Method for generating rotor fault diagnosis model based on multi-transformation domain and diagnosis method

Abstract

The present invention discloses a method for generating a rotor fault diagnosis model based on multiple transform domains and a diagnosis method. The former includes: collecting vibration signal data of a modular rotor fault of a gas turbine in multiple categories to construct a training and test data set; transforming each data to the short-time Fourier time-frequency domain using the STFT transform domain, transforming it to the Morlet wavelet time-frequency domain using the Morlet wavelet transform domain, and transforming it to the complex network phase space domain using the complex network recurrence plot method; inputting the training data set transformed by the three transform domains into a preset model, and using the CNN network layer for feature extraction; using the LSTM network to perform feature structure learning on the output features, and training to obtain a rotor fault model; using the test data set transformed by the three transform domains to verify and adjust the rotor fault model to obtain a rotor fault diagnosis model for fault detection. The present invention can realize automatic data processing, reduce manual input, and improve the efficiency and accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of gas turbine modular rotor fault diagnosis, and particularly relates to a method for generating a rotor fault diagnosis model and a diagnosis method based on multiple transform domains. Background Technique

[0002] A gas turbine is a common power generation unit and a large-scale power drive system. Due to its high efficiency and high reliability, gas turbines are widely used in many heavy industrial fields such as power plants, aeroengines, and large ships. The gas turbine rotor consists of multiple blades and disks, and the high-speed airflow generated by gas injection drives the rotor to rotate, and finally the energy is converted into electric power or power output. Its working environment is more severe and changeable. It will gradually age during long-term operation in a high-temperature and high-pressure environment, the remaining life will gradually decrease, and the potential possibility of failure will gradually increase. Moreover, due to the complexity and requirements of its high-speed rotating components, the gas turbine rotor is also prone to typical faults such as rubbing, imbalance, misalignment, and cracks. Once a fault occurs, it will not only cause huge economic losses, but even lead to catastrophic casualties and form a serious social impact.

[0003] However, for rotor fault diagnosis, traditional manual detection methods or expert database analysis methods require a large amount of manpower, and at the same time, the diagnosis efficiency and diagnosis accuracy are poor. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a method for generating a rotor fault diagnosis model and a diagnosis method based on multiple transform domains. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0005] In a first aspect, an embodiment of the present invention provides a method for generating a rotor fault diagnosis model based on multiple transform domains, and the method includes:

[0006] Collect vibration signal data of gas turbine modular rotor faults of multiple categories to construct a training data set and a test data set; wherein, the multiple categories include normal, rotor rubbing fault, rotor imbalance fault, and rotor misalignment fault;

[0007] For each data in the training data set and the test data set, respectively use STFT Fourier transform domain transformation to the short-time Fourier time-frequency domain, use Morlet wavelet transform domain transformation to the Morlet wavelet time-frequency domain, and use the complex network recurrence plot method to transform to the complex network phase space domain, and obtain three transformed training data sets and test data sets after the transform domain transformation;

[0008] Input the three transformed training data sets after the transform domain transformation into a preset model, and use the CNN network layer in the preset model for feature extraction to obtain CNN output features;

[0009] Utilize the LSTM network in the preset model to perform feature structure learning on the CNN output features, and use the learning results to train the preset model to obtain a rotor fault model;

[0010] Use the test data set after the three transform domain transformations to verify and adjust the rotor fault model, and obtain a rotor fault diagnosis model that meets the accuracy requirements; wherein, the rotor fault diagnosis model is used for rotor fault classification diagnosis.

[0011] In an embodiment of the present invention, transforming each data in the training data set and the test data set to the short-time Fourier time-frequency domain using the STFT Fourier transform domain includes:

[0012] For each data representing the gas turbine modular rotor fault vibration signal data in the training data set and the test data set Utilize the STFT Fourier transform domain transformation formula Transform it to the short-time Fourier time-frequency domain;

[0013] Among them, A k (t) represents the instantaneous amplitude; K represents the number of components; represents the instantaneous phase; f represents the frequency; τ represents the time parameter; s(τ) represents the original signal; g(τ - t) represents the window function; f(τ - t) represents the frequency distribution function corresponding to the window.

[0014] In an embodiment of the present invention, the window function is a Hamming window.

[0015] In an embodiment of the present invention, in the process of transforming each data in the training data set and the test data set to the Morlet wavelet time-frequency domain using the Morlet wavelet transform, the expression of the Morlet wavelet basis function adopted includes:

[0016]

[0017] Among them, ω0 represents the center frequency.

[0018] In an embodiment of the present invention, in the process of transforming each data in the training data set and the test data set to the complex network phase space domain using the complex network recurrence plot method, the recurrence plot modeling process includes:

[0019] Step 1: Set the sequence u k (k = 1, 2,..., N) and the delay time τ, and perform m-dimensional phase space reconstruction by embedding to obtain a new dynamic system as:

[0020] x i =(u i, u i+τ , …, u i+(m-1)τ )

[0021] i = 1, 2, …, N - (m - 1)τ

[0022] Step 2: Calculate the distance between two points i and j in the reconstructed phase space as follows:

[0023] S ij = ||x i - x j ||

[0024] i = 1, 2, …, N - (m - 1)τ

[0025] j = 1, 2, …, N - (m - 1)τ

[0026] Step 3: Calculate the recurrence value as follows:

[0027] R(i, j) = H(ε - S ij )

[0028] i = 1, 2, …, N - (m - 1)τ

[0029] j = 1, 2, …, N - (m - 1)τ

[0030] where, ||·|| represents the norm; H(·) represents the heaviside function; ε represents the edge - connecting threshold.

[0031] In a second aspect, an embodiment of the present invention provides a rotor fault diagnosis method based on multiple transform domains, and the method includes:

[0032] Obtain the vibration signal data of the modular rotor fault of the gas turbine to be diagnosed;

[0033] Transform the vibration signal data of the modular rotor fault of the gas turbine to be diagnosed into the short - time Fourier time - frequency domain using the STFT Fourier transform domain, into the Morlet wavelet time - frequency domain using the Morlet wavelet transform domain, and into the complex network phase space domain using the complex network recurrence graph method, respectively, to obtain the vibration signal data of the modular rotor fault of the gas turbine after being transformed in three transform domains;

[0034] Input the vibration signal data of the modular rotor fault of the gas turbine after being transformed in the three transform domains into the rotor fault diagnosis model to obtain the rotor fault diagnosis result; wherein, the rotor fault diagnosis model is pre - obtained according to the method for generating the rotor fault diagnosis model based on multiple transform domains described in the first aspect; the rotor fault diagnosis result includes any one of the multiple categories including normal, rotor rub fault, rotor imbalance fault, and rotor misalignment fault.

[0035] Compared with the prior art, the method for generating a rotor fault diagnosis model based on multiple transform domains and the rotor fault diagnosis method based on multiple transform domains proposed in the embodiments of the present invention have the following obvious advantages:

[0036] Enhance the feature expression ability of data: The short-time Fourier transform, Morlet wavelet transform, and complex network recurrence plot phase space transform are all effective data transformation methods. They can convert the original data from time-domain data to the frequency domain or other forms of feature domains, thereby better expressing the features of the data. By dividing different time windows, the capacity of data samples can be effectively increased, which is beneficial to the training and generalization of the CNN-LSTM deep learning model. This can improve the feature expression ability of the CNN-LSTM model and further enhance the accuracy of the model.

[0037] Fuse information at different scales: The short-time Fourier transform and Morlet wavelet transform can analyze data at different scales, thereby fusing information at different scales. The complex network phase space transform can analyze the variation law of data in the phase space. These transforms can help the CNN-LSTM model better capture multi-scale and various forms of information in the data.

[0038] Compensate for the deficiencies of CNN and LSTM: CNN is mainly used to capture the spatial features of data, while LSTM is mainly used to capture the temporal features of data. Inputting the transformed data into the CNN-LSTM model can make full use of the advantages of both models, thereby better analyzing the spatial and temporal features of the data.

[0039] In summary, performing short-time Fourier transform, Morlet wavelet transform, and complex network recurrence plot phase space transform on the data, and inputting the transformed data into CNN-LSTM for deep learning can effectively improve the feature expression ability of the model, fuse information at different scales, and make full use of the advantages of the CNN and LSTM models. It can improve the efficiency and accuracy of fault diagnosis, and can achieve automatic data processing, reducing the human input. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic flowchart of a method for generating a rotor fault diagnosis model based on multiple transform domains provided by an embodiment of the present invention;

[0041] Figure 2 It is a time-frequency diagram of the normal short-time Fourier transform of the rotor in an embodiment of the present invention;

[0042] Figure 3 It is a time-frequency diagram of the short-time Fourier transform of the rotor with rubbing in an embodiment of the present invention;

[0043] Figure 4 It is a time-frequency diagram of the short-time Fourier transform of the unbalanced rotor in an embodiment of the present invention;

[0044] Figure 5 It is the time-frequency diagram of short-time Fourier transform for rotor misalignment in the embodiment of the present invention;

[0045] Figure 6 It is the time-frequency diagram of Morlet wavelet transform for normal rotor in the embodiment of the present invention;

[0046] Figure 7 It is the time-frequency diagram of Morlet wavelet transform for rotor rub in the embodiment of the present invention;

[0047] Figure 8 It is the time-frequency diagram of Morlet wavelet transform for rotor imbalance in the embodiment of the present invention;

[0048] Figure 9 It is the time-frequency diagram of Morlet wavelet transform for rotor misalignment in the embodiment of the present invention;

[0049] Figure 10 It is the schematic diagram of the construction process of the power system in the embodiment of the present invention;

[0050] Figure 11 It is the normal complex network of the rotor for the fault vibration signal data of the modular rotor of the gas turbine in the embodiment of the present invention;

[0051] Figure 12 It is the complex network of the rotor for the fault vibration signal data of the modular rotor of the gas turbine with rotor imbalance in the embodiment of the present invention;

[0052] Figure 13 It is the complex network of the rotor for the fault vibration signal data of the modular rotor of the gas turbine with rotor imbalance in the embodiment of the present invention;

[0053] Figure 14 It is the complex network of the rotor for the fault vibration signal data of the modular rotor of the gas turbine with rotor misalignment in the embodiment of the present invention;

[0054] Figure 15 It is the schematic diagram of the flow of a rotor fault diagnosis method based on multiple transform domains provided by the embodiment of the present invention;

[0055] Figure 16 It is the experimental result of the accuracy rate of the rotor fault diagnosis method based on multiple transform domains in the embodiment of the present invention. Specific implementation manners

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] To facilitate the understanding of the solution of the embodiments of the present invention, a general description of the inventive concept will be given first. The embodiments of the present invention generate a rotor fault diagnosis model for the fault vibration signal data of a gas turbine modular rotor based on multiple transform domains, and use this model for fault diagnosis. Among them,

[0058] Faults such as rotor imbalance, misalignment, and rub-impact of a gas turbine are common gas turbine faults. The short-time Fourier transform (STFT) is a method for time-frequency analysis of signals, which can perform local analysis of signals in time and obtain the frequency information of signals in different time periods. The short-time Fourier transform can be realized by calculating the window function and Fourier transform of the signal. Different types of window functions can be selected, such as Hamming window, Gaussian window, etc., to adapt to the characteristics of different signals. The vibration signal of the gas turbine rotor is divided into multiple time windows, and the signal in each time window can be regarded as steady-state. For the signal in each time window, Fourier transform is performed to obtain its frequency distribution. For adjacent time windows, they can be connected through a certain amount of overlap and smoothing processing, so as to obtain the time-frequency distribution diagram of the entire signal. In the time-frequency diagram, the distribution of the signal at different frequencies and times can be seen. Faults such as rotor imbalance, misalignment, and rub-impact of the gas turbine rotor will generate vibration signals with different frequencies and amplitudes. Through the time-frequency distribution diagram, the distribution of these vibration signals in the time-frequency domain can be clearly seen, so as to determine the type and location of the fault. For example, if there is an imbalance fault in the gas turbine rotor, it will cause obvious peaks in the vibration signal at a specific frequency. Through the time-frequency distribution diagram, the position and amplitude of these peaks can be determined, so as to judge the imbalance situation of the rotor.

[0059] It can be seen that the short-time Fourier transform can decompose the signal in time and frequency, so that the time-frequency analysis of the vibration signal of the gas turbine rotor can be carried out, and further the diagnosis of faults such as rotor imbalance, misalignment, and rub-impact can be realized. The short-time Fourier transform can obtain the distribution of the vibration signal of the gas turbine rotor in the time-frequency domain, so as to diagnose faults such as rotor imbalance, misalignment, and rub-impact of the gas turbine.

[0060] The Morlet wavelet time-frequency analysis is also a signal analysis method for diagnosing faults such as rotor imbalance, misalignment, and rub-impact of a gas turbine. Wavelet analysis is a local analysis method, which can analyze signals at different times and frequencies, so it can better capture the local characteristics of signals. The Morlet wavelet is a complex wavelet, which has local properties in both the time domain and the frequency domain. For the vibration signal of the gas turbine rotor, through the Morlet wavelet time-frequency analysis, the signal can be decomposed in time and frequency, and the phase and amplitude information of different frequency components can be obtained.

[0061] First, perform wavelet transform on the vibration signal of the gas turbine rotor to obtain wavelet coefficients at different scales, that is, signal components at different frequencies. Then, use the Morlet wavelet as the mother wavelet to reconstruct the wavelet coefficients, obtaining the Morlet wavelet time-frequency analysis diagram, which is the distribution of the gas turbine rotor vibration signal in the time-frequency domain.

[0062] Through the Morlet wavelet time-frequency analysis diagram, the phase and amplitude information of different frequency components can be seen, thereby enabling the judgment of whether there are faults such as imbalance, misalignment, and rub-impact in the gas turbine rotor. For example, if there is an imbalance fault in the rotor, significant vibrations will occur at certain frequencies, corresponding to obvious peaks on the Morlet wavelet time-frequency analysis diagram. If there are misalignment or rub-impact faults in the rotor, obvious changes or distortions will occur in the frequency, which can also be identified through the Morlet wavelet time-frequency analysis diagram.

[0063] It can be seen that Morlet wavelet time-frequency analysis is a signal analysis method for diagnosing faults such as imbalance, misalignment, and rub-impact in the gas turbine rotor. By decomposing the distribution of the signal in the time-frequency domain, the phase and amplitude information of different frequency components are obtained, thereby judging the fault type and location of the gas turbine rotor.

[0064] For faults such as rotor rub, imbalance, and misalignment in gas turbines, different time-frequency characteristics will be generated in the vibration signals. For example, the rub fault usually occurs between the rotor and the bearing, resulting in an increase in the high-frequency noise of the rotor vibration. In time-frequency analysis, this fault will cause the high-frequency components in the frequency spectrum to increase, and obvious transients will appear in a short period. For example, the imbalance fault is usually caused by uneven rotor center of gravity or impeller damage. In time-frequency analysis, this fault will cause the components in the frequency spectrum to rotate and split, and obvious primary, secondary, tertiary, etc. harmonic components will appear in the spectrum. For example, in time-frequency analysis of the misalignment fault, this fault will cause the components in the frequency spectrum to shift or misalign, and periodic changes in amplitude and phase will appear in the time domain. The vibration signals generated by the faults of the gas turbine modular rotor have complex time-domain and frequency-domain characteristics, and these characteristics can be reflected as the topological characteristics of the complex network in complex network analysis. The vibration signals generated by the faults of the gas turbine modular rotor have complex time-domain and frequency-domain characteristics, and these characteristics can be reflected as the topological characteristics of the complex network in complex network analysis. For example, the rub fault will cause an increase in the frequency components of the rotor vibration signal, and these frequency components will be reflected as an increase in the degree of nodes in the network and an increase in the number of edges in the network in complex network analysis. For example, the imbalance fault will enhance the rotor vibration signal in a certain direction, and these enhanced signals will be reflected as an increase in the clustering coefficient of some regions of the network in complex network analysis. For example, the misalignment fault will cause periodic changes in the amplitude of the rotor vibration signal, and these changes will be reflected as a decrease in the clustering coefficient of some regions of the network and an increase in the average path length in complex network analysis.

[0065] On this basis, the embodiments of the present invention consider that more fault characteristics in different forms can be mined as much as possible through multi-transform domain transformation, which can improve the accuracy and reliability of rotor fault diagnosis. Therefore, for the faults of rotor rub, imbalance, and misalignment in gas turbines, a method for generating a gas turbine modular rotor fault diagnosis model based on multi-transform domain feature extraction is proposed, and a rotor fault diagnosis is implemented on this basis. Finally, the accuracy and reliability of the proposed rotor fault diagnosis method are verified through modular rotor experiments.

[0066] The following specifically describes the solutions of the embodiments of the present invention.

[0067] In the first aspect, the embodiments of the present invention provide a method for generating a rotor fault diagnosis model based on multi-transform domain. As Figure 1 shown, the method may include the following steps:

[0068] S1, collect vibration signal data of gas turbine modular rotor faults of multiple categories to construct a training data set and a test data set;

[0069] Among them, the multiple categories include normal, rotor rub fault, rotor imbalance fault, and rotor misalignment fault.

[0070] Specifically, an example of a dataset is given: During the signal acquisition process of a gas turbine modular rotor test bench, the rotor speed can be set to 1200 r / min, the sampling frequency can be set to 2048 Hz, and the sampling length can be 1 s. A total of 45 tests were carried out under different rotor states (normal, rotor rub, rotor imbalance, and rotor misalignment), and then a total of 180 sequence sample data were obtained for the gas turbine modular rotor fault dataset, including 4 fault types, namely rotor imbalance, rotor rub, rotor misalignment, and normal. The fault labels were set to 1, 2, 3, and 4 respectively. The above dataset was divided into a training dataset and a test dataset according to a ratio of 6:4, and the corresponding table of fault type labels is shown in Table 1.

[0071] Table 1 Corresponding table of fault type labels

[0072]

[0073]

[0074] Of course, the generation process of the dataset is not limited to the above example, and it can be understood that the vibration signal data of the gas turbine modular rotor collected is time-domain data.

[0075] S2. For each data in the training dataset and the test dataset, use the STFT Fourier transform to transform it into the short-time Fourier time-frequency domain, use the Morlet wavelet transform to transform it into the Morlet wavelet time-frequency domain, and use the complex network recurrence plot method to transform it into the complex network phase space domain, so as to obtain the training dataset and the test dataset after transformation in the three transformation domains;

[0076] The transformation processes of the three transformation domains are specifically described below.

[0077] 1) For the transformation from the STFT Fourier transform domain to the short-time Fourier time-frequency domain

[0078] The short-time Fourier transform (STFT) is used to perform data transformation on the training dataset and test dataset of the fault vibration signal data of the gas turbine modular rotor, converting the vibration signal of the modular rotor into a short-time Fourier time-frequency domain signal. For non-stationary and wide-sense stationary signals, it is often necessary to understand their frequency characteristics changing with time, so time-frequency analysis is required. The short-time Fourier transform process is that non-stationary and wide-sense stationary signals are multiplied by a time window function before transformation, and it is assumed that the signal is stationary within this finite time window. By sliding the window function over the signal, the Fourier transform of the signal in each window is calculated respectively. Finally, a set of spectrograms in time series is obtained and spliced along the time axis to obtain the short-time Fourier transform time-frequency diagram of the signal.

[0079] Specifically, each data in the training dataset and test dataset is respectively transformed to the short-time Fourier time-frequency domain using the STFT Fourier transform domain transformation, including:

[0080] For each data representing the fault vibration signal data of the gas turbine modular rotor in the training dataset and test dataset Using the STFT Fourier transform domain transformation formula Transform it to the short-time Fourier time-frequency domain;

[0081] Among them, A k (t) represents the instantaneous amplitude; K represents the number of components; represents the instantaneous phase; f represents the frequency; τ represents the time parameter; s(τ) represents the original signal; g(τ - t) represents the window function; f(τ - t) represents the frequency distribution function corresponding to the window. However, STFT is restricted by the time-frequency uncertainty principle, and the time resolution and frequency resolution cannot be optimal simultaneously, and it is impossible to represent the energy concentratedly.

[0082] In an optional implementation manner, in the short-time Fourier transform of the fault vibration signal data of the gas turbine modular rotor, the window function is a Hamming window.

[0083] Figures 2 to 5 The Hamming window size is 128, and the short-time Fourier transform time-frequency diagrams of the fault vibration signal data of different types of gas turbine modular rotors Figures 2 to 5 The corresponding categories are normal, rubbing, unbalance, and misalignment in sequence.

[0084] 2) Use the Morlet wavelet transform to transform to the Morlet wavelet time-frequency domain

[0085] As mentioned above, Morlet wavelet time-frequency analysis is a signal analysis method used to diagnose faults such as rotor imbalance, misalignment, and rub-impact in gas turbines. Wavelet analysis is a local analysis method that can analyze signals at different times and frequencies, so it can better capture the local characteristics of signals. The Morlet wavelet is a complex wavelet that has local properties in both the time domain and the frequency domain. For the vibration signals of gas turbine rotors, through Morlet wavelet time-frequency analysis, the signals can be decomposed in time and frequency, and the phase and amplitude information of different frequency components can be obtained.

[0086] Specifically, in the process of transforming each data in the training data set and the test data set to the Morlet wavelet time-frequency domain using the Morlet wavelet transform, the expression of the Morlet wavelet basis function includes:

[0087]

[0088] The basis function of the Morlet wavelet is composed of a complex trigonometric function multiplied by an exponential decay function. Here, ω0 represents the central frequency. In an optional implementation, the Morlet wavelet transform of the vibration signal data of the gas turbine modular rotor can set the scale parameter. Figures 6 to 9 It is the time-frequency diagram of the Morlet wavelet transform of the vibration signal data of the gas turbine modular rotor with different categories, and the scale coefficient is 256. Figures 6 to 9 The corresponding categories are normal, rub-impact, imbalance, and misalignment in sequence.

[0089] 3) Use the complex network recurrence plot method to transform to the complex network phase space domain

[0090] Use the complex network recurrence plot phase space transformation to establish complex network models for the training data set and the test data set of the vibration signal data of the gas turbine modular rotor respectively; the recurrence plot modeling method involves the delay time τ, the embedding dimension m of the phase space reconstruction, and the connection threshold ε.

[0091] In an optional implementation, in the process of transforming each data in the training data set and the test data set to the complex network phase space domain using the complex network recurrence plot method, the recurrence plot modeling process includes:

[0092] Step 1: Set the sequence u k (k = 1, 2, …, N) and the delay time τ, and perform m-dimensional phase space reconstruction by embedding to obtain a new dynamic system:

[0093] x i =(u i , u i+τ , …, u i+(m-1)τ )

[0094] i = 1, 2, …, N - (m - 1)τ

[0095] Please refer to Figure 10 the schematic diagram of the construction process of the shown power system x i The sequence u k (k = 1, 2, …, N) represents a gas turbine modular rotor fault vibration signal data, which is a time-domain data and has several numerical values. x i is a vector obtained by uniformly sampling m data from u k at intervals of the delay time τ. For example, x1 obtained from the first sampling consists of Figure 10 each blue point connected by the red line in Figure 10 x2 obtained from the second sampling consists of each blue point connected by the green line in i and so on. After N - (m - 1)τ samplings, it can be seen that compared with the previous time, each data in the subsequent sampling is shifted one place to the right, and all the obtained x

[0096] Step 2: Calculate the distance between two points i and j in the reconstructed phase space as:

[0097] S ij = ||x i - x j ||

[0098] i = 1, 2, …, N - (m - 1)τ

[0099] j = 1, 2, …, N - (m - 1)τ

[0100] In the embodiment of the present invention, for any x i and x j the distance is calculated. It can be understood that all the obtained distances S ij can form a distance matrix, where the element in the i-th row and j-th column is S ij .

[0101] Step 3: Calculate the recurrence value as:

[0102] R(i, j) = H(ε - S ij )

[0103] i = 1, 2, …, N - (m - 1)τ

[0104] j = 1, 2, …, N - (m - 1)τ

[0105] In the embodiment of the present invention, a connection threshold ε is set, for the element S in the distance matrix ij, if it is less than ε, the element R(i, j) can be correspondingly set to 1, otherwise the element R(i, j) is set to 0, then a new distance matrix represented by 0 or 1 can be obtained. If the element R(i, j) at the position (i, j) is 1, then the corresponding x i and x j are connected by an edge. If the element at the position (i, j) is 0, it means that the corresponding x i and x j are not connected by an edge. In this way, a recursive graph corresponding to each data set can be obtained.

[0106] Among them, ||·|| represents the norm; H(·) represents the heaviside function; ε represents the edge connection threshold.

[0107] Figures 11 to 14 are respectively the complex network models of the gas turbine modular rotor fault vibration signal data of different categories. The serial numbers in the figure represent the corresponding x i . Figures 11 to 14 They are in turn normal, rub, unbalance, misalignment. The delay time τ = 8, the embedding dimension m = 4, and the edge connection threshold ε = 0.8.

[0108] S3, input the training data sets after the three transform domains are transformed into a preset model, and use the CNN network layer in the preset model to extract features to obtain CNN output features;

[0109] The preset model in the embodiment of the present invention includes a CNN network layer and an LSTM network, which can be simply referred to as a CNN-LSTM network model.

[0110] The CNN convolutional neural network is a deep learning network model that uses non-linear mapping to reduce the dimensionality of local neighborhood data. It is an effective modeling tool for complex non-linear models and has advantages in multi-dimensional data processing. The convolutional neural network reduces the risk of overfitting while reducing the number of network parameters through local connection and parameter sharing. The CNN layer includes a convolution operation layer, a normalization layer, and an activation layer, which can extract image features by itself, including color, texture, shape, and the topological structure of image neurons, and is not affected by operations such as image scaling and distortion, and has good robustness. The role of the convolution operation layer is to extract features. The main parameters involved in the convolution operation layer include: the sliding window step size, the convolution kernel size, the convolution kernel elements, the padding edge, and the number of convolution kernels. Table 2 is the network structure of the CNN convolution layer.

[0111] Table 2

[0112]

[0113]

[0114] Those skilled in the art can understand that after the data transformed by the three transform domains is input into the CNN network layer for feature extraction, the corresponding CNN output features can be obtained.

[0115] S4. Use the LSTM network in the preset model to learn the feature structure of the CNN output features, and use the learning results to train the preset model to obtain a rotor fault model.

[0116] The LSTM long short-term memory network is a specific form of the RNN recurrent neural network. However, RNN encounters great difficulties when dealing with long-term dependencies of data (i.e., nodes far apart in the time series) because when calculating the connections between nodes far apart, it involves the chain multiplication of Jacobian matrices, which often leads to the problem of gradient disappearance or, to a certain extent, the problem of gradient explosion. The LSTM network blocks the above chain multiplication of Jacobian matrices by adding an input gate, a forget gate, and an output gate, which can effectively prevent gradient disappearance and at the same time increase the long-term state memory link to ensure the long-term dependency relationship of data. The calculation formulas for the input gate, forget gate, and output gate of the LSTM network are as follows:

[0117] f t = σ(W f · [h t-1 , x t + b f )

[0118] i t = σ(W i · [h t-1 , x t + b i )

[0119]

[0120]

[0121] o t = σ(W o · [h t-1 , x t + b0)

[0122] h t = o t * tanh(C t )

[0123] Among them, the first formula is the forget gate update formula; the second to fourth formulas are the input gate update formulas; the fifth and sixth formulas are the output gate update formulas.

[0124] Among them, the forget gate reads the short-term memory state h at the previous moment t-1and the current neuron input x t , through the sigmod activation operation σ, determine the forget gate f t The short-term memory state h of the previous moment t-1 and the current neuron input x t The degree of forgetfulness. t ∈[0,1], “1” means retain everything, “0” means forget everything; then the forget gate f t and the previous long-term memory state C t-1 Do the inner product and store it in the current long-term memory state C t ; While the forget gate decides to forget information, the input gate i t Determine the short-term memory state h at the previous moment t-1 and the current neuron input x t The input gate also selects a current long-term memory state through the tanh activation operation. AND input gate i t Do the addition and store it in the current long-term memory state C t ;Finally output gate o t Or output the short-term memory state h of the previous moment through sigmod operation t-1 and the current neuron input x t The output gate is also connected to the current long-term memory state C t The result after the tanh activation operation is processed as an inner product to generate the short-term memory state of the current moment. Repeat the above process at the next moment to achieve the synchronization of long and short memory states, thereby ensuring the long-term dependency of data. In the above process, the forget gate, input gate, and output gate all perform sigmod operations to determine their respective forgetting, input, and output of the short-term memory state h at the previous moment. t-1 and the current neuron input x t The degree of information is obtained through their respective weight matrices W f , W i , W C , W o and the bias vector b f , b i , b C , b0. The meanings of the specific symbols in the formula are explained in related technical terms and will not be explained in detail here.

[0125] Extract the features of the CNN network layer from the data in the training dataset, expand them to a size of 75 * 2048 (512 * 15 * 20 = 75 * 2048), extract the input and feed it into the LSTM network for feature structure learning. Finally, through the fully connected layer, when training the model, set a random seed and use the Adam optimizer of the random optimization method to perform training and learning to obtain a trained rotor fault model. For the specific training process, please understand it in combination with the training process of a general neural network, and no detailed description will be given here.

[0126] S5. Use the test dataset after the three transform domain transformations to verify and adjust the rotor fault model, and obtain a rotor fault diagnosis model that meets the accuracy requirements;

[0127] Among them, the rotor fault diagnosis model is used for rotor fault classification diagnosis.

[0128] After obtaining the rotor fault model using the training dataset in the embodiment of the present invention, use the test dataset to verify the rotor fault model. If the verification meets the accuracy requirements, the rotor fault model can be determined as a usable rotor fault diagnosis model for iterative use in real-time classification diagnosis. If the accuracy requirements are not met, the above process can be repeated, reset the random seed, and continuously train and verify iteratively until the rotor fault model meets the accuracy requirements to obtain the rotor fault diagnosis model.

[0129] Compared with the prior art, the method for generating a rotor fault diagnosis model based on multiple transform domains proposed in the embodiment of the present invention has the following obvious advantages:

[0130] Enhance the feature expression ability of data: The short-time Fourier transform, Morlet wavelet transform, and complex network recurrence plot phase space transform are all effective data transformation methods. They can convert the original data from time-domain data to the frequency domain or other forms of feature domains, so as to better express the features of the data. By dividing different time windows, the capacity of data samples can be effectively increased, which is beneficial to the training and generalization of the CNN-LSTM deep learning model. This can improve the feature expression ability of the CNN-LSTM model and thus improve the accuracy of the model.

[0131] Fuse information at different scales: The short-time Fourier transform and Morlet wavelet transform can analyze data at different scales, so as to fuse information at different scales. The complex network phase space transform can analyze the variation law of data in the phase space. These transformations can help the CNN-LSTM model better capture multi-scale and various forms of information in the data.

[0132] Compensating for the deficiencies of CNN and LSTM: CNN is mainly used to capture the spatial features of data, while LSTM is mainly used to capture the temporal features of data. Inputting the transformed data into the CNN-LSTM model can make full use of the advantages of both models, thus better analyzing the spatial and temporal features of data.

[0133] In summary, performing short-time Fourier transform, Morlet wavelet transform, and complex network recurrence plot phase space transform on the data, and inputting the transformed data into CNN-LSTM for deep learning can effectively improve the feature expression ability of the model, fuse information at different scales, and make full use of the advantages of CNN and LSTM models, which can improve the efficiency and accuracy of fault diagnosis, and can achieve automatic data processing and reduce human input.

[0134] In a second aspect, an embodiment of the present invention provides a rotor fault diagnosis method based on multiple transform domains. Please refer to Figure 15 , and this method includes the following steps:

[0135] S101, obtaining the gas turbine modular rotor fault vibration signal data to be diagnosed;

[0136] S102, respectively using STFT Fourier transform domain to transform the gas turbine modular rotor fault vibration signal data to be diagnosed into the short-time Fourier time-frequency domain, using Morlet wavelet transform domain to transform it into the Morlet wavelet time-frequency domain, and using the complex network recurrence plot method to transform it into the complex network phase space domain, obtaining the gas turbine modular rotor fault vibration signal data after transformation in three transform domains;

[0137] S103, inputting the gas turbine modular rotor fault vibration signal data after transformation in the three transform domains into the rotor fault diagnosis model to obtain the rotor fault diagnosis result;

[0138] Among them, the rotor fault diagnosis model is obtained in advance according to the method for generating a rotor fault diagnosis model based on multiple transform domains described in the first aspect; the rotor fault diagnosis result includes any one of the multiple categories including normal, rotor rubbing fault, rotor imbalance fault, and rotor misalignment fault.

[0139] Regarding the content of the rotor fault diagnosis model based on multiple transform domains, please refer to the relevant description in the first aspect and will not be elaborated here.

[0140] Please refer to Figure 16 As shown, experiments have confirmed that the fault diagnosis accuracy rate of the rotor fault diagnosis method based on multiple transform domains in this embodiment of the present invention is as high as 98.5%, and the false alarm rate is less than 3%, verifying the correctness and feasibility of the CNN-LSTM rotor fault diagnosis method based on multiple transform domains.

[0141] The rotor fault diagnosis method based on multiple transform domains provided by the embodiments of the present invention is a CNN-LSTM rotor fault diagnosis method based on multiple transform domains. It uses a pre-generated rotor fault diagnosis model based on multiple transform domains. In the fault diagnosis of the modular rotor of a gas turbine, the data is subjected to short-time Fourier transform, Morlet wavelet transform, and complex network recurrence plot phase space transform, and the transformed data is input into CNN-LSTM for deep learning, which can effectively improve the feature expression ability of the model, fuse information of different scales, and make full use of the advantages of CNN and LSTM models. This method can achieve accurate rotor fault diagnosis and can also realize automatic data processing, reducing the labor input.

[0142] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are all included in the protection scope of the present invention.

Claims

1. A rotor fault diagnosis model generation method based on multiple transformation domains, characterized in that: include: Collecting multiple categories of gas turbine simulated rotor fault vibration signal data to construct a training data set and a test data set; wherein the multiple categories include normal, rotor rubbing fault, rotor imbalance fault and rotor misalignment fault; Each data in the training data set and the test data set is transformed into the short-time Fourier time-frequency domain using the STFT Fourier transform domain, transformed into the Morlet wavelet time-frequency domain using the Morlet wavelet transform transform domain, and transformed into the complex network phase space domain using the complex network recursive graph method, to obtain the training data set and the test data set after the three transform domain transformations; Input the training data sets transformed by the three transform domains into a preset model, and use the CNN network layer in the preset model to perform feature extraction to obtain CNN output features; The CNN output features are used to perform feature structure learning using the LSTM network in the preset model, and the learning results are used to train the preset model to obtain a rotor fault model; The rotor fault model is verified and adjusted using the test data sets transformed by the three transform domains to obtain a rotor fault diagnosis model that meets the accuracy requirements; wherein the rotor fault diagnosis model is used for rotor fault classification diagnosis.

2. The rotor fault diagnosis model generation method based on multiple transformation domains according to claim 1 is characterized in that: Each data in the training data set and the test data set is transformed into the short-time Fourier time-frequency domain using the STFT Fourier transform domain, including: For each data in the training data set and the test data set representing the vibration signal data of the gas turbine modeled rotor fault Using the STFT Fourier transform domain transformation formula Transform it into the short-time Fourier time-frequency domain; Among them, A k (t) represents the instantaneous amplitude; K represents the number of components; Represents the instantaneous phase; f represents the frequency; τ represents the time parameter; s(τ) represents the original signal; g(τ-t) represents the window function; f(τ-t) represents the frequency distribution function corresponding to the window.

3. The rotor fault diagnosis model generation method based on multiple transformation domains according to claim 2 is characterized in that: The window function is a Hamming window.

4. The rotor fault diagnosis model generation method based on multiple transformation domains according to claim 1 is characterized in that: In the process of transforming each data in the training data set and the test data set from the transform domain to the Morlet wavelet time-frequency domain using the Morlet wavelet transform, the expression of the Morlet wavelet basis function used includes: Where ω0 represents the center frequency.

5. The rotor fault diagnosis model generation method based on multiple transformation domains according to claim 1 is characterized in that: In the process of transforming each data in the training data set and the test data set into the complex network phase space domain using the complex network recursive graph method, the recursive graph modeling process includes: Step 1: Set up the sequence k (k=1,2,…,N) and delay time τ, embedded in the m-dimensional phase space reconstruction, the new dynamic system is obtained as follows: x i =(u i ,u i+τ ,…,u i+(m-1)τ ) i=1,2,…,N-(m-1)τ Step 2: Calculate the distance between points i and j in the reconstructed phase space as: S ij =||x i -x j || i=1,2,…,N-(m-1)τ j=1,2,…,N-(m-1)τ Step 3: Calculate the recursive value as: R(i,j)=H(ε-S ij ) i=1,2,…,N-(m-1)τ j=1,2,…,N-(m-1)τ Among them, ||·|| represents the norm; H(·) represents the heaviside function; ε represents the edge threshold.

6. A rotor fault diagnosis method based on multiple transformation domains, characterized in that: include: Acquiring vibration signal data of a simulated rotor fault of a gas turbine to be diagnosed; The gas turbine model rotor fault vibration signal data to be diagnosed is transformed into a short-time Fourier time-frequency domain using an STFT Fourier transform domain, into a Morlet wavelet time-frequency domain using a Morlet wavelet transform transform domain, and into a complex network phase space domain using a complex network recursive graph method, respectively, to obtain gas turbine model rotor fault vibration signal data after three transform domain transformations; The gas turbine modeled rotor fault vibration signal data after the three transformation domain transformations are input into a rotor fault diagnosis model to obtain a rotor fault diagnosis result; wherein, the rotor fault diagnosis model is obtained in advance according to the rotor failure diagnosis model generation method based on multiple transformation domains according to any one of claims 1 to 5; the rotor fault diagnosis result includes any one of the multiple categories including normal, rotor rubbing fault, rotor imbalance fault and rotor misalignment fault.

Images