CN112347917B - Gas turbine fault diagnosis method, system, equipment and storage medium - Google Patents
Gas turbine fault diagnosis method, system, equipment and storage medium Download PDFInfo
- Publication number
- CN112347917B CN112347917B CN202011225289.9A CN202011225289A CN112347917B CN 112347917 B CN112347917 B CN 112347917B CN 202011225289 A CN202011225289 A CN 202011225289A CN 112347917 B CN112347917 B CN 112347917B
- Authority
- CN
- China
- Prior art keywords
- network model
- vibration signal
- model
- gas turbine
- deep
- 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.)
- Active
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 71
- 238000000034 method Methods 0.000 title claims abstract description 55
- 238000003860 storage Methods 0.000 title claims abstract description 8
- 239000002245 particle Substances 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000005070 sampling Methods 0.000 claims abstract description 23
- 238000005457 optimization Methods 0.000 claims abstract description 17
- 230000000694 effects Effects 0.000 claims abstract description 10
- 238000012360 testing method Methods 0.000 claims description 28
- 230000006870 function Effects 0.000 claims description 26
- 238000004422 calculation algorithm Methods 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 238000010276 construction Methods 0.000 claims description 6
- 238000013144 data compression Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 230000001133 acceleration Effects 0.000 claims description 5
- 108010076504 Protein Sorting Signals Proteins 0.000 claims description 4
- 238000012952 Resampling Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000007906 compression Methods 0.000 claims description 4
- 230000006835 compression Effects 0.000 claims description 4
- 230000011218 segmentation Effects 0.000 claims description 4
- 230000005284 excitation Effects 0.000 abstract description 13
- 238000010586 diagram Methods 0.000 description 10
- 230000009467 reduction Effects 0.000 description 10
- 210000002569 neuron Anatomy 0.000 description 7
- 238000012706 support-vector machine Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 229920002430 Fibre-reinforced plastic Polymers 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 238000012300 Sequence Analysis Methods 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 239000011151 fibre-reinforced plastic Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012567 pattern recognition method Methods 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
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
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention relates to a gas turbine fault diagnosis method, a system, equipment and a storage medium, which comprise the following steps: constructing a deep confidence network model; compressing an original vibration signal of the gas turbine, and inputting a deep confidence network model; optimizing the structural parameters of the deep confidence network model, and searching the optimal deep confidence network model with the best diagnosis effect; and performing fault diagnosis on the gas turbine according to the optimal deep belief network model. The deep confidence network model optimized based on the peak hold down sampling method and the particle swarm optimization can reduce sample data, reduce model training time and realize optimizing of network structure parameters. Compared with other shallow networks, the deep belief network model obtained by training the original vibration signal as input has better diagnosis performance and classification capability for the airflow excitation faults of the rotor system of the gas turbine.
Description
Technical Field
The invention relates to the technical field of gas turbine fault diagnosis, in particular to a gas turbine fault diagnosis method, system, equipment and storage medium based on a deep confidence network of original information compression.
Background
The gas turbine is used as an advanced power device and is widely applied to ships, aviation, vehicles and power generation equipment. The air flow excitation is used as the fault of the gas turbine caused by the working medium, has burst and instability, seriously affects the stable working range and the operation reliability of the gas turbine, the gas compressor generates a rotating stall group, howling occurs, the vibration of the gas compressor is increased, and the fault frequency of the gas compressor is mostly represented as a low-frequency component. Because the working condition of the gas turbine is complex in the running process and is always in a variable speed non-steady state, the fault diagnosis of the gas turbine is difficult. At present, traditional diagnosis methods such as time domain sequence analysis, frequency domain analysis, time frequency analysis and the like are still adopted for airflow excitation faults, the traditional diagnosis methods do not have the intelligent and rapid diagnosis capabilities, and high professional accumulation is required for diagnosticians, so that the research on intelligent diagnosis algorithms of the gas turbine is very important.
The deep belief network (Deep Belief Network, DBN) is a typical deep neural network, and can directly take an original vibration signal as input to extract more abstract high-level features from low-level data, so that effective classification is realized, the deep belief network has stronger feature extraction and nonlinear mapping capabilities than those of shallow networks such as an artificial neural network, a support vector machine and the like, uncertainty in a feature extraction process is avoided, and intellectualization and automation of fault diagnosis are realized. Although the current methods based on deep belief networks are relatively more, the intelligent DBN diagnosis model aiming at the characteristics of the vibration signals of the gas turbine is relatively lacking, and when the original vibration signals are taken as DBN input, the model training time is relatively long, and the intelligent optimization algorithm of the structural parameters of the DBN network is relatively lacking.
Disclosure of Invention
In view of the above problems, the present invention aims to provide a method, a system, a device and a storage medium for diagnosing a gas turbine fault based on a deep belief network, which can reduce sample data, reduce model training time and realize optimization of network structure parameters; and has better diagnostic performance and classification capability.
In order to achieve the above purpose, the present invention adopts the following technical scheme: a peak hold down sampling deep belief network gas turbine fault diagnosis method comprising the steps of: 1) Constructing a deep confidence network model; 2) Compressing an original vibration signal of the gas turbine, and inputting a deep confidence network model; 3) Optimizing the structural parameters of the deep confidence network model, and searching the optimal deep confidence network model with the best diagnosis effect as an optimized self-adaptive fault diagnosis model; 4) And collecting vibration signals from the actual test bed or the gas turbine test point, performing extremely poor normalization processing, compressing the signals by adopting peak value downsampling, inputting the obtained characteristics into a self-adaptive fault diagnosis model, obtaining a fault diagnosis result, and completing fault diagnosis.
Furthermore, the peak value of each section is used as a resampling value to realize the compression of the original vibration signal and preserve the impact characteristics of the signal.
Further, the peak hold down sampling method includes the steps of:
2.1 Assuming that the acquired vibration signal is a vibration signal s (t) containing P points, sampling frequency f 0 Then:
s(t)＝[t 0 ,t 1 ,…,t p ]，0≤p≤P-1；
2.2 If the vibration signal S (T) is divided into M data segments, the vibration signal S (T) is converted into S (T), and the signal S (T) after the segmentation conversion is:
S(T)＝[S(T 0 ),S(T 1 ),…,S(T m )]，0≤m≤M-1
in the formula ,s(tp ) A vibration signal representing the p-th point, S (T m ) A signal representing an mth data segment;
2.3 Extraction sequence S (T) m ) The maximum value of the medium absolute value represents the sequence S (T m ) Rearranging the maximum value of the absolute values of the data segments to obtain a down-sampled signal sequence.
Further, a particle swarm optimization method is adopted to optimize the structural parameters of the deep belief network model.
Further, the position and velocity update formula of the particles is:
in the formula ,
and />
Respectively representing the speed and the position of the particles q after the kth iteration; w represents an inertial weight; c 1 、c 2 Is an acceleration factor; r is (r) 1 、r 2 Is a random number between (0, 1); p is p q Is the optimal position of the particle q; p (P) g Is the optimal position of the population.
Further, the structure of the deep confidence network model is composed of a plurality of layers of limited Boltzmann machines and a layer of BP network, and features are automatically extracted from an original vibration signal to realize effective mapping of data; the model training process consists of pre-training and fine tuning.
Further, the fine tuning is: let the input sample be x, the DBN network is composed of l RBM layers, and the output of the RBM of the last layer is u l (x) Wherein, DBN is a deep belief network model, RBM is a limited Boltzmann machine; the DBN network output layer takes the soft max function as an activation function, and then the DBN network output is as follows:
wherein ,
for the output of the last layer of the network, soft max is the activation function of the output layer;
the average square error is adopted as a loss function, a reverse error propagation algorithm of a minimum mean square error principle is adopted to update the parameters of the whole network, and the loss function is as follows:
where E is the average square error of the forward propagation,
for output of output layer X i For desired output, N is the number of samples.
A peak hold down sampling deep belief network gas turbine fault diagnostic system, comprising: the system comprises a model construction module, a data compression module, an optimizing module and a diagnosis module;
the model construction module constructs a deep belief network model;
the data compression module compresses an original vibration signal of the gas turbine and inputs the compressed original vibration signal into a deep confidence network model;
the optimizing module optimizes the structural parameters of the deep confidence network model and searches the optimal deep confidence network model with the best diagnosis effect;
the diagnosis module collects vibration signals from actual test tables or measuring points of the gas turbine, compresses the signals by adopting peak value downsampling after performing range normalization processing, and then inputs obtained characteristics into the self-adaptive fault diagnosis model to obtain fault diagnosis results so as to finish fault diagnosis.
A computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods described above.
A computing apparatus, comprising: one or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods described above.
Due to the adoption of the technical scheme, the invention has the following advantages: 1. the deep confidence network model optimized based on the peak hold down sampling method and the particle swarm optimization can reduce sample data, reduce model training time and realize optimizing of network structure parameters. 2. Compared with other shallow networks, the deep confidence network model obtained by training the original vibration signal as input has better diagnosis performance and classification capability for the airflow excitation faults of the rotor system of the gas turbine.
Drawings
FIG. 1 is a schematic diagram of a deep belief network fault diagnosis flow.
Fig. 2 is a schematic diagram of a deep belief network architecture.
FIG. 3 is a schematic diagram of a gas turbine rotor system.
Fig. 4 is a time domain diagram of a receiver vibration signal.
FIG. 5 is a time-frequency diagram of a case vibration signal.
FIG. 6a is a normal vibration signal for a gas turbine;
FIG. 6b is a gas turbine airflow excitation signal;
FIG. 6c is an imbalance signal for a gas turbine.
FIG. 7 is a schematic diagram of the result of the particle swarm optimization iteration.
FIG. 8 is a schematic diagram of the test results of the model.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. All other embodiments, which are obtained by a person skilled in the art based on the described embodiments of the invention, fall within the scope of protection of the invention.
The peak hold down sampling deep confidence network gas turbine fault diagnosis method provided by the invention is characterized in that a deep confidence network is introduced into the field of gas turbine airflow excitation fault diagnosis, an input data is compressed by using a peak hold down sampling method aiming at the problem that when an original vibration signal is input as the deep confidence network, the number of layers of a limited Boltzmann machine of a deep confidence network model and the number of neurons of each layer are difficult to determine, a particle swarm algorithm is adopted to optimize the structural parameters of the deep confidence network, a model with the best diagnosis effect is searched, and then the obtained optimal deep confidence model is used for carrying out fault diagnosis on a gas turbine rotor system. The present invention will be described in detail with reference to the following examples.
In a first embodiment of the present invention, as shown in FIG. 1, a peak hold down sampling depth confidence network gas turbine fault diagnosis method is provided, comprising the steps of:
1) Constructing a deep confidence network model;
as shown in fig. 2, the Deep Belief Network (DBN) model is composed of a multi-layer limited boltzmann machine (Restricted Boltzmann Machines, RBM) and a layer of BP network, which can automatically extract features from the original vibration signal, and realize effective mapping of data. The model training process consists of pre-training and fine tuning, and specifically comprises the following steps:
1.1 Pre-training
The DBN pre-training process is to use the training set to unsupervised train all the limited Boltzmann machines. The limited Boltzmann machine (RBM) is a main structure of a DBN model, the RBM has two layers in total, the upper layer is a hidden layer, the lower layer is a display layer, v and h respectively represent states of the display layer and the hidden layer, and v= (v) 1 ,v 2 …,v I )，h＝(h 1 ,h 2 …h J ) The RBM model function E (v, h|θ) is:
wherein I and J are the numbers of neurons in the display layer and the hidden layer, respectively, θ= { w ij ,a i ,b j The model parameter of the limited Boltzmann machine comprises a weight w between an ith display element of a display layer and a jth hidden element of a hidden layer ij Bias a of display element i i And bias b of hidden element j j ，v i Represents the state of the ith display layer, h j Representing the state of the j-th hidden layer.
Based on the RBM model function, a joint probability distribution p (v, h|theta) of (v, h) is obtained:
p(v,h|θ)＝e -E(v,h|θ) /Z(θ) (2)
wherein ：
is a normalization factor.
The deep confidence network obtains a theta parameter set by adopting a maximum likelihood method on a training set, and adopts a random gradient ascent method to carry out the method on theta= { w ij ,a i ,b j And (3) updating.
1.2 Fine tuning)
The fine tuning is to use a back propagation algorithm to fine tune and optimize DBN model parameters through sample labels on the basis of model parameters obtained through pre-training. Let the input sample be x, the DBN network is composed of l RBM layers, and the output of the RBM of the last layer is u l (x) A. The invention relates to a method for producing a fibre-reinforced plastic composite The DBN network output layer takes a softmax function as an activation function, and then the DBN network output is as follows:
wherein ,
for output of the last layer of the network, softmax is the activation function of the output layer.
The average square error is adopted as a loss function, a reverse error propagation algorithm of a minimum mean square error principle is adopted to update the parameters of the whole network, and the loss function is as follows:
where E is the average square error of the forward propagation,
for output of output layer X i To be output, W l 、b l The weight and bias of the first layer are respectively represented, and N represents the number of samples.
Factors that have the greatest impact on the performance of the deep belief network model include the number of layers of the constrained boltzmann machine and the number of neurons per layer. The number of network layers and the number of neurons of the limited Boltzmann machine are excessive, so that overfitting is easy to cause, and the model effect is reduced. Meanwhile, when the original data is used as the input of the deep confidence network, the sample length is longer, so that the model training time is increased.
2) Compressing an original vibration signal of the gas turbine, and inputting a deep confidence network model;
when the original data is used as the input of the deep belief network, if the data length is too long, the model training time is too long, so that the data is very important to be compressed on the premise of retaining enough state information. In the embodiment, the peak value maintaining and downsampling method (Peak Hold Down Sampling, PHDS) is adopted to segment the original vibration signal with longer period at equal intervals, and the peak value of each segment is used as a resampling value, so that the original vibration signal is compressed, the data quantity is reduced, and the impact characteristic of the signal is maintained obviously.
The peak hold down sampling method specifically comprises the following steps:
2.1 Assuming that the acquired vibration signal is a vibration signal s (t) containing P points, sampling frequency f 0 Then:
s(t)＝[s(t 0 ),s(t 1 ),...,s(t p )]，0≤p≤P-1 (5)
2.2 If the vibration signal S (T) is divided into M data segments, the vibration signal S (T) is converted into S (T), and the signal S (T) after the segmentation conversion is:
S(T)＝[S(T 0 ),S(T 1 ),…,S(T m )]，0≤m≤M-1 (6)
in the formula ,s(tp ) A vibration signal representing the p-th point, S (T m ) A signal representing an mth data segment; p is the length of the original signal, r is the downsampling rate, i.e. the data reduction rate, q=p/r. After peak down-sampling the data, the down-sampling frequency f d ＝f 0 In order to ensure that the loss of the compressed data state information is as small as possible, the downsampling frequency is chosen to be 2-8 in this embodiment.
2.3 Extraction sequence S (T) m ) The maximum value of the medium absolute value represents the sequence S (T m ) Rearranging the maximum value of the absolute values of the data segments to obtain a down-sampled signal sequence.
3) Optimizing the structural parameters of the deep confidence network model, and searching the optimal deep confidence network model with the best diagnosis effect;
the number of layers of the limited Boltzmann machine in the deep belief network model has great influence on the model performance by the number of neurons in each layer, and the model trained by the method is not an optimal model in the prior art when parameters of the model are determined.
In this embodiment, a particle swarm optimization method (Particle Swarm Optimization, PSO) is used to optimize structural parameters of the deep belief network model, and the particle swarm optimization method is an optimizing method, so that intelligent optimization of model parameters can be realized, and the position of each particle is a parameter corresponding to a population optimal solution. The group evolution is completed through the cooperation between individuals and groups, so that the optimal solution of the nonlinear problem is obtained, and the global intelligent search of parameters is realized. The position and velocity update formula of the particles is:
in the formula ,
and />
Respectively representing the speed and the position of the particles q after the kth iteration; w represents an inertial weight; c 1 、c 2 Is an acceleration factor; r is (r) 1 、r 2 Is a random number between (0, 1); p is p q Is the optimal position of the particle q; p (P) g Is the optimal position of the population.
Finally, an optimized self-adaptive fault diagnosis model is obtained.
4) And collecting vibration signals from the actual test bed or the gas turbine test point, performing extremely poor normalization processing, compressing the signals by adopting peak value downsampling, inputting the obtained characteristics into a self-adaptive fault diagnosis model, obtaining a fault diagnosis result, and completing fault diagnosis.
In a second embodiment of the present invention, a peak-hold downsampling depth belief network gas turbine fault diagnosis system is provided that includes a model building module, a data compression module, an optimizing module, and a diagnosis module;
the model construction module constructs a deep belief network model;
the data compression module compresses an original vibration signal of the gas turbine and inputs the compressed original vibration signal into the deep belief network model;
the optimizing module optimizes the structural parameters of the deep confidence network model and searches the optimal deep confidence network model with the best diagnosis effect;
the diagnosis module collects vibration signals from the actual test bed or the gas turbine measuring point, compresses the signals by adopting peak value downsampling after performing extremely poor normalization processing, and then inputs the obtained characteristics into the self-adaptive fault diagnosis model to obtain a fault diagnosis result so as to finish fault diagnosis.
In a third embodiment of the invention, there is provided a computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods of the first embodiment.
In a fourth embodiment of the present invention, there is provided a computing device including: one or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods of the first embodiment.
Examples:
taking the diagnosis of the airflow excitation fault of a certain gas turbine as an example, the following treatment is carried out:
(1) Failure analysis
The structural principle of a certain type of double-rotor gas turbine of a certain company is shown in fig. 3, a rotor system consists of a low-pressure rotor and a high-pressure rotor, and a compressor and a turbine are connected together through corresponding shafts to form the rotor. In order to collect vibration data, speed sensors are respectively arranged at the radial positions of the casings of the low-pressure compressor and the high-pressure compressor. During experiments, the rotating speed of the high-pressure rotor is between 6000r/min and 7800r/min, the sampling frequency is 6000Hz, and the experiment time lasts 8000s.
When the gas turbine rotor rises from the slow car, the phenomenon of amplitude increase occurs at the low-pressure compressor measuring point and the high-pressure compressor measuring point, the vibration overrun occurs at the low-pressure rotor measuring point, and when the rotating speed is continuously increased, the vibration amplitude is reduced instead. And (3) cutting out a vibration overrun signal section from the low-voltage rotor measuring point measured vibration data, analyzing the vibration overrun signal section, wherein the vibration overrun signal section is obtained in the speed increasing process of the slow vehicle, and fig. 4 and 5 are respectively a time domain diagram and a short-time Fourier transform time-frequency diagram of the vibration signal.
When the rotor of the compressor rotates to stall, a plurality of lower rotating speeds occur in the flow field of the compressorCausing the air flow to excite. These rotating air masses will excite about f= (0.3-0.5) f N The vibration frequency of the whole engine is unstable, the ratio of the vibration frequency to the power frequency is less than 1, wherein f N Is the rotational frequency of the rotor. The combined vibration figures 4 and 5 can find that the rotor stall phenomenon occurs in the 1 and 2 time periods, the vibration amplitude changes, and the frequency component also has the airflow excitation frequency.
(2) Fault diagnosis
Vibration data obtained by testing a certain type of gas turbine are divided into three states, namely a normal state, an airflow excitation state and an unbalanced state, and in order to ensure that sample data contains enough state information, the sample length of each type of state is set to be 1024 data points. The constructed data sample set is composed of 1500 training samples and 600 test samples, the sample set is shown in table 1, and vibration signal time domain diagrams of each state of the sample set are shown in fig. 6a to 6 c.
Table 1 data sample set
First, using the range normalization method x on the sample data out ＝(x in -x min )/(x max -x min ) And normalizing the data to 0-1, then reducing the sample data by using a peak hold down sampling method, testing the sample data with different reduction factors by using a DBN model based on a particle swarm optimization method, and testing the influence of the different reduction factors on the model diagnosis accuracy, wherein the result is shown in Table 2.
Table 2 peak hold down sample test results
From table 2, it can be seen that the effect of the peak hold down sampling method on model construction after sample reduction is performed, when the reduction factor is 2, the training time of the model is reduced, the model test accuracy and the loss function change are not large, when the reduction factor exceeds 8, the loss function and the test accuracy change of the model are obvious, which proves that most of the state information in the sample can still be maintained after data reduction, but when the reduction factor is too large, the state information of the sample is lost, and therefore, the reduction factor should be set between 2 and 8. In order to balance the training time and the test accuracy, the reduction multiple is finally determined to be 2, the model training time is reduced by 10s, and the length of the reduced sample data is 512 data points.
Loading compressed data samples to train a deep confidence network, taking the testing accuracy of a testing data set as an objective function, and setting parameters of a particle swarm optimization method: the number of particles was 40, the acceleration factor was 20, and the number of iterations was 20. Wherein the search range of the hidden layer number n is [2,4], and the search range of the neuron number of each layer is [0,1000]. The optimal network layer number and the sum hidden element number of the DBN model are searched for in an iterative mode by using a particle swarm optimization method, the iterative result is shown in fig. 7, the highest test accuracy of the found model is 99.8% about 18 times in iteration, and the corresponding network model parameters are as follows: the hidden layer number is two, and the number of neurons of the hidden layer is 100 and 100 respectively, so that the DBN network structure is set to be 512-100-100-3. Determining the DBN network structure parameters through the PSO method can avoid local optimization and uncertainty of parameter setting.
Setting the parameter learning rate to be 0.1 and the iteration number to be 50, training the DBN model, and testing the model performance by using a test sample, wherein the test result is shown in figure 8. Comparing the predictive label diagnosed by the model with the actual sample label, 2 samples in the test data set are misclassified, and the DBN network can achieve an accuracy of about 99.8%. This demonstrates that the deep belief network can extract features well from the original vibration signal of the gas turbine, enabling good recognition and diagnosis.
To verify the advantages of the DBN method for gas turbine fault diagnosis, the DBN-based fault diagnosis method is compared with other pattern recognition methods, different models are trained by the same sample set, the performance of the models is tested by the same test set, and the test results are shown in Table 3.
TABLE 3 failure recognition rates for different diagnostic models
The first method trains a deep belief network model with the raw vibration signal as input. The second method uses the original vibration data as input to train the BP network model with hidden layer number of 3 and iteration number of 50. The third method is a Support Vector Machine (SVM) model that trains a radial basis function with raw vibration data as input. The deep confidence network model has the best diagnosis performance for the gas turbine airflow excitation faults through comparison discovery, because the DBN network has better feature extraction and mapping capability than the BP network and the SVM, and the fault diagnosis capability of the intelligent fault diagnosis model of the gas turbine deep confidence network is proved.
In summary, because the airflow excitation fault is a common fault of the gas turbine caused by a working medium, the invention establishes a deep confidence network fault diagnosis model optimized by a peak hold down-sampling algorithm and a particle swarm algorithm aiming at the airflow excitation fault of a certain gas turbine. And reducing vibration data by using a peak hold down sampling method, using the vibration data as input of a deep confidence neural network, reducing model training time, optimizing the deep confidence network structural parameters by adopting a particle swarm algorithm, and searching network structural parameters corresponding to a deep confidence model with the best diagnosis performance. The example result shows that the optimized model not only reduces the training time of the model and realizes the intelligent optimization of network structural parameters, but also effectively realizes the airflow excitation fault diagnosis of the gas turbine, and the test accuracy is about 99.8%.
The foregoing embodiments are only illustrative of the present invention, and the structure, dimensions, placement and shape of the components may vary, and all modifications and equivalents of the individual components based on the teachings of the present invention should not be excluded from the scope of protection of the present invention.
Claims (4)
1. A gas turbine fault diagnosis method, comprising the steps of:
1) Constructing a deep confidence network model;
2) Compressing an original vibration signal of the gas turbine, and inputting a deep confidence network model;
3) Optimizing the structural parameters of the deep confidence network model, and searching the optimal deep confidence network model with the best diagnosis effect as an optimized self-adaptive fault diagnosis model;
4) Collecting vibration signals from actual test tables or gas turbine test points, performing extremely poor normalization processing, compressing the signals by adopting peak value downsampling, inputting obtained characteristics into a self-adaptive fault diagnosis model, obtaining a fault diagnosis result, and completing fault diagnosis;
the peak value of each section is used as a resampling value to realize the compression of the original vibration signal and preserve the impact characteristic of the signal;
the peak hold down sampling method comprises the following steps:
2.1 Assuming that the acquired vibration signal is a vibration signal s (t) containing P points, sampling frequency f 0 Then:
s(t)＝[t 0 ,t 1 ,…,t p ]，0≤p≤P-1；
2.2 If the vibration signal S (T) is divided into M data segments, the vibration signal S (T) is converted into S (T), and the signal S (T) after the segmentation conversion is:
S(T)＝[S(T 0 ),S(T 1 ),…,S(T m )]，0≤m≤M-1
in the formula ,s(tp ) A vibration signal representing the p-th point, S (T m ) A signal representing an mth data segment;
2.3 Extraction sequence S (T) m ) The maximum value of the medium absolute value represents the sequence S (T m ) Rearranging the maximum value of the absolute values of the numerical values in all the data segments to obtain a down-sampled signal sequence;
optimizing structural parameters of the deep belief network model by adopting a particle swarm optimization method;
the position and velocity update formula of the particles is:
in the formula ,
and />
Respectively representing the speed and the position of the particles q after the kth iteration; w represents an inertial weight; c 1 、c 2 Is an acceleration factor; r is (r) 1 、r 2 Is a random number between (0, 1); p is p q Is the optimal position of the particle q; p (P) g Is the optimal position of the population;
the structure of the deep confidence network model consists of a plurality of layers of limited Boltzmann machines and a layer of BP network, and features are automatically extracted from an original vibration signal to realize effective mapping of data; the model training process consists of pre-training and fine tuning;
the fine tuning is as follows: let the input sample be x, the DBN network is composed of l RBM layers, and the output of the RBM of the last layer is u l (x) Wherein, DBN is a deep belief network model, RBM is a limited Boltzmann machine; the DBN network output layer takes the soft max function as an activation function, and then the DBN network output is as follows:
wherein ,
for networksThe output of the last layer, softmax, is the activation function of the output layer;
the average square error is adopted as a loss function, a reverse error propagation algorithm of a minimum mean square error principle is adopted to update the parameters of the whole network, and the loss function is as follows:
where E is the average square error of the forward propagation,
for output of output layer X i For the expected output, N is the number of samples; w (W) l Weight of layer I, b l Indicating the bias of the first layer.
2. A gas turbine fault diagnosis system, comprising: the system comprises a model construction module, a data compression module, an optimizing module and a diagnosis module;
the model construction module constructs a deep belief network model;
the data compression module compresses an original vibration signal of the gas turbine and inputs the compressed original vibration signal into a deep confidence network model;
the optimizing module optimizes the structural parameters of the deep confidence network model and searches the optimal deep confidence network model with the best diagnosis effect;
the diagnosis module collects vibration signals from an actual test bed or a gas turbine measuring point, compresses the signals by adopting peak value downsampling after performing extremely poor normalization processing, and then inputs the obtained characteristics into the self-adaptive fault diagnosis model to obtain a fault diagnosis result so as to finish fault diagnosis;
the peak value of each section is used as a resampling value to realize the compression of the original vibration signal and preserve the impact characteristic of the signal;
the peak hold down sampling method comprises the following steps:
2.1 Assuming that the acquired vibration signal is a vibration signal s (t) containing P points, sampling frequency f 0 Then:
s(t)＝[t 0 ,t 1 ,…,t p ]，0≤p≤P-1；
2.2 If the vibration signal S (T) is divided into M data segments, the vibration signal S (T) is converted into S (T), and the signal S (T) after the segmentation conversion is:
S(T)＝[S(T 0 ),S(T 1 ),…,S(T m )]，0≤m≤M-1
in the formula ,s(tp ) A vibration signal representing the p-th point, S (T m ) A signal representing an mth data segment;
2.3 Extraction sequence S (T) m ) The maximum value of the medium absolute value represents the sequence S (T m ) Rearranging the maximum value of the absolute values of the numerical values in all the data segments to obtain a down-sampled signal sequence;
optimizing structural parameters of the deep belief network model by adopting a particle swarm optimization method;
the position and velocity update formula of the particles is:
in the formula ,
and />
Respectively representing the speed and the position of the particles q after the kth iteration; w represents an inertial weight; c 1 、c 2 Is an acceleration factor; r is (r) 1 、r 2 Is a random number between (0, 1); p is p q Is the optimal position of the particle q; p (P) g Is the optimal position of the population;
the structure of the deep confidence network model consists of a plurality of layers of limited Boltzmann machines and a layer of BP network, and features are automatically extracted from an original vibration signal to realize effective mapping of data; the model training process consists of pre-training and fine tuning;
the fine tuning is as follows: let the input sample be x, the DBN network is composed of l RBM layers, and the output of the RBM of the last layer is u l (x) Wherein, DBN is a deep belief network model, RBM is a limited Boltzmann machine; the DBN network output layer takes the soft max function as an activation function, and then the DBN network output is as follows:
wherein ,
for the output of the last layer of the network, softmax is the activation function of the output layer;
the average square error is adopted as a loss function, a reverse error propagation algorithm of a minimum mean square error principle is adopted to update the parameters of the whole network, and the loss function is as follows:
where E is the average square error of the forward propagation,
for output of output layer X i For the expected output, N is the number of samples; w (W) l Weight of layer I, b l Indicating the bias of the first layer.
3. A computer readable storage medium storing one or more programs, wherein the one or more programs comprise instructions, which when executed by a computing device, cause the computing device to perform the method of claim 1.
4. A computing device, comprising: one or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing the method of claim 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011225289.9A CN112347917B (en) | 2020-11-05 | 2020-11-05 | Gas turbine fault diagnosis method, system, equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011225289.9A CN112347917B (en) | 2020-11-05 | 2020-11-05 | Gas turbine fault diagnosis method, system, equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112347917A CN112347917A (en) | 2021-02-09 |
| CN112347917B true CN112347917B (en) | 2023-05-16 |
Family
ID=74429554
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011225289.9A Active CN112347917B (en) | 2020-11-05 | 2020-11-05 | Gas turbine fault diagnosis method, system, equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112347917B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113835412A (en) * | 2021-09-13 | 2021-12-24 | 上海工业自动化仪表研究院有限公司 | Gas turbine control system sensor fault diagnosis method based on data driving |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3255837A1 (en) * | 2016-06-09 | 2017-12-13 | Alcatel Lucent | Method and system for automating troubleshooting of a network |
| CN108376286A (en) * | 2018-02-26 | 2018-08-07 | 上海理工大学 | A kind of equipment fault diagnosis apparatus and system |
| CN109655259A (en) * | 2018-11-23 | 2019-04-19 | 华南理工大学 | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks |
| CN109933881A (en) * | 2019-03-06 | 2019-06-25 | 武汉大学 | A kind of Fault Diagnosis of Power Electronic Circuits method based on optimization deepness belief network |
| CN110232402A (en) * | 2019-05-14 | 2019-09-13 | 沈阳化工大学 | A kind of adaptive depth confidence network Fault Diagnosis of Roller Bearings of optimization |
| CN110232435A (en) * | 2019-04-30 | 2019-09-13 | 沈阳化工大学 | A kind of adaptive depth confidence network Fault Diagnosis of Roller Bearings |
| CN110297480A (en) * | 2019-06-17 | 2019-10-01 | 重庆仲澜科技有限公司 | The TE process failure diagnosis method of deepness belief network model based on parameter optimization |
| CN110852017A (en) * | 2019-10-08 | 2020-02-28 | 湖南省计量检测研究院 | Hydrogen fuel cell fault diagnosis method based on particle swarm optimization and supporting vector machine |
| CN111175054A (en) * | 2020-01-08 | 2020-05-19 | 沈阳航空航天大学 | Aeroengine fault diagnosis method based on data driving |
-
2020
- 2020-11-05 CN CN202011225289.9A patent/CN112347917B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3255837A1 (en) * | 2016-06-09 | 2017-12-13 | Alcatel Lucent | Method and system for automating troubleshooting of a network |
| CN108376286A (en) * | 2018-02-26 | 2018-08-07 | 上海理工大学 | A kind of equipment fault diagnosis apparatus and system |
| CN109655259A (en) * | 2018-11-23 | 2019-04-19 | 华南理工大学 | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks |
| CN109933881A (en) * | 2019-03-06 | 2019-06-25 | 武汉大学 | A kind of Fault Diagnosis of Power Electronic Circuits method based on optimization deepness belief network |
| CN110232435A (en) * | 2019-04-30 | 2019-09-13 | 沈阳化工大学 | A kind of adaptive depth confidence network Fault Diagnosis of Roller Bearings |
| CN110232402A (en) * | 2019-05-14 | 2019-09-13 | 沈阳化工大学 | A kind of adaptive depth confidence network Fault Diagnosis of Roller Bearings of optimization |
| CN110297480A (en) * | 2019-06-17 | 2019-10-01 | 重庆仲澜科技有限公司 | The TE process failure diagnosis method of deepness belief network model based on parameter optimization |
| CN110852017A (en) * | 2019-10-08 | 2020-02-28 | 湖南省计量检测研究院 | Hydrogen fuel cell fault diagnosis method based on particle swarm optimization and supporting vector machine |
| CN111175054A (en) * | 2020-01-08 | 2020-05-19 | 沈阳航空航天大学 | Aeroengine fault diagnosis method based on data driving |
Non-Patent Citations (6)
| Title |
|---|
| 刘秀丽 ; 徐小力 ; .基于深度信念网络的风电机组齿轮箱故障诊断方法.可再生能源.2019,(第12期),全文. * |
| 王应晨 ; 段修生 ; .深度学习融合模型在机械故障诊断中的应用.振动.测试与诊断.2019,(第06期),全文. * |
| 王永杰 ; .基于深度置信网络的故障诊断方法.电子技术与软件工程.2018,(第12期),全文. * |
| 石宪 ; 钱玉良 ; .基于改进深度信念网络的燃气轮机故障诊断.上海电力大学学报.2020,(第02期),全文. * |
| 闫丽萍 ; 董学智 ; 张永军 ; 王涛 ; 高庆 ; 谭春青 ; .基于深度置信网络的燃气轮机气路故障诊断方法.工程热物理学报.2020,(第04期),全文. * |
| 陈自强 ; 程健 ; 季文强 ; 翟红雨 ; .基于深度置信网络风电机组变桨系统的故障诊断.测控技术.2017,(第05期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112347917A (en) | 2021-02-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Qin et al. | The optimized deep belief networks with improved logistic sigmoid units and their application in fault diagnosis for planetary gearboxes of wind turbines | |
| CN112161784B (en) | Mechanical fault diagnosis method based on multi-sensor information fusion migration network | |
| CN112418013B (en) | Complex working condition bearing fault diagnosis method based on meta-learning under small sample | |
| CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
| CN112163640A (en) | Planetary gearbox fault diagnosis method and system based on deep convolution migration learning | |
| CN112101085A (en) | Adaptive intelligent fault diagnosis method based on importance weighted domain impedance | |
| CN113375941A (en) | Open set fault diagnosis method for high-speed motor train unit bearing | |
| CN115758212A (en) | Mechanical equipment fault diagnosis method based on parallel network and transfer learning | |
| CN111753891A (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
| CN114048688A (en) | Method for predicting service life of bearing of wind power generator | |
| CN113095179A (en) | Metametric learning driven helicopter planetary gearbox fault diagnosis method | |
| CN113268833A (en) | Migration fault diagnosis method based on deep joint distribution alignment | |
| CN112504682A (en) | Chassis engine fault diagnosis method and system based on particle swarm optimization algorithm | |
| CN112364706A (en) | Small sample bearing fault diagnosis method based on class imbalance | |
| CN112347917B (en) | Gas turbine fault diagnosis method, system, equipment and storage medium | |
| CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
| CN112651426A (en) | Fault diagnosis method for rolling bearing of wind turbine generator | |
| CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| Xue et al. | Rolling bearing fault diagnosis method based on self-calibrated coordinate attention mechanism and multi-scale convolutional neural network under small samples | |
| CN115859077A (en) | Multi-feature fusion motor small sample fault diagnosis method under variable working conditions | |
| CN114386452B (en) | Nuclear power circulating water pump sun gear fault detection method | |
| CN114330881A (en) | Data-driven fan blade icing prediction method and device | |
| Qi et al. | Feature classification method of frequency cepstrum coefficient based on weighted extreme gradient boosting | |
| CN112146882B (en) | Bearing fault diagnosis method based on transfer learning vibration signal image recognition |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |