Hydroelectric generating set fault identification method and system based on ICEEMDAN-VMD-GAF and PCNN-BiGRU

Abstract

The invention discloses a hydroelectric generating set fault identification method and a hydroelectric generating set fault identification system based on ICEEMDAN-VMD-GAF and PCNN-BiGRU. The method comprises the following steps: firstly, collecting a vibration signal, decomposing the signal by using an ICEEMDAN algorithm, and extracting an IMF component; secondly, secondary decomposition and reconstruction are carried out on IMF components through a VMD and a wavelet denoising method; then, converting the IMF reconstruction component into a two-dimensional image by using a GAF algorithm; and finally, extracting image features by using a PCNN network, and inputting the image features into a BiGRU network for classification so as to realize fault identification. According to the method, the advantages of ICEEMDAN-VMD-GAF and PCNN-BiGRU are combined, fault features can be more effectively extracted from complex noise signals, one-dimensional vibration data are converted into two-dimensional image data, and the PCNN-BiGRU capability is fully utilized, so that the recognition accuracy is improved, and the prediction accuracy and the fault degree diagnosis capability are enhanced.

Description

Technical Field

[0001] The invention belongs to the field of hydropower unit fault diagnosis, and in particular relates to a hydropower unit fault identification method and system based on ICEEMDAN-VMD-GAF and PCNN-BiGRU. Background Art

[0002] Hydropower units are the core equipment in the process of hydropower energy conversion. Maintaining safe, stable and reliable operation of the units is the first priority of hydropower station operation. This is not only related to the safety of the power station itself, but also directly determines whether the power station can continuously and stably provide power energy to the power grid. Therefore, the safety and stability of hydropower units have always been a hot topic of concern in the industry and academia. In recent years, with the continuous improvement of my country's advanced manufacturing level, the installed capacity of hydropower stations has shown an explosive growth trend. Hydropower generation equipment is developing in the direction of large-scale, integrated and intelligent, and the internal structure is becoming more and more complex. For a long time, my country's hydropower unit operation and maintenance system has adopted a parallel mode of planned maintenance and post-maintenance. On the one hand, according to the service life and working conditions of the unit equipment, a regular maintenance plan is formulated to inspect, maintain and maintain the equipment. On the other hand, when the unit equipment fails or an emergency occurs, maintenance work is immediately carried out to eliminate the safety hazards of the unit equipment and ensure the safe operation of the unit. This model has reduced the probability of failure of hydropower units in my country for a long period of time, but its defects such as long maintenance period, lack of targeted maintenance, high labor costs and ignoring early signs of unit failure have also limited the application upper limit of this maintenance model.

[0003] In recent years, with the rapid development of sensor technology, computer technology and artificial intelligence technology, the state maintenance system with massive state operation data as the core has been applied in the operation and maintenance of multi-energy equipment such as hydropower units. At present, deep learning models are gradually attracting more and more attention from scholars in the field of turbine fault diagnosis. In this field, convolutional neural networks (CNNs) have attracted much attention. As an end-to-end deep learning model, it greatly simplifies the fault diagnosis process. However, because the vibration signal of the turbine belongs to one-dimensional data, and the existing convolutional network model construction is based on two-dimensional data, the core concept of CNN includes parameter sharing, local perception and pooling, which makes it more suitable for processing two-dimensional data. Although the traditional hydropower unit fault detection technology based on time-frequency domain analysis can play a certain role, its accuracy still needs to be improved. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a hydropower unit fault identification method and system based on ICEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise)-VMD (Variational Mode Decomposition)-GAF (Gram Angular Field) and PCNN-BiGRU.

[0005] The technical solution provided by the present invention is as follows:

[0006] A hydropower unit fault identification method based on ICEEMDAN-VMD-GAF and PCNN-BiGRU includes the following steps:

[0007] S1. Collect vibration signals of hydropower units under different working conditions;

[0008] S2, decomposing the vibration signal by ICEEMDAN algorithm to obtain IMF components;

[0009] S3, using VMD and wavelet denoising method to decompose and reconstruct the IMF components to obtain IMF reconstructed components;

[0010] S4, transforming the IMF reconstruction components into a two-dimensional image using the GAF algorithm;

[0011] S5. Use two-dimensional images to train the neural network for fault identification of hydropower units.

[0012] Furthermore, in step S2, before performing signal decomposition, white noise is superimposed on the collected original vibration signal, and the white noise is decomposed from the original vibration signal by the EMD algorithm:

[0013] x i (t) = x(t) + n k (t),

[0014] Where n k (t) is the white noise obtained by EMD decomposition, k is the decomposition sequence; x i (t) is the vibration signal of the hydropower unit after superimposing white noise, i is the data sequence; x(t) is the original vibration signal of the hydropower unit.

[0015] Furthermore, in step S2, the vibration signal of the hydropower unit after superimposing white noise is decomposed by ICEEMDAN. When the j-th order residual component becomes a monotonic function and the IMF component can no longer be screened out, the screening is stopped. At this time, j is the number of selected modal components:

[0016]

[0017] In the formula, IMF j For the ICEEMDAN algorithm, from x i The jth IMF component generated by the decomposition in (t), e i (t) is the average trend of the signal.

[0018] Further, in step S3, the IMF components are first decomposed using the VMD algorithm, which specifically includes the steps of: performing Hilbert transform on the IMF components to obtain their analytical signals so as to obtain a one-way spectrum; moving the spectrum of the IMF components to the baseband position by multiplying the Hilbert transformed IMF components with the exponential function of the center frequency; applying the Gaussian function to the moved spectrum to calculate its bandwidth as an optimization target; minimizing the sum of the bandwidths of each IMF component under a constrained variational model, updating the center frequency and the component signal, and iterating continuously until the convergence condition is met or the maximum number of iterations is reached, and finally obtaining a number of decomposed IMF components.

[0019] Furthermore, in step S3, a wavelet threshold shrinkage denoising method is used to perform secondary decomposition and signal reconstruction, specifically including the steps of: performing multi-layer wavelet decomposition on the IMF component after VMD decomposition to obtain low-frequency approximate coefficients and high-frequency detail coefficients of each layer; processing the wavelet coefficients to shrink coefficients less than a set threshold to zero, and subtracting the threshold from coefficients greater than the set threshold; using the processed wavelet coefficients to reconstruct the signal through an inverse wavelet transform to obtain an IMF reconstructed component.

[0020] Furthermore, step S4 specifically includes:

[0021] S401, Data Standardization:

[0022] For all data points of the IMF reconstruction component, calculate its minimum value min(x) and maximum value max(x);

[0023] For each data point x in the IMF reconstruction component i , apply the following normalization formula to convert it into normalized form:

[0024]

[0025] S402, polar coordinate conversion:

[0026] Normalize the value of each data point As the point r i :

[0027] Calculate the phase angle θ for each data point i :

[0028] S403, Gram matrix generation:

[0029] The polar coordinate transformation results in the generation of two different Gram matrices: GASF matrix and GADF matrix; GASF matrix element GASF i,j and GADF matrix elements GASDF i,jThe calculation formulas are:

[0030] GASF i,j =cos(θ i +θ j ),

[0031] GADF i,j =cos(θ i -θ j ),

[0032] In the formula, θ i and θ j are the phase angles of the ith and jth data points respectively;

[0033] S404, two-dimensional image generation: directly convert the GASF matrix and the GADF matrix into grayscale images, where the values ​​of the matrix elements correspond to the grayscale levels of the image pixels; each IMF reconstruction component generates a pair of images, namely, a GASF image and a GADF image; and fuse the GASF image and the GADF image into one image.

[0034] Furthermore, in step S5, the hydropower unit fault identification neural network combines the PCNN network and the BiGRU network, firstly using the PCNN network to extract the convolution features of the two-dimensional image, and then inputting the extracted convolution features into the BiGRU network for classification, thereby realizing fault identification.

[0035] Furthermore, the different operating conditions include normal operating conditions and fault operating conditions, and the fault operating conditions include dynamic and static friction, rotor imbalance, hydraulic imbalance, excessive bearing clearance, rotor misalignment, and eccentric vortex in the tailwater pipe.

[0036] A hydropower unit fault identification system based on the above method includes the following modules:

[0037] Vibration data acquisition module, used to collect vibration signals of hydropower units under different working conditions;

[0038] ICEEMDAN decomposition module, used to perform ICEEMDAN decomposition on the collected vibration signal to obtain the corresponding IMF components;

[0039] The secondary decomposition and reconstruction module is used to perform VMD decomposition and wavelet denoising decomposition on the IMF components, and reconstruct the IMF according to the decomposed signals to obtain the IMF reconstructed components;

[0040] GAF image generation module, used to transform the IMF reconstruction components into a two-dimensional image through the GAF algorithm;

[0041] The network building and training module is used to build a hydropower unit fault recognition network and train the hydropower unit fault recognition network using the two-dimensional images output by the GAF image generation module.

[0042] Preferably, the vibration data acquisition module uses a CRY2301 noise sensor to collect vibration data of the hydropower unit under various working conditions, with a sampling frequency of 10240 Hz and a single sampling time of 1 s.

[0043] Compared with the prior art, the present invention has at least the following beneficial effects:

[0044] 1. The present invention combines the advantages of ICEEMDAN-VMD-GAF and PCNN-BiGRU, and can more effectively extract fault features from complex noise signals, thereby improving recognition accuracy; the noise signal of the hydropower unit is decomposed by ICEEMDAN-VMD to extract multiple intrinsic mode components (IMFs), and then the IMF components are transformed by using the Gram angular field (GAF), and the input one-dimensional vibration data is converted into two-dimensional image data as the input of PCNN-BiGRU.

[0045] 2. The present invention decomposes the signal through ICEEMDAN-VMD, which can remove part of the noise and retain the fault characteristics. The difference from the traditional EMD method is that ICEEMDAN integrates adaptive noise and complete integration strategy, thereby improving the stability and accuracy of the decomposition process, so that the model can still maintain a high recognition accuracy in a strong background noise environment.

[0046] 3. The present invention uses GAF to extract features from the original vibration signal of the hydropower unit after noise reduction, extracts it into a two-dimensional image containing more information, fuses the two-dimensional GASF and GADF images, and makes full use of the information of different Gram angle fields, which helps to fully describe the characteristics of the data, and then extracts more abundant and distinctive features, thereby improving the accuracy of classification and recognition. Make the input method conform to the characteristics of CNN that are more suitable for processing two-dimensional data.

[0047] 4. By combining CNN and BIGRU vibration prediction methods, the present invention fully exploits the temporal prediction capability of BiGRU, enhances the accuracy of prediction, and strengthens the diagnosis of fault degree by convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0049] Figure 1 The present invention is a schematic diagram of a flow chart of a method for identifying a hydropower unit fault provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative work are all within the scope of protection of the present invention.

[0051] Embodiment 1

[0052] This embodiment provides a hydropower unit fault identification method based on ICEEMDAN-VMD-GAF and PCNN-BiGRU. Figure 1 As shown, the method mainly includes the following steps:

[0053] S1. The original vibration signal of the hydropower unit is decomposed by the ICEEMDAN algorithm to obtain the intrinsic mode components representing the different frequency characteristics of the signal to reduce the complexity and randomness of the original signal.

[0054] Table 1 Noise sensor CRY2301 parameters

[0055]

[0056] The vibration signal data of the hydropower unit includes two working conditions: normal and fault. There are many types of faults, such as dynamic and static friction, rotor imbalance, hydraulic imbalance, excessive bearing clearance, rotor misalignment, eccentric vortex belt in the tailwater pipe, etc. In this embodiment, the CRY2301 noise sensor is used to collect the vibration data of the hydropower unit under various working conditions. The sampling frequency is 10240Hz and the single sampling time is 1s. CRY2301 has built-in frequency-time weighted filtering, which can perform acoustic analysis functions such as noise statistical analysis, octave analysis, and FFT analysis. The parameters of the noise sensor are shown in Table 1.

[0057] ICEEMDAN (Intrinsic Computing Expressive Empirical Mode Decomposition With Adaptive Noise) is a new adaptive signal decomposition algorithm that can handle nonlinear and non-stationary signals and has good local time-frequency analysis capabilities. The algorithm combines empirical mode decomposition (EMD) with adaptive noise control technology, and introduces adaptive mean shift (AEMD) and adaptive extreme value large jump (AEMD-H) on the basis of EMD to achieve signal decomposition. The main steps of the ICEEMDAN method include:

[0058] 1. Perform AEMD decomposition on the original signal to obtain a set of IMFs;

[0059] 2. Perform AEMD-H decomposition on each IMF to obtain more detailed IMFs;

[0060] 3. Perform adaptive noise control on the decomposed IMFs to remove noise interference;

[0061] 4. Reconstruct the original signal.

[0062] The ICEEMDAN method uses adaptive noise control technology to reduce the impact of noise, and introduces AEMD and AEMD-H to achieve multi-level decomposition of the signal, so as to better capture the local time-frequency characteristics of the signal. Before performing ICEEMDAN decomposition, this embodiment first superimposes white noise on the collected vibration data of the hydropower unit to eliminate the modal aliasing problem. Compared with the EEMD (Ensemble Empirical Mode Decomposition) algorithm that directly adds Gaussian white noise during the decomposition process, this embodiment selects the white noise component decomposed from the original vibration data of the hydropower unit by the EMD algorithm, that is,

[0063] x i (t) = x(t) + n k (t),

[0064] Where n k (t) is the white noise obtained by EMD decomposition, k is the decomposition sequence; x i (t) is the vibration data of the hydropower unit after superimposing white noise, i is the data sequence; x(t) is the original vibration data of the hydropower unit.

[0065] The vibration data of the hydropower unit after superimposing white noise is decomposed by ICEEMDAN. When the j-th order residual component becomes a monotonic function and the IMF component can no longer be screened out, the screening is stopped. At this time, j is the number of selected modal components.

[0066]

[0067] In the formula, IMF j For x i (t) The jth modal component generated by ICEEMDAN, e i (t) is the average trend of the signal.

[0068] S2. Use VMD algorithm and wavelet threshold shrinkage denoising method to perform secondary decomposition and reconstruction on IMF components, and then extract the main information of the components to solve the influence of high-frequency noise on the prediction accuracy of subsequent models. The IMF components processed by VMD and wavelet threshold shrinkage contain richer fault feature information. Specifically:

[0069] S201, VMD decomposition

[0070] 1. Perform Hilbert transform on the IMF component to obtain its analytical signal in order to obtain a one-way spectrum.

[0071] 2. The spectrum of the IMF component is moved to the baseband position by multiplying the Hilbert transformed IMF component with the exponential function of the center frequency.

[0072] 3. Apply the Gaussian function to the shifted spectrum and calculate its bandwidth as the optimization target.

[0073] 4. Under the constrained variational model, minimize the sum of the bandwidths of each IMF component, update the center frequency and component signal, and iterate continuously until the convergence condition is met or the maximum number of iterations is reached, and finally obtain several decomposed IMF components.

[0074] S202, wavelet threshold shrinkage denoising

[0075] Wavelet threshold denoising is a nonlinear method for signal denoising. It sets the noise component in the signal to zero or attenuate by applying a threshold to the wavelet coefficients.

[0076] 1. Select appropriate wavelet basis functions (such as Daubechies wavelet or Haar wavelet) and decomposition layers. It is generally recommended that the decomposition layers be between 3 and 5 layers in order to fully separate the signal from the noise. Perform multi-layer wavelet decomposition on the IMF components after VMD decomposition to obtain the low-frequency approximate coefficients and high-frequency detail coefficients of each layer.

[0077] 2. For wavelet coefficient processing, coefficients less than the threshold are shrunk to zero, while coefficients greater than the threshold are subtracted from the threshold.

[0078] 3. Using the processed wavelet coefficients, reconstruct the signal through inverse wavelet transform to obtain the IMF component after noise reduction.

[0079] S3. Use the GAF algorithm to transform the IMF components, convert the one-dimensional IMF component data into two-dimensional image data, and generate a pair of images for each group of samples, namely, a GADF image and a GASF image.

[0080] The GAF (Gramian Angular Field) algorithm is an unsupervised learning algorithm that uses autocorrelation in time series to create two-dimensional images. The main steps include data standardization, polar coordinate transformation, Gram matrix generation, and generation of GASF and GADF images.

[0081] S301. Data Standardization

[0082] 1. Calculate the maximum and minimum values:

[0083] For all data points of the IMF component, calculate its minimum value min(x) and maximum value max(x);

[0084] 2. Normalized calculation:

[0085] For each data point x in the IMF component i , apply the following normalization formula:

[0086]

[0087] in, are the normalized data points.

[0088] The time series data in the IMF component are converted into a standardized form so that its range is limited to [-1, +1] to ensure the consistency of data in the subsequent polar coordinate and GAF ​​matrix generation.

[0089] S302, Polar coordinate conversion

[0090] 1. Define the radius:

[0091] Normalized value for each data point This is the r of the point i :

[0092] 2. Calculate the phase angle:

[0093] The phase angle θ of each data point i Use the following formula to calculate: Note that i The range is [0,π], which is used to ensure that the phase angle of each point corresponds to its relative position in the data sequence.

[0094] In this step, the normalized time series data Convert to polar coordinates and generate the radius and phase angle of each data point. The purpose of this is to convert one-dimensional time series data into polar coordinates with time series information.

[0095] S303, Gram matrix generation

[0096] The polar coordinate transformation results are used to generate two different Gram matrices: GASF matrix and GADF matrix, which represent the cumulative trend and difference characteristics of the data respectively.

[0097] 1. Generate GASF (Gram Angular Summation Field) matrix:

[0098] The GASF matrix is ​​used to capture the cumulative trend characteristics of time series data. The element GASF i,j The calculation method is:

[0099] GASF i,j =cos(θ i +θ j ),

[0100] Among them, θ i and θ j are the phase angles of the i-th and j-th data points, respectively. Through the sum of the phase angles of two points, the GASF matrix can accumulate the correlation of each pair of data points and reflect the overall trend of the data.

[0101] 2. Generate GADF (Gram Angular Difference Field) matrix:

[0102] The GADF matrix is ​​used to capture the difference characteristics of time series data. The element GASDF i,j The calculation method is:

[0103] GADF i,j =cos(θ i -θ j ),

[0104] Similarly, θ i and θ j are the phase angles of the i-th and j-th data points, respectively. Through the difference in the phase angles of the two points, the GADF matrix can describe the relative changes between the data points and reflect the characteristics of the changes in the data details.

[0105] S304: Generate a two-dimensional image

[0106] By visualizing the GASF and GADF matrices, the one-dimensional time series data is converted into two-dimensional images, which will be used as input data for the convolutional neural network.

[0107] 1. Graphical operation

[0108] The GASF and GADF matrices are directly converted to grayscale images, where the values ​​of the matrix elements correspond to the grayscale levels of the image pixels. Higher matrix values ​​appear as brighter pixels, and lower values ​​appear as darker pixels.

[0109] 2. Image pair generation:

[0110] Each IMF component generates a pair of images, namely GASF image and GADF image. These image pairs contain rich frequency domain and time series features and are suitable for use as CNN input.

[0111] Through the above steps, the time series data of the IMF component is successfully converted into GASF and GADF two-dimensional image data. The GASF and GADF images are fused into one image, that is, the IMF component is finally converted into a fused image.

[0112] S4. Through the above steps, we can obtain a large number of fused images corresponding to the vibration signals of the hydropower unit under different working conditions. These fused images can be used to train the convolutional neural network, and then the trained convolutional neural network can be used to identify the faults of the hydropower unit.

[0113] In this embodiment, the PCNN network is combined with the BiGRU network, and the PCNN network is used to extract the convolution features of the fused image, and then the extracted convolution features are input into the BiGRU network for fault identification (classification). PCNN (Pulse-Coupled Neural Network) is an artificial neural network model built on the principle of biological visual system, with unique dynamic characteristics and connection mechanism, and exhibits powerful image processing capabilities by simulating the pulse behavior of biological neurons. The BiGRU network, i.e., the bidirectional gated recurrent unit network, is a deep learning model based on the gated recurrent unit (GRU). BiGRU processes sequence data by combining the forward and backward GRU units, so that the front and back information of the sequence data can be captured at the same time. The experimental results show that the network model constructed in this embodiment shows a high accuracy rate in identifying hydropower unit faults, and its recognition performance is significantly better than several other existing recognition models.

[0114] Embodiment 2

[0115] Based on the above method, this embodiment provides a hydropower unit fault identification system, which mainly includes:

[0116] 1. Vibration data acquisition module: This module is responsible for real-time collection of vibration data of hydropower units under different operating conditions. The system uses a high-precision CRY2301 noise sensor, and the sampling frequency is set to 10240Hz to ensure that sufficiently detailed vibration information can be captured. The collected data includes normal and various fault conditions, such as dynamic and static friction, rotor imbalance, etc.

[0117] 2.ICEEMDAN decomposition module: This module performs ICEEMDAN decomposition on the collected vibration signal. ICEEMDAN is an adaptive noise complete set empirical mode decomposition algorithm that can handle nonlinear and non-stationary signals and has good local time-frequency analysis capabilities. Through this algorithm, the vibration signal is decomposed into multiple intrinsic mode components (IMFs) to reduce the complexity and randomness of the signal.

[0118] 3. Secondary decomposition and reconstruction module: Based on the ICEEMDAN decomposition, this module further uses the VMD (variational mode decomposition) algorithm to decompose the IMF components, and performs secondary decomposition and signal reconstruction through the wavelet threshold shrinkage denoising method. The VMD algorithm optimizes the decomposition result by minimizing the sum of the bandwidths of each IMF component, while the wavelet denoising method effectively removes high-frequency noise and retains fault feature information.

[0119] 4. GAF image generation module: This module converts the IMF components after secondary decomposition and reconstruction into two-dimensional images through the GAF (Gram Angular Field) algorithm. The GAF algorithm converts one-dimensional time series data into two-dimensional image data containing time series information through steps such as data standardization, polar coordinate conversion, and Gram matrix generation.

[0120] 5. Network construction and training module: This module is responsible for building a hydropower unit fault recognition network and using the two-dimensional images output by the GAF image generation module for training. The system uses a network structure that combines PCNN (pulse coupled neural network) and BiGRU (bidirectional gated recurrent unit network). PCNN is used to extract convolutional features of images, while BiGRU is used to process sequence data for fault classification.

[0121] Through the collaborative work of these modules, the hydropower unit fault identification system can efficiently identify and classify the fault types of hydropower units, thereby improving the accuracy and efficiency of fault diagnosis.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may also be combined, the steps may be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A hydropower unit fault identification method based on ICEEMDAN-VMD-GAF and PCNN-BiGRU, characterized in that: The following steps are involved: S1. Collect vibration signals of hydropower units under different working conditions; S2, decomposing the vibration signal by ICEEMDAN algorithm to obtain IMF components; S3, using VMD and wavelet denoising method to decompose and reconstruct the IMF components to obtain IMF reconstructed components; S4, transforming the IMF reconstruction components into a two-dimensional image using the GAF algorithm; S5. Use two-dimensional images to train the neural network for fault identification of hydropower units.

2. The method for identifying faults of a hydroelectric generator set according to claim 1, characterized in that: In step S2, before signal decomposition, white noise is superimposed on the collected original vibration signal, and the white noise is decomposed from the original vibration signal by the EMD algorithm: x i (t)=x(t)+n k (t), Where n k (t) is the white noise obtained by EMD decomposition, k is the decomposition sequence; x i (t) is the vibration signal of the hydropower unit after superimposing white noise, i is the data sequence; x(t) is the original vibration signal of the hydropower unit.

3. The method for identifying faults of a hydroelectric generator set according to claim 2, characterized in that: In step S2, the vibration signal of the hydropower unit after superimposing white noise is decomposed by ICEEMDAN. When the j-th order residual component becomes a monotonic function and the IMF component can no longer be screened out, the screening is stopped. At this time, j is the number of selected modal components: In the formula, IMF j For the ICEEMDAN algorithm, from x i The jth IMF component generated by the decomposition in (t), e i (t) is the average trend of the signal.

4. The method for identifying faults of a hydroelectric generator set according to claim 1, characterized in that: In step S3, the IMF component is first decomposed using the VMD algorithm, which specifically includes the following steps: performing Hilbert transform on the IMF component to obtain its analytical signal so as to obtain a one-way spectrum; moving the spectrum of the IMF component to the baseband position by multiplying the Hilbert transformed IMF component with the exponential function of the center frequency; applying the Gaussian function to the moved spectrum to calculate its bandwidth as an optimization target; Under the constrained variational model, the sum of the bandwidths of each IMF component is minimized, and the center frequency and component signal are updated and iterated continuously until the convergence condition is met or the maximum number of iterations is reached, and finally several decomposed IMF components are obtained.

5. The method for identifying faults of a hydroelectric generator set according to claim 4, characterized in that: In step S3, a wavelet threshold shrinkage denoising method is used to perform secondary decomposition and signal reconstruction, which specifically includes the following steps: performing multi-layer wavelet decomposition on the IMF component after VMD decomposition to obtain low-frequency approximate coefficients and high-frequency detail coefficients of each layer; processing the wavelet coefficients to shrink the coefficients less than the set threshold to zero, and subtract the threshold from the coefficients greater than the set threshold; using the processed wavelet coefficients to reconstruct the signal through inverse wavelet transform to obtain the IMF reconstructed component.

6. The method for identifying faults of a hydroelectric generator set according to claim 1, characterized in that: Step S4 specifically includes: S401, Data Standardization: For all data points of the IMF reconstruction component, calculate its minimum value min(x) and maximum value max(x); For each data point x in the IMF reconstruction component i , apply the following normalization formula to convert it into normalized form: S402, polar coordinate conversion: Normalize each data point As the point r i : Calculate the phase angle θ for each data point i : S403, Gram matrix generation: The polar coordinate transformation results in the generation of two different Gram matrices: GASF matrix and GADF matrix; GASF matrix element GASF i,j and GADF matrix elements GASDF i,j The calculation formulas are: GASF i,j =cos(θ i +θ j ), GADF i,j =cos(θ i -θ j ), In the formula, θ i and θ j are the phase angles of the ith and jth data points respectively; S404, two-dimensional image generation: directly convert the GASF matrix and the GADF matrix into grayscale images, where the values ​​of the matrix elements correspond to the grayscale levels of the image pixels; each IMF reconstruction component generates a pair of images, namely, a GASF image and a GADF image; and fuse the GASF image and the GADF image into one image.

7. The method for identifying faults of a hydroelectric generator set according to claim 1, characterized in that: In step S5, the hydropower unit fault identification neural network combines the PCNN network and the BiGRU network. The PCNN network is first used to extract the convolution features of the two-dimensional image, and then the extracted convolution features are input into the BiGRU network for classification, thereby realizing fault identification.

8. The method for identifying faults of a hydroelectric generator set according to claim 1, characterized in that: The different operating conditions include normal operating conditions and fault operating conditions. The fault operating conditions include dynamic and static friction, rotor imbalance, hydraulic imbalance, excessive bearing clearance, rotor misalignment, and eccentric vortex belts in the tailwater pipe.

9. A hydropower unit fault identification system based on the method according to any one of claims 1 to 8, characterized in that: Includes the following modules: Vibration data acquisition module, used to collect vibration signals of hydropower units under different working conditions; ICEEMDAN decomposition module, used to perform ICEEMDAN decomposition on the collected vibration signal to obtain the corresponding IMF components; The secondary decomposition and reconstruction module is used to perform VMD decomposition and wavelet denoising decomposition on the IMF components, and reconstruct the IMF according to the decomposed signals to obtain the IMF reconstructed components; GAF image generation module, used to transform the IMF reconstruction components into a two-dimensional image through the GAF algorithm; The network building and training module is used to build a hydropower unit fault recognition network and train the hydropower unit fault recognition network using the two-dimensional images output by the GAF image generation module.

10. The hydropower unit fault identification system according to claim 9, characterized in that: The vibration data acquisition module uses a CRY2301 noise sensor to collect vibration data of the hydropower unit under various working conditions, with a sampling frequency of 10240 Hz and a single sampling time of 1 s.

Images