Steam turbine vibration fault diagnosis method and system

Abstract

The invention discloses a method for diagnosing vibration faults of steam turbines, and relates to the technical field of vibration fault diagnosis of steam turbines. The method performs denoising, diffusion and amplification on vibration data of normal operation of the steam turbine and vibration data of various faults to obtain a denoising diffusion model, a fault diagnosis model and a seed library. Then, samples of actual faults are processed by the denoising diffusion model, the fault diagnosis model and the seed library to obtain a fault type distribution P(A|B) and a probability distribution P(B) of the noise level under the current operating state. Then, the fault diagnosis result under the current operating condition is obtained by calculation. The invention also discloses a system for diagnosing vibration faults of steam turbines, which comprises a sensor, a computing server and an industrial computer. By executing the above process, more fault samples can be generated to improve the diagnosis effect, and noise interference can be avoided to improve the diagnosis stability.

Description

Technical Field

[0001] The invention relates to the technical field of steam turbine vibration fault diagnosis, and in particular to a steam turbine vibration fault diagnosis method and a system thereof. Background Art

[0002] With the rapid growth of residential electricity consumption, the installed capacity of thermal power plants has also been increasing. As an important equipment in thermal power plants, steam turbines are crucial to ensure the safe operation of steam turbine units. Bearings are key components of steam turbines, so fault diagnosis of bearings is very important.

[0003] In recent years, the mechanical equipment of steam turbine units has been developing in a more complex direction, and conventional fault diagnosis methods are difficult to meet. With the development of computer technology and sensor technology, data-driven intelligent diagnosis has become possible. At present, data-driven intelligent diagnosis methods still have the disadvantages of high false alarm rate and poor versatility. Therefore, solving the data problem in vibration fault diagnosis has become a top priority to improve the accuracy of intelligent diagnosis, and data enhancement is the main solution to this problem.

[0004] Generally speaking, the method to solve the data problem in vibration fault diagnosis is as follows: first, use rotor dynamics to generate simulation data from a mechanism perspective, but the data generated by this type of method has poor generalization and is easily affected by noise; then use a generative adversarial network (GAN) for data enhancement, which is very easily affected by the noise of samples in the original data set. At the same time, the algorithm has poor stability and is difficult to implement in actual industrial scenarios with more noise.

[0005] In summary, in the current diagnosis of turbine bearing vibration faults, the existing methods are easily affected by the difficulty in collecting fault samples and the insufficient number of samples, resulting in poor results. At the same time, the existing data enhancement methods are easily affected by noise and have poor stability. Summary of the invention

[0006] The purpose of the present invention is to provide a turbine vibration fault diagnosis method and system thereof to solve the problems existing in the above-mentioned prior art, to generate more fault samples to improve the diagnosis effect, to avoid noise interference, and to improve the diagnosis stability.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] The present invention provides a method for diagnosing vibration faults of a steam turbine, comprising the following steps:

[0009] S1. Collect the normal operation vibration data and fault vibration data of the steam turbine to obtain the data set X = {X 0 ,X 1 ,X 2 …X n-1}, where X 0 For normal operation vibration data, X 1 ,X 2 ...X n-1 is the fault vibration data of n-1 different fault types;

[0010] S2. For each type of data in the data set X obtained in step S1, a denoising diffusion model is trained. Through training, n denoising diffusion models can be obtained, where the model set is D = {D 0 ,D 1 ,D 2 …D n-1};

[0011] S3. Use the denoising diffusion model set D obtained in step S2 to expand the data set X one by one. Each model D obtains m data sets with different noise levels. Where i = 0, 1, 2...n-1, j = 1, 2...m;

[0012] S4. For the data set obtained in step S3 m noise level data are used to train m vibration fault diagnosis models. Taking noise level 1 as an example, the training data set is {X 0,1 ,X 1,1 ,X 2,1 …X n-1,1}, the trained model is M 1 , thus obtaining the fault diagnosis model set M = {M j}, where j = 1, 2...m, and at the same time, we get m sets of hidden variables H = H under different noise levels i,j , a seed library is generated through the hidden variable set H of the normal running samples in the set;

[0013] S5. Run the steam turbine, read the turbine vibration sample every 1 second, use the m fault diagnosis models obtained in step S4 to perform fault diagnosis on the data, and obtain the fault type distribution P(A|B) of the current steam turbine bearing under different noise levels, where B is a random variable representing the noise level and A is a random variable representing the fault type;

[0014] S6. Run the steam turbine and read the turbine vibration sample every 60 seconds to obtain the sample x s , x s Matching with the seed library generated in step S4, through which the probability distribution P(B) of the noise level under the current operating state can be estimated;

[0015] S7. Calculate the fault diagnosis result under the current operating condition using the probability distribution obtained in step S5 and step S6.

[0016] Preferably, in step S2, the method of the denoising diffusion model is: gradually adding noise to the turbine vibration data to simulate the process of gradually diffusing the data distribution of the turbine vibration data to the Gaussian distribution, and then learning the process of gradually restoring the turbine vibration data distribution from the Gaussian distribution through the denoising diffusion model.

[0017] Preferably, in step S3, the method for expanding the data set X is: expanding data sets of different categories using a denoising diffusion model, and obtaining samples at different noise levels by performing multi-step sampling in Gaussian noise.

[0018] Preferably, in step S4, the fault diagnosis model under different noise levels is trained as a neural network model based on the data under different noise levels. The neural network model uses a convolutional neural network layer for feature extraction. The input of each neural network model is a vibration signal sample, and the output is the probability that the current sample is normal or under a certain type of abnormal fault under the corresponding noise level.

[0019] Preferably, in step S4, the seed library is generated by: 0,1 ,X 0,2 ,...,X 0,m By using the neural network model for feature extraction, m groups of features can be obtained. Each group of features corresponds to samples under different noise levels. Each group of features is clustered by kmeans in turn. Through kmeans clustering, k cluster centers can be obtained. These k cluster centers are the seeds under the current noise level. For all m noise levels, a total of m×k seeds can be obtained.

[0020] Preferably, in step S6, x s The matching method with the seed library generated in S4 is as follows: first, the sample x is matched by the neural network model. s Perform feature extraction, then calculate the cosine similarity between the sample features and the seeds in the seed library, and take the maximum value of the cosine similarity between each group of seeds and sample features as the similarity between the current sample and the group of seeds:

[0021] S j =min 0≤z≤k (x·s j,z )

[0022] Where x is the sample, s z is the zth seed of the jth group of seeds in the seed bank, S j is the similarity between the sample and the jth group of seeds;

[0023] The softmax function is used to obtain the noise level distribution of the equipment under the current operating conditions:

[0024] .

[0025] Preferably, in step S7, the fault diagnosis result under the current operating condition is calculated using the full probability formula:

[0026]

[0027] Where P i is the probability of the i-th type of fault under the current operating conditions, P(A i |B j ) is the probability of the i-th type of failure under the j-th fault level given by the diagnostic model, P(B j ) is the probability that the current operating condition is the jth noise level obtained through seed library matching.

[0028] Preferably, in step S1, the method for collecting normal operating vibration data and fault vibration data of the turbine is: installing a sensor at the turbine bearing seat, so that the turbine is in normal working state and various fault conditions to collect normal operating vibration data and fault vibration data through the sensor; the calculations in step S1, step S2, step S3 and step S4 are all performed by the computing server, and the process data in step S1, step S2, step S3 and step S4 are all stored in the disk of the computing server; the operations and calculations in step S5, step S6 and step S7 are all performed by the industrial computer, and the process data in step S5, step S6 and step S7 are all stored in the disk of the industrial computer.

[0029] The present invention also provides a turbine vibration fault diagnosis system, including a sensor, a computing server and an industrial computer, wherein the sensor is used to be fixed on a turbine bearing seat to measure the vibration data of the turbine bearing, a neural network model training framework and a first database are arranged inside the computing server, the computing server is used for neural network model training, the industrial computer has a built-in denoising diffusion model inference module, a fault diagnosis model inference module and a second database, the industrial computer can obtain the probability of turbine failure by calling the denoising diffusion model inference module and the fault diagnosis model inference module, the sensor is electrically connected to the computing server and the industrial computer, and the computing server is electrically connected to the industrial computer.

[0030] Preferably, the sensor is a Huige HG6800A integrated vibration transmitter, the computer server is an Alibaba Cloud GPU cloud server gn6v, and the industrial computer is a Lenovo industrial computer ECI-430 platform Nvidia 3060TI independent graphics version.

[0031] Compared with the prior art, the present invention has achieved the following technical effects:

[0032] The present invention provides a method and system for diagnosing vibration faults of a steam turbine. The method utilizes the characteristic of a denoising diffusion model that can gradually reconstruct the distribution of required data from Gaussian noise, performs data enhancement on fault samples, and the intermediate results of the enhancement process can be used as reference data under different noise levels when the steam turbine is running, thereby generating a noise level seed library. In the actual reasoning process, the fault diagnosis model obtained by training with sufficient data can be used for diagnosis, and the current operating state can also be matched with the noise level seed library to obtain the distribution of the current noise level. The final diagnosis result is given by combining the fault diagnosis result with the noise level distribution, and more fault samples are generated, which alleviates the problem of difficulty in obtaining abnormal data and improves the diagnosis effect. At the same time, the present invention takes noise into consideration in the model, and can give more reliable diagnosis results in an industrial environment with more noise, thereby improving the diagnosis stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 This is a flow chart of the data preparation phase in Example 1;

[0035] Figure 2 This is a flow chart of the on-site diagnosis phase in Embodiment 1;

[0036] Figure 3 This is the steam turbine vibration fault diagnosis system in the second embodiment.

[0037] 1- turbine bearing seat, 2- sensor, 3- computing server, 4- industrial computer. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The purpose of the present invention is to provide a turbine vibration fault diagnosis method and system thereof to solve the problems existing in the above-mentioned prior art, to generate more fault samples to improve the diagnosis effect, to avoid noise interference, and to improve the diagnosis stability.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Embodiment 1

[0042] This embodiment provides a method for diagnosing vibration faults of a steam turbine. Figures 1-2 As shown, the steps are as follows:

[0043] (I) Data preparation stage: Figure 1 As shown,

[0044] S1. Install a non-contact eddy current sensor at the turbine bearing seat, and collect normal operation vibration data and fault vibration data through the sensor when the turbine is in normal working state and various fault conditions, and obtain the data set X = {X 0 ,X 1 ,X 2 …X n-1}, and stored in the disk of the computing server, where X 0 For normal operation vibration data, X 1 ,X 2 ...X n-1 The fault vibration data of n-1 different fault types can also be directly obtained from external data sources;

[0045] S2. On the computing server, a denoising diffusion model is trained for each type of data in the data set X obtained in step S1. Through training, n denoising diffusion models can be obtained, where the model set is D = {D 0 ,D 1 ,D 2 …D n-1} and store it in the disk of the computing server;

[0046] Specifically, the method of the denoising diffusion model is: gradually adding noise to the turbine vibration data to simulate the process of gradually diffusing the data distribution of the turbine vibration data to the Gaussian distribution, and then learning the process of gradually restoring the distribution of the turbine vibration data from the Gaussian distribution through the denoising diffusion model, so that the trained model can have the ability to reconstruct the required distribution from the Gaussian distribution;

[0047] S3. On the computing server, the denoising diffusion model set D obtained in step S2 is used to amplify the data set X one by one. Each model D obtains m data sets with different noise levels. And store it in the disk of the computing server, where i = 0, 1, 2...n-1, j = 1, 2...m;

[0048] Specifically, the method of expanding the data set X is as follows: expanding the data sets of different categories using the denoising diffusion model, and obtaining samples under different noise levels by performing multi-step sampling in Gaussian noise;

[0049] S4. On the computing server, for the data set obtained in step S3 m noise level data are used to train m vibration fault diagnosis models. Taking noise level 1 as an example, the training data set is {X 0,1 ,X 1,1 ,X 2,1 …X n-1,1}, the trained model is M 1 , thus obtaining the fault diagnosis model set M = {M j}, where j = 1, 2...m, and at the same time, we get m sets of hidden variables H = H under different noise levels i,j , a seed library is generated through the hidden variable set H of the normal running samples in the set, and stored in the disk of the computing server;

[0050] In specific implementation, the fault diagnosis model under different noise levels is trained as a neural network model based on data under different noise levels. The neural network model uses a convolutional neural network layer for feature extraction. The input of each neural network model is a vibration signal sample, and the output is the probability that the current sample is normal or under a certain type of abnormal fault under the corresponding noise level;

[0051] Specifically, the seed library is generated by: taking the enhanced normal sample data set {X 0,1 ,X 0,2 ,...,X 0,m Using the neural network model for feature extraction, m groups of features can be obtained, each group of features corresponds to samples under different noise levels, and each group of features is clustered by kmeans in turn. Through kmeans clustering, k cluster centers can be obtained. These k cluster centers are the seeds under the current noise level. For all m noise levels, a total of m×k seeds can be obtained;

[0052] (II) On-site diagnosis stage: Figure 2 As shown, after the preparation phase, the denoising diffusion model D and the fault diagnosis model set M = {M j} and the seed library are copied to the industrial computer,

[0053] S5. The steam turbine is running, and the industrial computer uses a sensor to read the vibration sample of the steam turbine every 1 second, and uses the m fault diagnosis models obtained in step S4 to perform fault diagnosis on the data, and obtains the fault type distribution P(A|B) of the current steam turbine bearing under different noise levels, and stores it in the disk of the industrial computer, where B is a random variable representing the noise level, and A is a random variable representing the fault type;

[0054] S6. When the steam turbine is running, the industrial computer uses the sensor to read the vibration sample of the steam turbine every 60 seconds and obtains the sample x s The industrial computer first uses the neural network model to train the sample x s Perform feature extraction, then calculate the cosine similarity between the sample features and the seeds in the seed library, and take the maximum value of the cosine similarity between each group of seeds and sample features as the similarity between the current sample and the group of seeds:

[0055] S j =min 0≤z≤k (x·s j,z )

[0056] Where x is the sample, s z is the zth seed of the jth group of seeds in the seed bank, S j is the similarity between the sample and the jth group of seeds;

[0057] The softmax function is used to obtain the noise level distribution of the equipment under the current operating conditions:

[0058]

[0059] That is the probability distribution P(B) of the noise level under the current operating state, and it is stored in the disk of the industrial computer.

[0060] It is worth noting that the above matching method can update the noise level of the detection equipment under the current operating conditions in real time and match the current sampling results with the dataset seed library.

[0061] S7. Use P(A|B) obtained in step S5 and P(B) obtained in step S6 to calculate the total probability formula

[0062]

[0063] Calculate the fault diagnosis result under the current operating conditions, where P i is the probability of the i-th type of fault under the current operating conditions, P(A i |B j ) is the probability of the i-th type of failure under the j-th fault level given by the diagnostic model, P(B j ) is the probability that the current operating condition is the jth noise level obtained through seed library matching.

[0064] A method for diagnosing vibration faults of a steam turbine provided in the above embodiment utilizes the characteristic that a denoising diffusion model can gradually reconstruct the distribution of required data from Gaussian noise, and performs data enhancement on fault samples. Meanwhile, the intermediate results of the enhancement process can be used as reference data under different noise levels when the steam turbine is running, thereby generating a noise level seed library. In the actual reasoning process, it is possible to use the fault diagnosis model obtained by training with sufficient data for diagnosis, and it is also possible to match the current operating state with the noise level seed library to obtain the distribution of the current noise level. The final diagnosis result is given by combining the fault diagnosis result with the noise level distribution, and more fault samples are generated, alleviating the problem of difficulty in obtaining abnormal data, so as to improve the diagnosis effect. Meanwhile, the present invention takes noise into consideration in the model, and can give more reliable diagnosis results in industrial environments with more noise, thereby improving the diagnosis stability.

[0065] Embodiment 2

[0066] This embodiment provides a steam turbine vibration fault diagnosis system, such as Figure 3 As shown, the turbine vibration fault diagnosis system includes a sensor 2, a computing server 4 and an industrial computer 3. The sensor 2 is used to be fixed on the turbine bearing seat 1 to measure the turbine bearing vibration data. The computing server 4 is internally provided with a neural network model training framework and a first database. The computing server 4 is used for neural network model training. The industrial computer 3 has built-in denoising diffusion model reasoning module, fault diagnosis model reasoning module and a second database. The industrial computer 3 can obtain the probability of turbine failure by calling the denoising diffusion model reasoning module and the fault diagnosis model reasoning module. The sensor 2 is electrically connected to the computing server 4 and the industrial computer 3, and the computing server 4 is electrically connected to the industrial computer 3.

[0067] Specifically, sensor 2 is the Huige HG6800A integrated vibration transmitter, computer server 4 is the Alibaba Cloud GPU cloud server gn6v, and industrial computer 3 is the Lenovo industrial computer ECI-430 platform Nvidia3060TI independent graphics version.

[0068] During specific implementation, the sensor 2, the computing server 4 and the industrial computer 3 are connected via a data line, and steps S1 to S7 in the first embodiment are executed.

[0069] A steam turbine vibration fault diagnosis system provided in the above embodiment utilizes the characteristic of a denoising diffusion model that can gradually reconstruct the distribution of required data from Gaussian noise, and performs data enhancement on fault samples. At the same time, the intermediate results of the enhancement process can be used as reference data under different noise levels when the steam turbine is running, thereby generating a noise level seed library. In the actual reasoning process, it is possible to use the fault diagnosis model obtained by training with sufficient data for diagnosis, and at the same time, it is possible to match the current operating state with the noise level seed library to obtain the distribution of the current noise level. The final diagnosis result is given by combining the fault diagnosis result with the noise level distribution, generating more fault samples, alleviating the problem of difficulty in obtaining abnormal data, and improving the diagnosis effect. At the same time, the present invention takes noise into consideration in the model, and can provide more reliable diagnosis results in industrial environments with more noise, thereby improving the diagnosis stability.

[0070] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for diagnosing vibration faults of steam turbines. Features: The steps include: S1. Collect the normal operation vibration data and fault vibration data of the steam turbine to obtain the data set ,in To run vibration data normally, for Fault vibration data of different fault types; S2. For the data set obtained in step S1 Each type of data in the denoising diffusion model is trained, and the training can be obtained denoising diffusion models, where the model set is ; S3. Use the denoising diffusion model set obtained in step S2 For the dataset Each model is expanded one by one. We get m data sets with different noise levels. ,in , ; S4. For the data set obtained in step S3 m noise level data are used to train m vibration fault diagnosis models. Taking noise level 1 as an example, the training data set is The trained model is , thus obtaining the fault diagnosis model set ,in , and obtain m sets of hidden variables at different noise levels , Generate a seed library through the hidden variable set of normal operating samples in the set H; S5. Run the steam turbine, read the steam turbine vibration sample every 1 second, use the m fault diagnosis models obtained in step S4 to perform fault diagnosis on the data, and obtain the fault type distribution of the current steam turbine bearing under different noise levels. , where B is a random variable representing the noise level, and A is a random variable representing the fault type; S6. Run the turbine and read the turbine vibration sample every 60 seconds , get the sample ,Will Matching with the seed library generated in step S4 can estimate the probability distribution of the noise level under the current operating state. ; S7. Calculate the fault diagnosis result under the current operating condition using the probability distribution obtained in step S5 and step S6.

2. The steam turbine vibration fault diagnosis method according to claim 1, Features: In step S2, the method of the denoising diffusion model is: gradually adding noise to the turbine vibration data to simulate the process of gradually diffusing the data distribution of the turbine vibration data to the Gaussian distribution, and then using the denoising diffusion model to learn the process of gradually restoring the turbine vibration data distribution from the Gaussian distribution.

3. The steam turbine vibration fault diagnosis method according to claim 1, Features: In step S3, the data set The method of amplification is: amplify different categories of data sets using a denoising diffusion model, and obtain samples under different noise levels by performing multi-step sampling in Gaussian noise.

4. The steam turbine vibration fault diagnosis method according to claim 1, Features: In step S4, the fault diagnosis model under different noise levels is trained as a neural network model based on the data under different noise levels. The neural network model uses a convolutional neural network layer for feature extraction. The input of each neural network model is a vibration signal sample, and the output is the probability that the current sample is normal or under a certain type of abnormal fault under the corresponding noise level.

5. The steam turbine vibration fault diagnosis method according to claim 4, Features: In step S4, the seed library is generated by: taking the enhanced normal sample data set By using the neural network model for feature extraction, m groups of features can be obtained. Each group of features corresponds to samples under different noise levels. Each group of features is clustered by kmeans in turn. K cluster centers can be obtained through kmeans clustering. These k cluster centers are the seeds under the current noise level. For all m noise levels, a total of seeds.

6. The steam turbine vibration fault diagnosis method according to claim 5, Features: In step S6, The matching method with the seed library generated in S4 is: firstly, the sample is matched by the neural network model. Perform feature extraction, then calculate the cosine similarity between the sample features and the seeds in the seed library, and take the maximum value of the cosine similarity between each group of seeds and sample features as the similarity between the current sample and the group of seeds: S j =min 0≤z≤k (x·s j,z ), where x is the sample, is the zth seed of the jth group of seeds in the seed bank, is the similarity between the sample and the jth group of seeds; The softmax function is used to obtain the noise level distribution of the equipment under the current operating conditions: .

7. The steam turbine vibration fault diagnosis method according to claim 1, Features: In step S7, the fault diagnosis result under the current operating condition is calculated using the full probability formula: In the formula, is the probability of the i-th type fault under the current operating condition, is the probability of the i-th type of fault under the j-th fault level given by the diagnostic model, is the probability that the current operating condition is the jth noise level obtained through seed library matching.

8. The steam turbine vibration fault diagnosis method according to claim 1, Features: In step S1, the method for collecting normal operation vibration data and fault vibration data of the steam turbine is: installing a sensor at the bearing seat of the steam turbine, so that the steam turbine is in a normal working state and various fault conditions to collect normal operation vibration data and fault vibration data through the sensor; the calculations in step S1, step S2, step S3 and step S4 are all performed by a computing server, and the process data in step S1, step S2, step S3 and step S4 are all stored in the disk of the computing server; the operations and calculations in step S5, step S6 and step S7 are all performed by an industrial computer, and the process data in step S5, step S6 and step S7 are all stored in the disk of the industrial computer.

9. A steam turbine vibration fault diagnosis system applied to the steam turbine vibration fault diagnosis method according to any one of claims 1 to 8, Features: include: A sensor, a computing server and an industrial computer, wherein the sensor is used to be fixed on a turbine bearing seat to measure turbine bearing vibration data, a neural network model training framework and a first database are arranged inside the computing server, the computing server is used for neural network model training, the industrial computer has built-in denoising diffusion model reasoning module, fault diagnosis model reasoning module and a second database, the industrial computer can obtain the probability of turbine failure by calling the denoising diffusion model reasoning module and the fault diagnosis model reasoning module, the sensor is electrically connected to the computing server and the industrial computer, and the computing server is electrically connected to the industrial computer.

Images