0140] The kernel function K(t) of the time-varying characteristics of the postsynaptic potential is:

[0141]

[0142] Where: τ m and τ s is the decay time constant; V nor is the normalized value.

[0143] In the training phase, SNN uses the supervised SpikeProp algorithm to adjust the synaptic connection weight value through error function feedback. The error function formula is:

[0144]

[0145] In the formula, is the time when neuron j actually emits a pulse, is the time when neuron j is expected to fire a pulse.

[0146] The error function is optimized according to the gradient descent method, and the synaptic weight value is updated based on the chain rule:

[0147]

[0148] In the formula, For the The synaptic weight update value between the input neuron and the output neuron; η is the learning rate, which is set to 0.001 during training; To be updated The synaptic weights between the input neurons and the output neurons.

[0149] In the step 7, 100s of inlet guide vane system fault data are obtained under the condition of a sampling frequency of 100 Hz, and the vibration signal of the inlet guide vane system to be diagnosed is processed according to steps 2 to 5 and input into the trained SNN model. The input neuron continuously sends a pulse sequence to the output neuron to stimulate the output neuron membrane voltage to rise. When the membrane voltage of a neuron in the output layer reaches the threshold and sends a pulse, while other neurons have not reached the threshold and do not send pulses, the fault category of the vibration signal to be diagnosed is obtained, wherein the neuron membrane voltage threshold is set to 1.

[0150] like Figure 4 As shown in the figure, the sensor is used to collect the vibration signals of a gas turbine inlet guide vane system under normal and three sudden fault conditions. Then, 100s of data are taken at a sampling frequency of 100Hz and the fault mechanism is analyzed to obtain the change trend of the vibration signal amplitude. Through steps 2 to 5, it is converted into a pulse signal that can be recognized by the SNN model. The membrane voltage of the output neuron represented by the fault F2 receives the pulse stimulation from the input neuron within 100s and reaches the threshold voltage at about 78s. The other three output neurons are regulated by the SpikeProp algorithm, which inhibits the weight of the synapse connecting them to the input neuron and fails to reach the threshold voltage, thereby judging the fault category.

[0151] The above is only a description of a specific embodiment of the present invention, and the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for fault diagnosis of gas turbine inlet guide vane system based on feature information fusion, characterized in that: include: Step 1, collecting the vibration signal data of the gas turbine inlet guide vane system and performing fault mechanism analysis on the data to obtain the change trend of the vibration signal amplitude at the time of the fault; Step 2: Optimize the parameters of variational mode decomposition using swarm intelligence algorithm; Step 3: Use the optimized parameters to perform variational modal decomposition on the vibration signal to obtain k intrinsic modal function components, and select the intrinsic modal function components that are sensitive to fault information based on kurtosis-mutual information entropy; Step 4: Use a multi-feature entropy algorithm to extract fault features in the time-frequency domain, construct a state feature vector, and normalize it; Step 5: Encode the normalized feature vector into a pulse time series to obtain a training set; Step 6: Input the training set into the spiking neural network model for training, construct the spiking neural network model, and then use the SpikeProp algorithm to optimize the spiking neural network; Step 7, sampling the inlet guide vane system fault data and repeating steps 2 to 5 to obtain a pulse signal that can be recognized by the pulse neural network model as a test set; inputting it into the trained pulse neural network model to obtain the membrane voltage of the output neuron, when the membrane voltage of a certain neuron reaches the threshold and emits a pulse, while other neurons have not reached the threshold and do not emit pulses, the fault category of the vibration signal to be diagnosed is obtained; In step 2, the parameters of variational mode decomposition are optimized using the dolphin group algorithm in the swarm intelligence algorithm. Step 2 is divided into: Set the parameter range of the modal decomposition number k and the penalty factor α, and initialize the population; decompose the signal through VMD to obtain multiple intrinsic mode function components, and select the minimum envelope entropy as the fitness function; Then enter the search phase to find the optimal individual fitness value; enter the call and receive phase to find the optimal solution in the neighborhood; Entering the predation stage, the individual updates its position to the optimal solution in the neighborhood and calculates the optimal fitness value; when the number of iterations is met, the optimal solution of k and α is output, where the fitness function F fit The formula is: F fit =min(E en ) (2) In the formula, E en is the envelope entropy, which indicates the sparsity of the component signal; E en The calculation formula is: Where τ(s) is the envelope signal of the vibration signal X(t) after Hilbert demodulation; s is the normalized value of τ(s); s=1,2,...,n.

2. A method for diagnosing faults of a gas turbine inlet guide vane system based on feature information fusion according to claim 1, characterized in that: The variation trend of the vibration signal amplitude at the time of the fault in step 1 is as follows: the vibration signal generated by the gas turbine inlet guide vane system is transmitted from the blade root to the casing surface and then to the vibration sensor; wherein the vibration signal amplitude will suddenly change at the time of the fault, and the vibration signal amplitude r a for: In the formula, ω c is the natural frequency, Hz; ζ is the damping ratio; Ω is the speed, r / s; is the vibration phase, rad; l is the distance between the center of mass before and after the guide vane breaks, cm; t is the time, unit is s.

3. A method for diagnosing faults of a gas turbine inlet guide vane system based on feature information fusion according to claim 1, characterized in that: In the step 3, the vibration signal is subjected to variational modal decomposition using the optimized parameters. The variational modal decomposition decomposes the vibration signal into k intrinsic modal function components, ensuring that the decomposition sequence is a modal component with a limited bandwidth of an intermediate frequency, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original signal. The variational modal decomposition realizes the frequency domain division of the signal, thereby obtaining the effective decomposition components of the vibration signal, and finally obtaining the optimal solution to the variational problem; Among them, after the vibration signal is decomposed by variational mode, k intrinsic mode function components are obtained, and the constrained variational model formula for the vibration signal is established as follows: In the formula, u k (t) is the modal component; ω k is the intermediate frequency of each modal component; k is the number of IMFs; δ(t) is the impulse function; j is the imaginary unit; t is time, in seconds; is the harmonic signal; * is the convolution calculation; st is the constraint condition; In order to find the optimal solution to the constraint problem of the variational model in equation (4), the quadratic penalty function and Lagrange multiplier method are used to transform it into the unconstrained variational model shown in equation (5): Where λ(t) is the Lagrange multiplier; represents the 2-norm; <·> represents the vector inner product; In order to solve the unconstrained variational model, the saddle point of the Lagrangian expression is sought by the alternating direction multiplier method until the iteration stop condition is met. That is, the optimal solution is obtained for the variational model; where ε is the accuracy.

4. A method for diagnosing faults of a gas turbine inlet guide vane system based on feature information fusion according to claim 1, characterized in that: In step 3, the kurtosis-mutual information entropy is used to screen the intrinsic mode function components that are sensitive to fault information. The formula is: E mi =H mi (x1)+H mi (X(t))-H mi (x1,X(t)) (7) In the formula, K c is the kurtosis, indicating the severity of the fault; is the average amplitude of signal X(t); k is the number of IMFs; E mi is the mutual information entropy, indicating the correlation between two signals; x1 is any IMF component; H mi (·) is the entropy value of the signal.

5. A method for gas turbine inlet guide vane system fault diagnosis based on feature information fusion according to claim 1, characterized in that: The step 4 is divided into: Step 4.1, the IMF components after variational mode decomposition are {u k (t)}={u1(t),u2(t),...u k (t)}, calculate the instantaneous energy entropy H energy : Where r i is the percentage of the energy of the i-th intrinsic mode function component to the total energy; i = {1, 2, …, k} is the number of sensitive intrinsic mode function components; Step 4.2: Among the inherent characteristics of the signal, the singular value has excellent stability. The energy change of the signal within the sampling time can be expressed by the singular value entropy. The components of the intrinsic mode function are formed into an eigenvector matrix [u1(t),u2(t),…u k (t)] Τ , calculate the singular values ​​of the characteristic matrix V singular , thus obtaining the singular entropy, the formula is: In the formula is the percentage of the singular value of the i-th sensitive eigenmode function component to the total number of singular values; Step 4.3: The power spectrum entropy represents the change of signal energy in the frequency domain. The intrinsic mode function components are Fourier transformed to obtain the power spectrum value χ. According to the energy conservation law of signal time-frequency domain transformation, the power spectrum entropy H is obtained. power , the formula is: Where p i is the weight of the power spectrum of the i-th sensitive intrinsic mode function component in the total power spectrum; Step 4.4: Substitute the intrinsic mode function component u k (t) Reconstruct the phase space, u k (t)′ is the reconstructed intrinsic mode function component, and the matrix is ​​obtained: Where m is the embedding dimension; υ is the time delay; g is the number of vectors in the reconstructed phase space; a i is the probability of occurrence of i reconstructed eigenmode function components; Step 4.5: In order to make different feature vectors y in the same interval before encoding, the feature vectors composed of four entropy values ​​are normalized to the maximum and minimum values ​​so that they are normalized to [0,1]. The formula is: Where x p * is the normalized value of the pth feature; x p is the pth feature; x min is the minimum value of the pth feature; x max is the maximum value of the pth feature.

6. A method for diagnosing faults of a gas turbine inlet guide vane system based on feature information fusion according to claim 5, characterized in that: In step 5, the normalized feature vector features are converted into a pulse time series that can be recognized by the input neurons through the group coding method. The group coding uses γ overlapping Gaussian receptive field coding neurons to map a single feature to a high-dimensional feature space. The activation value formula of the coding neuron is: In the formula, A γ is the activation value of the γth encoding neuron corresponding to the eigenvalue; μ γ is the center value of the Gaussian receptive field of the γth encoding neuron; σ h is the width of the Gaussian receptive field of the encoding neuron.

7. A method for gas turbine inlet guide vane system fault diagnosis based on feature information fusion according to claim 1, characterized in that: In step 6, the number of neurons in the input layer of the spiking neural network is the product of the number of state types and the dimension of the feature vector, and the number of neurons in the output layer is the number of fault categories; the synaptic weights between the neurons in the input layer and the output layer are initialized to obey a uniform distribution between 0 and 1; the neuron model of the spiking neural network is an impulse response model, and the formula is: Where V(t) is the neuron voltage in mV; f is the number of synapses; ω f is the weight of the fth synapse; t f is the arrival time of the fth synaptic input pulse, in seconds; V rest is the resting potential; K(·) is the kernel function of the impulse response model; K(tt f ) is the kernel function of the arrival time of the input pulse of the f-th synapse; The kernel function K(t) of the time-varying characteristics of the postsynaptic potential is: Where: τ m and τ s is the decay time constant; V nor is the normalized value; During the training phase, the spiking neural network uses the SpikeProp algorithm to adjust the synaptic connection weight value through error function feedback. The error function formula is: In the formula, is the time when neuron j actually emits a pulse, is the time when neuron j is expected to emit a pulse; The error function is optimized according to the gradient descent method, and the synaptic weight value is updated based on the chain rule: In the formula, For the The synaptic weight update value between the input neuron and the output neuron; η is the learning rate; To be updated The synaptic weights between the input neurons and the output neurons.

Images