A Fault Diagnosis Method for Gas Turbine Electric Actuator

Through the fault diagnosis method combined with Gaussian hybrid model and one-dimensional convolutional neural network, the problem of insufficient feature extraction in fault diagnosis of electric actuators of gas turbines is solved, and more efficient and accurate fault identification is achieved.

CN114154254BInactive Publication Date: 2024-11-26NORTH CHINA ELECTRIC POWER UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111001163.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-08-30
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing fault diagnosis methods for electric actuators of gas turbines rely on signal processing technology and expert experience, and it is difficult to intelligently extract fault signals from massive data, resulting in a gradual fault hiddenness. Failure to detect it in time may lead to a vicious cycle and threaten the life of system equipment.

Method used

The Gaussian hybrid model (GMM) is used to initially cluster the traffic signals collected by the sensor, combine the one-dimensional convolutional neural network (1DCNN) model for self-learning, and build the GMM-1DCNN fault diagnosis model, and achieve faults through preprocessing and fine-tuning of network weights. Feature extraction and classification.

Benefits of technology

It improves the efficiency and accuracy of fault identification, significantly improves the accuracy of fault diagnosis of electric actuators of gas turbines, and solves the problem of insufficient feature extraction in the existing methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114154254B_ABST
    Figure CN114154254B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for fault diagnosis of a gas turbine electric actuator, which belongs to the technical field of gas turbine fault diagnosis. The method comprises step 1: pre-processing the collected flow signal samples in a standardized manner; step 2: identifying the working conditions and performing preliminary clustering on the data set in step 1, and then training the function model of the Gaussian probability density; step 3: constructing a self-learning network model, training the self-learning network model, extracting the data features of the flow signal samples and saving the network weights of the self-learning network model; step 4: constructing a state classification and identification network model of the electric actuator, and loading the network weight parameters; step 5: fine-tuning the network weights of the GMM‑1DCNN of the electric actuator. The present invention has strong advantages in diagnostic accuracy and diagnostic efficiency, and avoids the drawbacks of large data processing workload and difficulty in extracting effective features caused by directly processing data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of gas turbine fault diagnosis, and in particular to a gas turbine electric actuator fault diagnosis method. Background Art

[0002] Gas turbine is a highly efficient and complex power mechanical device. As an internal combustion engine, it mainly generates heat energy through the combustion of air-fuel mixture, and the turbine converts the heat energy into mechanical energy to generate power. Gas turbine has a wide range of uses. It is used in many fields such as ship power, aerospace, energy generation, vehicle manufacturing, etc., and has become the core of industrial technology in the 21st century.

[0003] As a complex system engineering with high nonlinear integration, the actuator in the gas turbine control system is the key functional system to ensure its stable performance and reliable operation.

[0004] At present, common faults of electric actuators include progressive faults, minor faults, sudden faults, etc. Most mechanical equipment faults are progressive faults. Due to their hidden nature, these faults pose a greater threat to the system. If not discovered in time, they will lead to a vicious cycle and endanger the life of the system equipment. The traditional method for progressive fault diagnosis is as follows: acquisition of fault signals, data preprocessing, fault feature extraction, feature selection and optimization, and fault classification. The existing fault feature extraction method relies too much on signal processing technology and expert experience. Therefore, how to intelligently extract the features of fault signals from massive data is the key to the development of fault diagnosis technology. Summary of the invention

[0005] The object of the present invention is to provide a method for diagnosing a fault of a gas turbine electric actuator, which is characterized by comprising the following steps:

[0006] Step 1: Use sensors to collect flow signal samples under three operating conditions including variable load, variable speed and variable opening as data sets, and perform standardized preprocessing on the collected flow signal samples;

[0007] Step 2: Use the Gaussian mixture model, i.e. GMM, to identify the working condition of the data set in step 1, perform preliminary clustering on the data, obtain the function model of Gaussian probability density, and then use the expectation maximization algorithm to train the function model of Gaussian probability density to obtain the final parameters of GMM;

[0008] Step 3: A self-learning network model based on a one-dimensional convolutional neural network model, i.e., 1DCNN, is constructed in each sub-operating interval after GMM processing. The self-learning network model is trained using training samples in each sub-operating interval, the data features of the flow signal samples are extracted, and the network weights of the self-learning network model are saved.

[0009] Step 4: Construct a state classification and recognition network model of the electric actuator combining 1DCNN and Softmax classifier, namely GMM-1DCNN, and load the network weight parameters saved in step 3;

[0010] Step 5: Use the preprocessed real flow signal samples and the state label corresponding to each flow signal sample to fine-tune the network weights of the GMM-1DCNN of the electric actuator to achieve fault diagnosis of the electric actuator under different working conditions.

[0011] In the step 1, the number of flow signal samples under each operating condition is 500, including 300 training samples and 200 test samples; there are 10 fault modes detected under each operating condition, including power failure, motor failure, travel transmission gear damage, severe wear of worm gear, hardened and dry packing or over-tightening of packing gland, external leakage, internal leakage, valve blockage, deposits on valve or valve seat and valve stem bending.

[0012] The expectation maximization algorithm in step 2 includes the following steps:

[0013] Step 2.1: Obtain the estimated parameters (α u , μ u ,Σ u ) initial value; where α u is the mixing coefficient, μ u is the mean of Gaussian distribution, Σ u is the Gaussian distribution covariance matrix;

[0014] Step 2.2: Calculate for each sample x v The posterior probability that (v=1,2,3…,m) belongs to the u-th Gaussian model

[0015]

[0016] In the formula, Z v To generate x v Gaussian distribution, P(X v |μ u ,Σ u ) is the probability density function;

[0017] Step 2.3: Calculate the expected equation of the likelihood function of the sample data to reach the maximum value and obtain the GMM parameters for the next iteration;

[0018] Step 2.4: Repeat steps 2.2 and 2.3 until the algorithm converges to obtain the final parameters of GMM; otherwise, repeat steps 2.2 and 2.3.

[0019] When the initial error of training the self-learning network model in step 3 is large, a high fixed learning rate is used; after the model error decreases, the learning rate adopts a natural exponential decay method of the reference error; wherein the natural exponential decay method of the reference error is as follows:

[0020]

[0021] In the formula, α is the exponentially decaying learning rate; lr is the initial learning rate; dr is the decay rate of the learning rate; step is the current number of iterations; 1 epoch is equal to training once with all the samples in the training set.

[0022] In step 3, the 1DCNN includes a convolution layer, a pooling layer and a fully connected layer; wherein the convolution layer and the pooling layer appear alternately to extract fault features; the convolution kernel in the convolution layer is convolved with the input data to extract the features of the input data; the pooling layer is arranged after the convolution layer, and the pooling operation is to obtain the maximum value, minimum value or average value within the sub-operating condition interval.

[0023] The cost function J(θ) of the Softmax classifier in step 4 is:

[0024]

[0025] In the formula, x=(x1,x2,...,x m ) is the input, and its label is j=(j1,j2,...,j m ), j i ∈{1,2,...,k}; θ represents the parameters of the Softmax classifier model; l represents the starting point of the normalization operation, and m represents the range of the defined variable. Describes the normalization operation; n represents the dimension of the feature map; 1{y i =j} is an indicator function, whose value is related to the truth of the internal expression. If the internal expression is true, it is 1, otherwise it is 0; λ is the weight attenuation coefficient.

[0026] The GMM-1DCNN model in step 5 includes two establishment processes: identifying the working condition of the data based on GMM and clustering the collected data; establishing a sub-working condition 1DCNN model based on the sub-working condition interval data set, reconstructing the input data, and completing model training.

[0027] The beneficial effects of the present invention are:

[0028] 1. The present invention firstly uses GMM to reasonably reconstruct the electric actuator operation data collected by the sensor, reduces the time step of the data and increases the feature quantity contained in each time step, divides the mixed working condition into sub-working conditions with different opening degrees, and improves the ability of adaptive feature extraction and the efficiency of the training process;

[0029] 2. Aiming at the one-dimensional characteristics of traffic signals, the present invention improves the traditional two-dimensional convolutional neural network into a 1DCNN suitable for processing one-dimensional signals, ensuring no information loss when extracting fault features;

[0030] 3. The GMM-1DCNN fault diagnosis model established by the present invention first realizes the initial operating condition division, and then uses a one-dimensional convolutional neural network to complete model training and testing in each sub-operating condition; it improves the fault identification efficiency in terms of time and accuracy, solves the problem of insufficient feature extraction of existing mature algorithms, and significantly improves the accuracy of gas turbine electric actuator fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a method for diagnosing faults of a gas turbine electric actuator provided by the present invention;

[0032] Figure 2 It is the GMM working condition classification diagram provided by the present invention;

[0033] Figure 3 is a diagram of training error and learning rate variation of the fault diagnosis model provided by the present invention;

[0034] Figure 4 It is a ten-fold cross validation diagram of the fault diagnosis model provided by the present invention;

[0035] Figure 5 It is a comparison chart of the diagnosis results of the GMM-1DCNN fault diagnosis method provided by the present invention and the other four methods. DETAILED DESCRIPTION

[0036] The present invention provides a method for diagnosing faults of an electric actuator of a gas turbine, which will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0037] Figure 1 FIG. 1 is a flow chart of a method for diagnosing a fault of a gas turbine electric actuator provided by the present invention. Figure 1 As shown, the gas turbine electric actuator fault diagnosis method includes the following steps:

[0038] Step 1: Use sensors to collect flow signal samples under various operating conditions (variable load, variable speed, variable opening) as data sets. The number of flow signal samples under each operating state is 500, including 300 training samples and 200 test samples. A total of 10 fault modes are detected, and the obtained flow signal samples are standardized and preprocessed;

[0039] Step 2: Use the Gaussian mixture model (GMM) to identify the working conditions of all data, perform preliminary clustering on the data, obtain the function model of Gaussian probability density, and use the expectation maximization algorithm for training to obtain the final parameters of GMM;

[0040] Step 3: Construct a self-learning network model based on a one-dimensional convolutional neural network model (1DCNN) in each sub-operating interval after GMM processing, and use the training samples in each sub-operating interval to train the network model. When the error is large in the early stage of model training, a higher fixed learning rate is used; after the model error decreases, the learning rate adopts the natural exponential decay method of the reference error to extract the data features of the flow signal samples and save the network weights of the self-learning network model;

[0041] Step 4: Construct an electric actuator state classification and recognition network model (GMM-1DCNN) combining 1DCNN and Softmax classifier, and load the network weights and other parameters saved in step 3;

[0042] Step 5: Use the preprocessed real flow signal samples and the state label corresponding to each flow signal sample to fine-tune the network weights of the GMM-1DCNN model of the electric actuator. The electric actuator state classification and recognition network model after weight fine-tuning can realize electric actuator fault diagnosis under different working conditions.

[0043] Failure modes in step 1 include power failure, motor failure, damaged travel gears, severely worn worm gears, hardened or dried packing, overtightened packing gland, external leakage, internal leakage, blocked valve, deposits on valve or seat, and bent valve stem.

[0044] The expectation maximization algorithm in step 2 is as follows:

[0045] Step 2.1: Initialization: Obtain the estimated parameters (α u , μ u ,Σ u ), α u is the mixing coefficient, μ u is the mean of Gaussian distribution, Σ u is the Gaussian distribution covariance matrix;

[0046] Step 2.2: Expectation step: Calculate for each sample x v The posterior probability P(Z) that (v=1,2,3…,m) belongs to the u-th Gaussian model v =u|X v ):

[0047]

[0048] In the formula, Z v To generate x v Gaussian distribution, P(X v |μ u ,Σ u ) is the probability density function;

[0049] Step 2.3: Maximum step: Calculate the expected equation of the likelihood function of the sample data to reach the maximum value and obtain the GMM parameters for the next iteration;

[0050] Step 2.4: Stop updating condition: Repeat the expected step and maximum step until the algorithm converges to obtain the final parameters of GMM. Otherwise, repeat steps 2.2 and 2.3.

[0051] Figure 2 The GMM working condition classification diagram provided by the present invention is shown in Figure 2. The working condition of the electric actuator can be divided into three sub-working conditions. Figure 2 shown.

[0052] In step 3, 1DCNN is mainly composed of convolutional layers, pooling layers, and fully connected layers. Convolutional layers and pooling layers appear alternately to extract fault features. In the convolutional layer, the convolution kernel performs convolution operations with the input data to extract the features of the input data, which is called a feature map. The pooling layer follows the convolutional layer, and its basic operation is the same as the convolution operation, which is also the sliding of the pooling kernel on the input data. Different from the convolution operation, the pooling operation takes the maximum value, minimum value, or average value in the area.

[0053] The natural exponential decay of the reference error in step 3 is as follows:

[0054]

[0055] In the formula, α is the exponentially decaying learning rate; lr is the initial learning rate; dr is the decay rate of the learning rate; step is the current number of iterations; 1 epoch is equal to training once with all the samples in the training set.

[0056] Figure 3 The diagram is a diagram showing the training error and learning rate variation of the fault diagnosis model provided by the present invention. The rapid convergence of the training error and learning rate demonstrates the correctness of the method described in the present invention.

[0057] The cost function J(θ) of the Softmax classifier model in step 4 is:

[0058]

[0059] In the formula, x=(x1,x2,...,x m ) is the input, and its label is j=(j1,j2,...,j m ), j i ∈{1,2,...,k}; θ represents the parameters of the Softmax classifier model; Describes the normalization operation; n represents the dimension of the feature map; 1{y i =j} is an indicator function, whose value is related to the truth of the internal expression. If the internal expression is true, it is 1, otherwise it is 0; λ is the weight attenuation coefficient.

[0060] The GMM-1DCNN model in step 5 includes two establishment processes: identifying the working condition of the data based on GMM and clustering the collected data; establishing the sub-working condition 1DCNN model based on the sub-working condition data set, reconstructing the input data, and completing the model training.

[0061] Figure 4 This is a ten-fold cross validation diagram of the fault diagnosis model provided by the present invention. The performance of the GMM-1DCNN fault diagnosis model can be verified by ten-fold cross validation. Due to the large number of data set samples, the present invention uses the test set data for verification. The results of the ten-fold cross validation of the GMM-1DCNN fault diagnosis model are shown in Figure 1. Figure 4 shown.

[0062] Figure 5 This is a comparison chart of the diagnostic results of the GMM-1DCNN fault diagnosis method provided by the present invention and the other four methods. It can be seen from the figure that the diagnostic efficiency of GMM-1DCNN is higher than that of the other four methods, and has a higher diagnostic accuracy.

[0063] This embodiment is only a preferred specific implementation of the present invention, but 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 diagnosing faults of a gas turbine electric actuator, characterized in that: The following steps are involved: Step 1: Use sensors to collect flow signal samples under three operating conditions including variable load, variable speed and variable opening as data sets, and perform standardized preprocessing on the collected flow signal samples; Step 2: Use the Gaussian mixture model GMM to identify the working conditions of the data set in step 1, perform preliminary clustering on the data, obtain the function model of the Gaussian probability density, and then use the expectation maximization algorithm to train the function model of the Gaussian probability density to obtain the final parameters of the GMM; The expectation maximization algorithm in step 2 includes the following steps: Step 2.1: Obtain the estimated parameters (α u , μ u ,Σ u ) initial value; where α u is the mixing coefficient, μ u is the mean of Gaussian distribution, Σ u is the Gaussian distribution covariance matrix; Step 2.2: Calculate for each sample x v The posterior probability that (v=1,2,3…,m) belongs to the u-th Gaussian model In the formula, Z v To generate x v Gaussian distribution, P(X v |μ u ,Σ u ) is the probability density function; Step 2.3: Calculate the expected equation of the likelihood function of the sample data to reach the maximum value and obtain the GMM parameters for the next iteration; Step 2.4: Repeat steps 2.2 and 2.3 until the algorithm converges to obtain the final parameters of GMM; otherwise, repeat steps 2.2 and 2.3; Step 3: Construct a self-learning network model based on the one-dimensional convolutional neural network model 1DCNN in each sub-operating interval after GMM processing, use the training samples in each sub-operating interval to train the self-learning network model, extract the data features of the flow signal samples and save the network weights of the self-learning network model; Step 4: Construct the state classification and recognition network model GMM-1DCNN of the electric actuator combining 1DCNN and Softmax classifier, and load the network weight parameters saved in step 3; The cost function J(θ) of the Softmax classifier in step 4 is: In the formula, x=(x1,x2,...,x m ) is the input, and its label is j=(j1,j2,...,j m ), j i ∈{1,2,...,k}; θ represents the parameters of the Softmax classifier model; l represents the starting point of the normalization operation, and m represents the range of the defined variable. Describes the normalization operation; n represents the dimension of the feature map; 1{y i =j} is an indicative function, whose value is related to the truth of the internal expression. If the internal expression is true, it is 1, otherwise it is 0; λ is the weight attenuation coefficient; Step 5: Use the preprocessed real flow signal samples and the state label corresponding to each flow signal sample to fine-tune the network weights of the GMM-1DCNN of the electric actuator to achieve fault diagnosis of the electric actuator under different working conditions.

2. The method for diagnosing faults of a gas turbine electric actuator according to claim 1, characterized in that: In the step 1, the number of flow signal samples under each operating condition is 500, including 300 training samples and 200 test samples; there are 10 fault modes detected under each operating condition, including power failure, motor failure, travel transmission gear damage, severe wear of worm gear, hardened and dry packing or over-tightening of packing gland, external leakage, internal leakage, valve blockage, deposits on valve or valve seat and valve stem bending.

3. The method for diagnosing faults of a gas turbine electric actuator according to claim 1, characterized in that: When the initial error of training the self-learning network model in step 3 is large, a high fixed learning rate is used; after the model error decreases, the learning rate adopts a natural exponential decay method of the reference error; wherein the natural exponential decay method of the reference error is as follows: In the formula, α is the exponentially decaying learning rate; lr is the initial learning rate; dr is the decay rate of the learning rate; step is the current number of iterations; 1 epoch is equal to training once with all the samples in the training set.

4. The method for diagnosing faults of a gas turbine electric actuator according to claim 1, characterized in that: In step 3, the 1DCNN includes a convolution layer, a pooling layer and a fully connected layer; wherein the convolution layer and the pooling layer appear alternately to extract fault features; the convolution kernel in the convolution layer is convolved with the input data to extract the features of the input data; the pooling layer is arranged after the convolution layer, and the pooling operation is to obtain the maximum value, minimum value or average value within the sub-operating condition interval.

5. The method for diagnosing faults of a gas turbine electric actuator according to claim 1, characterized in that: The GMM-1DCNN model in step 5 includes two establishment processes: identifying the working condition of the data based on GMM and clustering the collected data; establishing a sub-working condition 1DCNN model based on the sub-working condition interval data set, reconstructing the input data, and completing model training.