A monitoring method for hydro-generator sets based on swing data mining

Abstract

The present invention discloses a monitoring method for a hydro-turbine generator set based on swing data mining, including the establishment of a model: obtaining the swing information of the hydro-turbine set to be monitored; conducting diversified analysis of signals to obtain multivariate quantitative data of the signals; then conducting standardization processing on the above analysis results; performing dimensionality reduction processing on multivariate information entropy feature vectors; training a monitoring model based on the auto-encoding features of normal samples and abnormal samples; establishing multiple groups of models within different water head ranges; and then applying the model for monitoring according to actual needs. The present invention analyzes the swing signal of the daily monitoring of the hydro-turbine generator set, constructs its multivariate feature set through signal processing and analysis, conducts unsupervised learning on it based on deep learning theory and obtains the swing feature set, and then establishes a monitoring model of the unit based on steady-state data, and finally realizes the monitoring of the hydro-turbine generator set.

Description

Technical Field

[0001] The invention belongs to the technical field of hydraulic machinery state monitoring and fault diagnosis, and in particular relates to a monitoring method for a hydro-generator set based on swing data mining. Background Art

[0002] In recent years, renewable energy represented by photovoltaics and wind power will still be the main source of growth in electricity consumption. As the core equipment for energy conversion and grid regulator, the importance of monitoring and maintenance of hydropower generator sets is self-evident.

[0003] Hydroelectric generators are indispensable and important equipment for the hydropower industry. Their working principle is to convert various water flow energies into mechanical energy of rotor rotation, and then convert them into corresponding electrical energy through electromagnetic induction. Once a fault occurs in a hydroelectric generator set, it will lead to serious catastrophic consequences. Therefore, it is required to detect unit abnormalities as early as possible in daily operation so that effective preventive measures can be taken before the fault occurs. There are the following difficulties in the condition monitoring of hydroelectric generator sets: First, the individual working environments of hydroelectric generator sets are different and vary greatly from each other, so it is difficult to directly establish a quantitative monitoring model through a certain type of characteristic value; at the same time, the amount of information in the monitoring data of hydroelectric generator sets is large, and it is difficult to conduct in-depth and comprehensive research on it. It is difficult and inefficient to rely solely on manual selection; in addition, there are many types of data analysis and feature calculation methods for existing unit monitoring signals, but there is no consensus on the pros and cons of performance. The selection of specific methods still relies heavily on subjective experience and is blind. These factors restrict the development and improvement of monitoring technology for hydroelectric generator sets. In order to detect or predict the abnormal state of the hydro-turbine generator set as early as possible and guide the scientific development of maintenance work, powerful data mining methods and monitoring means are currently needed to realize the analysis and application of massive data of the hydro-turbine generator set. Therefore, the current monitoring of the hydro-turbine generator set needs to be further improved. Summary of the invention

[0004] The purpose of the present invention is to provide a monitoring method for a hydro-generator set based on swing data mining, which solves the problem that the current data analysis and application of hydro-generators and monitoring methods need to be further improved.

[0005] The technical solution adopted by the present invention is:

[0006] A monitoring method for a hydro-turbine generator set based on swing data mining includes establishing a monitoring model for the hydro-turbine generator set, and specifically includes the following steps:

[0007] Step 1: Obtain the swing signals at the upper guide, lower guide and water guide of the hydro-generator set to be monitored, wherein each swing signal includes signals collected from two mutually perpendicular directions;

[0008] Step 2: preprocessing the swing signal including smoothing filtering and standardization to obtain a preprocessed signal;

[0009] Step 3: Analyze the swing signal, use the adaptive analysis method to obtain the intrinsic mode component group, use the short-time Fourier transform to obtain the corresponding spectrum diagram, use the continuous wavelet transform to obtain the corresponding spectrum diagram, and use the symbolic analysis method to convert the original signal into a symbol sequence. At this point, the preprocessed signal is analyzed into four new information forms, and four types of analytical signals are obtained;

[0010] Step 4: According to the following formula (1), the four types of analytical signals are quantized with a unified quantization rule. The symbolized sequence S is obtained after quantization. i (y i ), where y i represents the resolved component;

[0011]

[0012] Where N is the length of the analytical signal, μ is the analytical component signal y i The mean of , α is the conditional parameter;

[0013] After the quantitative calculation is completed, all different forms of information will be normalized to the same interval, achieving the formal unification of various analytical results;

[0014] Step 5: Calculate the multivariate quantized data and solve the corresponding multivariate information entropy feature vector; for the symbol sequence S i (y i ) Construct substring set W i , W i With S i (y i ) is shown in the following formula (2):

[0015] W i =(S i ,S i+1 ,S i+2 ) (2),

[0016] W i There are a total of sixty-four types, and then the probability of each type appearing is calculated, recorded as P j , and then calculate its corresponding information entropy through the following formula (3):

[0017]

[0018] Step 6: Construct an autoencoder network to reduce the dimension of the multi-source information entropy feature vector. The autoencoder network consists of an input layer, four hidden layers, and an output layer. The number of nodes in hidden layer 1 is the same as that in hidden layer 4, and the number of nodes in hidden layer 2 is the same as that in hidden layer 3.

[0019] Step 7: After the construction is completed, the autoencoder network is trained. The construction judgment basis of the autoencoder network is to ensure that the multivariate information entropy after decoding is consistent with the multivariate information entropy feature vector, and the autoencoder feature is obtained through the autoencoder;

[0020] Step 8: Extract the autoencoder features from the autoencoder network, including normal samples and various abnormal samples, to achieve the purpose of dimensionality reduction, and build a monitoring model based on the autoencoder features;

[0021] Step 9: Perform dimensionality reduction calculations on the sample data in the stable operation area and various abnormal sample data at the same time to obtain their auto-encoding features, use the K-Means clustering algorithm to analyze normal samples and abnormal samples, extract the corresponding cluster center coordinates of the two, set the cluster center distance judgment function in the model, and judge the subsequent input auto-encoding features to complete the construction of the monitoring model;

[0022] Step 10: Based on the cluster center coordinates of the normal samples, calculate the distance judgment value between the normal samples and the cluster center coordinates, and use it as the judgment condition for judging whether the generator set is normal. The clustering results, judgment conditions, and cluster center coordinates constitute the monitoring model of the hydro-turbine generator set.

[0023] The monitoring method of the hydro-generator monitoring model includes:

[0024] When the monitoring model of the hydro-turbine generator set is used for monitoring, the self-encoding features of the swing information to be monitored are imported into the model, and the distance judgment value between the two is calculated. If the value is less than the normal feature range, it is considered that the unit to be detected is normal. The larger the distance value, the farther the information to be monitored is from the steady-state data range.

[0025] When further judging the fault type, the self-encoding features of the corresponding fault type samples are needed to form the corresponding cluster center coordinates to further judge the fault type.

[0026] The present invention is also characterized in that:

[0027] In steps 9-10, the monitoring model has the ability to update and correct. As the historical sample data of the unit operation becomes richer, the autoencoding features of the normal operation samples can be added, and the cluster center coordinates can be updated to cope with operating conditions such as unit repair, reconstruction, and replacement of components. However, when updating the cluster center coordinates, the existing parameters of the autoencoder network structure are not changed, and the dimensionality reduction process is not recalculated.

[0028] In step 10, the specific type of abnormal unit status or fault diagnosis is implemented, relying on the corresponding abnormal unit status or fault samples, otherwise the function cannot be implemented, but the function of monitoring whether the unit is abnormal always exists.

[0029] In steps 8-9, the autoencoder network realizes the dimensionality reduction of the multivariate information entropy feature vector, and extracts the autoencoder features from the middle of the autoencoder. After completing the establishment and training of the autoencoder network, the network model is saved for subsequent dimensionality reduction of the monitoring signal, and its internal parameters and structure are no longer changed.

[0030] Step 3, the adaptive component refers to the component obtained after the preprocessed signal is decomposed and calculated through EMD, EWT and VMD. The spectrum diagram refers to the time-frequency spectrum obtained after the preprocessed signal completes STFT and continuous wavelet transform. The symbolized sequence refers to the simplified representation obtained after the preprocessed signal is requantized. For each method, the parameters will be changed to obtain a series of different results to ensure full utilization of information.

[0031] In step 6, when the autoencoding characteristics are not ideal, the autoencoding network structure is adjusted according to the encoding and decoding results, and the adjustment refers to increasing or decreasing the number of hidden layer nodes.

[0032] In step 4, multivariate quantitative data refers to quantifying different forms of information in the same way. Since different forms of information need to be standardized and normalized first, data differences can be eliminated to achieve unified application of quantification rules.

[0033] In steps 1-10, for a specific hydro-turbine generator set, it is necessary to consider the change in head during the model building stage, that is, to divide the head change range into 3-5 different ranges, repeat steps 1-10 in each range, and finally obtain 3-5 monitoring models for different head ranges. The data to be analyzed must be consistent with the head range when the model is established.

[0034] The beneficial effects of the present invention are as follows: the monitoring method of a hydro-turbine generator set based on swing data mining of the present invention (1) realizes unsupervised analysis of swing information, realizes autonomous feature acquisition, and gets rid of the guidance and constraints of artificial experience in the process of calculating features;

[0035] (2) Provide a new method for calculating the characteristics of the swing signal, realize a personalized monitoring method for hydro-generator units, and obtain different characteristics for different units;

[0036] (3) By constructing a multivariate feature set of signals and integrating a variety of signal processing and transformation methods, the swing data can be fully mined from multiple angles;

[0037] (4) Introduce machine learning methods to establish models, improve the generalization ability and nonlinear learning ability of the model, and enhance the automation level of turbine monitoring;

[0038] (5) Finally, the status monitoring of the hydro-generator set is realized based on the swing characteristics;

[0039] (6) Apply deep learning methods in the process of obtaining feature values ​​to improve data mining and feature representation capabilities;

[0040] (7) The monitoring model uses a mean clustering method with higher computational efficiency to improve computational efficiency and speed, which is easy to implement in applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a model schematic diagram of a monitoring method of a hydro-generator set based on swing data mining of the present invention;

[0042] Figure 2 It is a schematic diagram of the principle of constructing an auto-encoding network in a monitoring method of a hydro-generator set based on swing data mining according to the present invention;

[0043] Figure 3 It is a schematic diagram of the principle of a monitoring model in a monitoring method for a hydro-turbine generator set based on swing data mining according to the present invention. DETAILED DESCRIPTION

[0044] A monitoring method for a hydro-generator set based on swing data mining of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0045] like Figure 1 , Figure 2 as well as Figure 3 As shown, the present invention provides a monitoring method for a hydro-turbine generator set based on swing data mining, including a unit to be monitored, swing information, a preprocessed signal, an adaptive component, an STFT spectrum diagram, a wavelet analysis spectrum diagram, a symbolized sequence, multivariate quantized data, a multivariate information entropy feature vector, an autoencoder network, an autoencoder feature, a monitoring model, an input layer, a hidden layer, an output layer, a decoded multivariate information entropy, unit swing data under different water heads, a K-means clustering model, a normal sample clustering center, and an abnormal sample clustering center.

[0046] Specifically: identify the unit to be tested and obtain the swing signal of the unit during steady-state operation; conduct data mining research on the swing signal and perform preprocessing to obtain the preprocessing signal; the analytical methods here include four categories, namely, adaptive analysis of the preprocessing signal, time-frequency analysis, and symbolic sequence analysis to obtain adaptive components, STFT spectrum diagrams, wavelet analysis spectrum diagrams, and symbolic sequences respectively; then, through a unified quantization rule, the above results are combined to form multivariate quantized data, and then the corresponding multivariate information entropy feature vector is calculated; an autoencoder network is established, including an input layer, a hidden layer, and an output layer, and a deep learning method is used to obtain the autoencoding features of the swing signal, and the decoded multivariate information entropy obtained by decoding the autoencoder is used to verify the validity of the autoencoding features; relying on the established The autoencoder network model extracts the autoencoder features of the multivariate information entropy feature vector; cluster analysis is performed using the autoencoder features of normal samples and abnormal samples to form a monitoring model, which includes a K-means clustering model and a normal sample clustering center; after completing the above steps, the autoencoder network is used to obtain the autoencoder features of the monitoring swing signal, and the normal sample clustering center is used to determine whether the input signal is within the normal range; the fault data type is refined to obtain the clustering center of abnormal samples under various fault modes, thereby realizing fault classification, but this function must have known category fault data samples of the unit to be tested; the unit swing data under different water heads are obtained within different water head ranges, and then multiple groups of models are established. In monitoring and diagnosis, the discriminant model corresponding to the water head input of the unknown signal is used.

[0047] The steps of a monitoring method for a hydro-generator set based on swing data mining of the present invention are as follows:

[0048] Obtain the swing information of the turbine unit to be monitored, mainly the swing information of the upper guide, lower guide and water guide. Each signal should contain two mutually perpendicular directions, including steady-state data and abnormal data under different water heads. Preprocess the data to achieve filtering and noise reduction of the swing signal, divide the samples, and sort the samples according to different water heads to form preprocessed signals.

[0049] Conduct diversified analysis of signals, obtain multivariate quantitative data of signals, and conduct comprehensive pre-processing signal analysis from multiple angles, including adaptive analysis in the time domain, spectrum analysis in the frequency domain, continuous wavelet transform spectrum, signal symbolization sequence, etc.

[0050] The above analysis results are then standardized and converted into multivariate quantitative data according to a unified quantification rule. The information entropy value is calculated according to a unified method to form a multivariate information entropy feature vector, which is used as the original feature set to characterize the unit status.

[0051] The multivariate information entropy feature vector is subjected to dimensionality reduction. The dimensionality reduction method adopts the autoencoder method. The constructed autoencoder network includes an input layer, a hidden layer, and an output layer. The multivariate information entropy feature vector passes through the input layer and the hidden layer and is reduced to an autoencoder feature.

[0052] The autoencoder feature passes through the hidden layer and the output layer to become the multivariate information entropy after decoding. The difference between the multivariate information entropy after decoding and the multivariate information entropy feature vector is used as a criterion to judge whether the autoencoder network is successfully established. If successful, the autoencoder network is retained and used to calculate the autoencoder feature of the new input signal to achieve the dimensionality reduction of the multivariate information entropy feature vector.

[0053] The monitoring model is trained based on the auto-encoding features of normal samples and abnormal samples. Here, the K-means clustering method is used to establish the monitoring model. After completing the cluster analysis, the coordinates of the normal sample cluster center are extracted from the model as the core judgment basis of the monitoring model.

[0054] Multiple groups of models are established within different water head ranges to adapt to different water head ranges, that is, monitoring models are established for a turbine within different water head ranges. The monitoring status includes: stable operation status and various abnormal working conditions. When each unit uses the present invention again, a new monitoring model needs to be re-established;

[0055] When using the monitoring model, the unknown swing information to be detected is processed to obtain its auto-encoding features. According to the different water heads, it is input into the trained model. According to the distance between its auto-encoding features and the coordinates of the normal sample cluster center, it is judged whether it is in a steady state. According to the size of the distance, the abnormality of the unit can be judged, so as to determine whether the unit has potential dangers or abnormal phenomena, and prompt the monitoring personnel to shut down the unit for inspection in time;

[0056] The fault or abnormality type can also be determined based on the cluster center coordinates of the abnormal samples. However, since the number of abnormal samples is relatively small in practice, the method of using the present invention needs to be selected based on the actual situation.

[0057] The following is a further detailed description of a monitoring method for a hydro-generator set based on swing data mining according to the present invention through specific embodiments.

[0058] A monitoring method for hydro-turbine generator sets based on swing data mining, the signal acquisition and preprocessing scheme is as follows: identify the unit to be detected, obtain the swing signal of the unit during steady-state operation, obtain the swing signal of the upper guide, lower guide and water guide of the unit to be monitored, each signal needs to contain signals collected from two mutually perpendicular directions, here it is necessary to clarify the water head corresponding to the information obtained, so as to establish different models according to different water head ranges; preprocess the swing signal, which mainly includes smoothing filtering and standardization, and the smoothing process is mainly based on the sliding average method and the Savitzky-Golay smoothing filter method, aiming to achieve smoothing under the premise of a larger window length; standardization is performed to prevent special points and extreme points in the signal from having an adverse effect on the data, and zero-mean normalization is used here:

[0059]

[0060] Where X(i) is the i-th data in the standardized signal, x i is the i-th data in the original signal, is the original signal mean, and σ is the standard deviation of the original signal.

[0061] A monitoring method for a hydro-turbine generator set based on swing data mining, the signal analysis scheme is as follows: the standardized signal is analyzed mainly from three aspects: adaptive analysis, time-frequency analysis, and symbolic sequence analysis; the adaptive analysis method is empirical mode decomposition, each pre-processed signal is analyzed into a group of intrinsic mode components, and 1-3 components are selected;

[0062] The methods used for time-frequency analysis are short-time Fourier transform and continuous wavelet transform. Short-time Fourier transform is a Fourier transform of the standardized signal s(τ) multiplied by a window function g(τ-t) centered at t, thereby obtaining the local time-varying spectrum characteristic image of the signal;

[0063]

[0064] Where t is time, f is vibration frequency, s(τ) represents the original signal, g(τ-t) represents the window function, e -j2πfτ The frequency components decomposed from the standardized signal.

[0065] Continuous wavelet transform is to obtain the time-frequency image of the corresponding wavelet time information and frequency characteristics by performing operations such as translation and scaling of the wavelet basis function on the standardized signal;

[0066]

[0067]

[0068] Among them, w(a,b) represents the wavelet transform coefficient; ψ a，b (t) is the wavelet basis function transformed by the scale factor a and the expansion factor b.

[0069] In order to make the time-frequency image features obtained by the two time-frequency analysis methods more prominent, the image processing algorithm is used to convert the three primary color pixels of the time-frequency image into the corresponding 8-level grayscale image by averaging. The conversion rules are as follows:

[0070] I(x,y)=1 / 3*I_R(x,y)+1 / 3*I_G(x,y)+1 / 3*I_B(x,y) (5),

[0071] Where: I(x,y) is the grayscale image obtained, I_R(x,y) is the pixel point of the red channel in the time-frequency image, I_G(x,y) is the pixel point of the green channel in the time-frequency image, and I_B(x,y) is the pixel point of the blue channel in the time-frequency image.

[0072] Symbolic sequence analysis uses the extreme value separation method to obtain the extreme value sequence of the signal. The local maximum and local minimum of the original signal X(i) are extracted to form the extreme value sequence Y(i). The binary conversion is performed on it, with the mean value P0 of the preprocessed signal as the boundary. The conversion rule is:

[0073]

[0074] Subsequently, Y(i) is traversed continuously, and three consecutive data points are taken as a unit and converted into decimal. That is, for the i-th Y(i), Y(i), Y(i+1), and Y(i+2) are combined into a 3-bit binary number, which is converted into a decimal number. All numbers in Y(i) are converted in the same way to obtain a decimal sequence with a length of N-2. After completing the above research, a series of parsed results will be obtained for subsequent feature set calculations.

[0075] A monitoring method for a hydro-turbine generator set based on swing data mining, the unified quantization and feature calculation process of the analytical signal is as follows: if the symbolic sequence S is obtained by quantization i (y i ), where y i Represents the parsed component, which represents the eigenmode component, the gray value sequence of the time-frequency diagram, the symbolized decimal sequence, etc.

[0076]

[0077] Among them, the length of the analytical signal is N, where 1≤i≤N, where μ represents the analytical component signal y iThe mean of α is the conditional parameter, which is initially 0.1. After the quantization calculation is completed, all different information forms will be normalized to the same interval to achieve the formal unification of various analytical results. The multivariate quantized data is calculated to solve the corresponding multivariate information entropy feature vector. For the symbol sequence S i (y i ) construct a substring set W of length 3 i , W i With S i (y i ) is as follows, W i The length of the sequence is N-2:

[0078] W i =(S i ,S i+1 ,S i+2 ) (8),

[0079] Visible W i There are 64 types in total, and then the probability of each type appearing is calculated, recorded as P j , where 1≤j≤64, and then calculate its corresponding information entropy

[0080]

[0081] A method for monitoring a hydro-turbine generator set based on swing data mining, wherein the autoencoder network construction and feature dimension reduction process are as follows: constructing an autoencoder network to reduce the dimension of a multi-source information entropy feature vector, the autoencoder network consists of one input layer, four hidden layers 1-4, and one output layer, wherein the number of nodes of hidden layer 1 is the same as that of hidden layer 4, and the number of nodes of hidden layer 2 is the same as that of hidden layer 3; after the construction is completed, the autoencoder network is trained, the construction judgment basis of the autoencoder network is to ensure that the multivariate information entropy after decoding is consistent with the multivariate information entropy feature vector, and the autoencoder feature is obtained through the autoencoder; the autoencoder feature is extracted from the autoencoder network, which should include normal samples and various abnormal samples to achieve the purpose of dimension reduction, and the input layer, hidden layer 1 and hidden layer 2 of the trained autoencoder network structure are stored for calculating the relevant features of the newly collected signal at any time; the number of nodes in each layer of the autoencoder network can be increased or decreased according to the actual situation, but its symmetric structure remains unchanged, and the basic framework remains unchanged.

[0082] A monitoring method for a hydro-turbine generator set based on swing data mining, wherein the model establishment process is as follows: dimensionality reduction calculation is performed on sample data in a stable operation area and various abnormal sample data at the same time to obtain their auto-encoding features, a K-Means clustering algorithm is adopted to analyze normal samples and abnormal samples, and the corresponding cluster center coordinates of the two are extracted, a cluster center distance judgment function is set in the model, and subsequent input auto-encoding features are judged to complete the construction of a monitoring model; based on the cluster center coordinates of normal samples, the distance judgment value between the normal samples and the cluster center coordinates is calculated, and it is used as a judgment condition for judging whether the unit is normal, and the clustering results, judgment conditions, and cluster center coordinates constitute a monitoring model; the monitoring model is applied to The self-encoding features of the swing information to be monitored are imported into the model, and the distance judgment value between the two is calculated. If the value is less than the normal feature range, it is considered that the unit to be detected is normal. The larger the distance value, the farther the information to be monitored is from the steady-state data range. To further judge the fault type, the self-encoding features of the corresponding fault type samples are required to form the corresponding cluster center coordinates, and then judge the fault type. The normal samples involved in the above content refer to the swing signals of the unit in the stable operation area, and the abnormal samples involved in the above content refer to the signals when the unit enters the strong vibration operation area or the restricted operation area, and also include faults such as friction, misalignment, and imbalance of the mechanical main shaft, including abnormal conditions of low-frequency vortex belts through the tail water pipe.

[0083] A monitoring method for a hydro-turbine generator set based on swing data mining, wherein the model is used in the following manner: the working head variation range of the unit to be detected is divided into 3-5 intervals, the data used and the model to be established need to be within the corresponding interval, and the monitored signal must also be input into the model of the corresponding interval to be meaningful; for a certain diagnostic model, all abnormal samples can be set as one category, and normal samples can be set as another category, and the model is used to determine whether the unit is in an abnormal state; normal samples can also be set as one category, and each abnormal state can be set as multiple categories, and in this case, the model can not only monitor, but also have a certain abnormal type recognition and diagnosis capability.

[0084] Based on the swing data, the present invention proposes a monitoring method for a hydro-generator set based on the swing data mining. The reason why the present invention selects the swing data is that it is widely used as a conventional monitoring method of current hydropower stations. The swing data has a strong characterization ability for the mechanical structure of the unit, and this type of data is easy to obtain; the present invention combines the deep learning method to construct a monitoring model for the hydro-generator set, which to a certain extent gets rid of the constraints of artificial experience on feature selection;

[0085] The present invention establishes a rich feature representation form through information expansion and analysis of the swing information, and relies on the autoencoder method in deep learning theory to achieve dimensionality reduction, thereby achieving a highly concise representation of the swing information and establishing a monitoring model to achieve the abnormal state monitoring task of the turbine unit;

[0086] The present invention analyzes the swing signal of daily monitoring of the hydro-turbine generator set, constructs its multivariate feature set through signal processing and analysis, conducts unsupervised learning on it based on deep learning theory and obtains the swing feature set, and then establishes a unit monitoring model according to the steady-state data, ultimately realizing the monitoring of the hydro-turbine generator set.

Claims

1. A monitoring method for a hydro-generator set based on swing data mining, characterized in that: Including the establishment of a monitoring model for a hydro-generator set, specifically including the following steps: Step 1: Obtain the swing signals at the upper guide, lower guide and water guide of the hydro-generator set to be monitored, wherein each swing signal includes signals collected from two mutually perpendicular directions; Step 2: preprocessing the swing signal including smoothing filtering and standardization to obtain a preprocessed signal; Step 3: Analyze the swing signal, use the adaptive analysis method to obtain the intrinsic mode component group, use the short-time Fourier transform to obtain the corresponding spectrum diagram, use the continuous wavelet transform to obtain the corresponding spectrum diagram, and use the symbolic analysis method to convert the original signal into a symbol sequence. At this point, the preprocessed signal is analyzed into four new information forms, and four types of analytical signals are obtained; Step 4: According to the following formula (7), the four types of analytical signals are quantized with a unified quantization rule. If the symbolized sequence S is obtained after quantization i (y i ), where y i represents the resolved component; Where N is the length of the analytical signal, where 1≤i≤N, and μ is the analytical component signal y i The mean of , α is the conditional parameter, initially 0.1; After the quantitative calculation is completed, all different forms of information will be normalized to the same interval, achieving the formal unification of various analytical results; Step 5: Calculate the multivariate quantized data and solve the corresponding multivariate information entropy feature vector; for the symbol sequence S i (y i ) Construct substring set W i , W i With S i (y i ) is shown in the following formula (8): W i =(S i ,S i+1 ,S i+2 ) (8), Determine W i The total number of types, and then calculate the probability of each type appearing, recorded as P j , and then calculate its corresponding information entropy through the following formula (9): Step 6: Construct an autoencoder network to reduce the dimension of the multi-source information entropy feature vector. The autoencoder network consists of an input layer, four hidden layers, and an output layer. The number of nodes in hidden layer 1 is the same as that in hidden layer 4, and the number of nodes in hidden layer 2 is the same as that in hidden layer 3. Step 7: After the construction is completed, the autoencoder network is trained. The construction judgment basis of the autoencoder network is to ensure that the multivariate information entropy after decoding is consistent with the multivariate information entropy feature vector, and the autoencoder feature is obtained through the autoencoder; Step 8: Extract the autoencoder features from the autoencoder network, including normal samples and various abnormal samples, to achieve the purpose of dimensionality reduction, and build a monitoring model based on the autoencoder features; Step 9: Perform dimensionality reduction calculations on the sample data in the stable operation area and various abnormal sample data at the same time to obtain their auto-encoding features, use the K-Means clustering algorithm to analyze normal samples and abnormal samples, extract the corresponding cluster center coordinates of the two, set the cluster center distance judgment function in the model, and judge the subsequent input auto-encoding features to complete the construction of the monitoring model; Step 10: Based on the cluster center coordinates of the normal samples, the distance judgment value between the normal samples and the cluster center coordinates is calculated, and it is used as the judgment condition for judging whether the generator set is normal. The clustering results, judgment conditions, and cluster center coordinates constitute a monitoring model for the hydro-generator set; The monitoring method of the hydro-generator monitoring model includes: When the monitoring model of the hydro-generator unit is used for monitoring, the self-encoding characteristics of the swing information to be monitored are imported into the model, and the distance judgment value between the two is calculated. If the value is less than the normal characteristic range, it is considered that the unit to be detected is normal. The larger the distance value, the farther the information to be monitored is from the steady-state data range. When further judging the fault type, the self-encoding features of the corresponding fault type samples are needed to form the corresponding cluster center coordinates to further judge the fault type.

2. A monitoring method for a hydro-generator set based on swing data mining according to claim 1, characterized in that: In steps 9-10, the monitoring model has the ability to update and correct. As the historical sample data of the unit operation becomes richer, the autoencoding features of the normal operation samples are added, and the cluster center coordinates are updated to cope with the operating conditions of the unit repair, reconstruction, and replacement of components. However, when updating the cluster center coordinates, the existing parameters of the autoencoder network structure are not changed, and the dimensionality reduction process is not recalculated.

3. The monitoring method of a hydro-generator set based on swing data mining according to claim 1 is characterized in that: In step 10, the specific type of abnormal unit status or fault diagnosis is implemented, relying on the corresponding abnormal unit status or fault samples, otherwise the function cannot be implemented, but the function of monitoring whether the unit is abnormal always exists.

4. The monitoring method of a hydro-generator set based on swing data mining according to claim 1 is characterized in that: In steps 8-9, the autoencoder network realizes the dimensionality reduction of the multivariate information entropy feature vector, and extracts the autoencoder features from the middle of the autoencoder. After completing the establishment and training of the autoencoder network, the network model is saved for subsequent dimensionality reduction of the monitoring signal, and its internal parameters and structure are no longer changed.

5. The monitoring method of a hydro-generator set based on swing data mining according to claim 1 is characterized in that: In step 3, the adaptive component refers to the component obtained after the preprocessed signal is decomposed and calculated through EMD, EWT and VMD, the spectrum diagram refers to the time-frequency spectrum obtained after the preprocessed signal completes STFT and continuous wavelet transform, and the symbolized sequence refers to the representation obtained after the preprocessed signal is requantized; for each method, the parameters will be changed to obtain a series of different results.

6. The monitoring method of a hydro-generator set based on swing data mining according to claim 1 is characterized in that: In step 6, when the autoencoding characteristics are not ideal, the autoencoding network structure is adjusted according to the encoding and decoding results, and the adjustment refers to increasing or decreasing the number of hidden layer nodes.

7. The monitoring method of a hydro-generator set based on swing data mining according to claim 1 is characterized in that: In step 4, multivariate quantitative data refers to quantifying different forms of information in the same way. Since different forms of information need to be standardized first, data differences will be eliminated to achieve unified application of quantification rules.

8. The monitoring method of a hydro-generator set based on swing data mining according to claim 1 is characterized in that: In steps 1-10, for a specific hydro-turbine generator set, it is necessary to consider the change in head during the model building stage, that is, to divide the head change range into 3-5 different ranges, repeat steps 1-10 in each range, and finally obtain 3-5 monitoring models for different head ranges. The data to be analyzed must be consistent with the head range when the model is established.

Images