A method for early warning of eccentric wear of guide bearing bushes of hydropower units

Abstract

The invention discloses an early warning method for eccentric wear of a guide bearing bush of a hydropower unit, and relates to the field of electronic technology. According to characteristic signals such as vibration swing, vibration, bush temperature, etc. installed on a hydropower generator unit, the invention adopts wavelet packet filtering and Hilbert transform method to perform envelope demodulation to identify impact characteristics caused by eccentric wear, combines with an SVR machine learning method to identify bush temperature anomalies, and realizes identification, early warning, and detection of eccentric wear of the guide bearing bush of the unit by comparing the impact orientation and the bush temperature change at the corresponding orientation.

Description

Technical Field

[0001] The invention relates to the field of electronic technology, and in particular to a method for early warning of eccentric wear of a guide bearing bush of a hydroelectric generator set. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] The so-called eccentric wear is a kind of dynamic-static friction, which refers to the friction occurring in a certain local position. The friction refers to the abnormal contact between the rotating parts and the stationary parts. In large rotating machinery such as hydro-turbine generator sets and steam turbine generator sets, the friction between the rotating parts and the stationary parts is a common fault. The main reason is that the dynamic-static clearance of the unit becomes smaller for some reason. The friction causes the rotor to produce very complex vibrations, which is an important reason for the instability of the rotor system. In mild cases, it causes the unit to vibrate strongly, and in severe cases, it can cause permanent bending of the shaft, or even the destruction of the entire shaft system.

[0004] Rub / eccentric wear is a concomitant fault, which is a secondary effect caused by other faults. Rub / eccentric wear contact can occur in radial, axial or both directions. When the actual rub contact occurs in a small part of the vibration cycle, it is called local rub. When the contact occurs throughout the entire vibration cycle and is continuous, it is called full-circle rub. For rub faults, local rub is the most common form of rub, which includes three physical phenomena: collision, friction and change in shaft stiffness.

[0005] For hydro-turbine generator sets, the friction occurs in the following places:

[0006] 1) Eccentric wear between the main shaft and the guide bearing caused by the non-concentricity of the bearing shell and the eccentricity of the main shaft.

[0007] The first aspect is the installation deviation of the unit bearing itself. When the unit bearing clearance is uneven or eccentric, when the unit swing is greater than the minimum clearance, the bearing will be worn eccentrically. For units with three guide bearings, when the three guides are not concentric, the bearing will also be worn eccentrically at the non-concentric parts.

[0008] The second aspect is that for the rigidly supported thrust bearing structure, when the horizontal adjustment deviation of the thrust bearing is large, it will also cause the bearing bearing to wear eccentrically during the operation of the unit;

[0009] The third aspect is the bearing displacement caused by the displacement of the rotation center of the guide bearing due to the fixed-direction force during the operation of the unit. This force mainly comes from two parts. One is the imbalance of the fixed-direction magnetic pull caused by the non-concentricity of the stator and rotor of the unit, and the other is the hydraulic imbalance caused by the uneven water flow along the circumference of the impeller.

[0010] 2) Friction between the unit's impeller blades and the impeller chamber.

[0011] The first aspect is that the displacement of the unit shaft causes the gap between the blades and the runner chamber to become smaller, resulting in friction between the runner chamber and the runner. This friction usually occurs during the unit shutdown process or in the vibration area or non-coordinated area. When the unit is subjected to water hammer or unstable water flow impact, a significant axial displacement will cause the gap between the blades and the runner chamber to become smaller and cause friction;

[0012] The second aspect is that the runner chamber is deformed (oval or other irregular shapes) due to long-term uneven water flow or other reasons. On the one hand, it will cause radial hydraulic imbalance, resulting in radial displacement of the runner rotation center, causing the gap between the blades and the runner chamber to become locally smaller, thus causing friction;

[0013] The third aspect is that due to the deviation of the guide vane installation angle of the unit or the inconsistent operation of the guide vane operating mechanism, such as the guide vane being stuck or blocked by foreign objects, the unit generates a more obvious radial water thrust during the load adjustment process, causing the gap between the blade and the impeller chamber to partially decrease, resulting in friction.

[0014] 3) Friction between the upper crown and the labyrinth ring of the mixed flow unit impeller.

[0015] For mixed flow turbines, the gap between the runner crown and the labyrinth ring is small, usually only 2-4mm. Under long-term water flow impact and sediment wear, the labyrinth ring will become loose or even fall off, causing friction between the runner crown and the labyrinth ring.

[0016] The dynamics of a rubbing rotor are complicated by the fact that rubbing causes nonlinear changes in both force and dynamic stiffness. When rubbing occurs, contact forces appear and disappear suddenly. When the rotating part contacts the stationary part, the stationary part pushes the rotor, while the rotating part pushes the stationary part with equal but opposite forces. This contact force can be decomposed into radial and tangential (friction) components. The radial force acts in the direction of the rotor center and thus significantly accelerates the rotating part away from the contact point. The radial force changes during the dwell time of the contact cycle and is proportional to the instantaneous radial acceleration a (F = Ma). After a short time, the rotating part deviates due to its inertia, the average radial velocity decreases, and after the radial velocity reaches zero, the rotating part rebounds and the rubbing contact ends.

[0017] During contact, tangential friction occurs, which is equal to the instantaneous magnitude of the radial force multiplied by the friction factor of the contact surface. The direction of the tangential friction force is opposite to the surface speed of the rotating part, generating torque on the rotating part while trying to accelerate the rotating part in the opposite direction of movement. This is why in rubbing, reverse components are usually generated in the full spectrum. The side effect of tangential friction is that it acts as a medium to transfer rotational energy to radial vibration. For local rubbing, if a slight friction contact is generated, it will not significantly change the system force and the overall stiffness of the system, but as the degree of rubbing increases, the change in force and stiffness will become more obvious. The radial and tangential friction forces caused by rubbing are essentially impact forces, which are characterized by sudden appearance, great force, and then sudden disappearance. The effect is similar to hitting the rotor shaft with a hammer. When the rubbing is severe to a certain extent, this impact force can produce an impact response, which includes multiple free radial vibration modes in the rotating part. When the rotating part bounces off the contact, it will vibrate at one or more natural frequencies in the form of free vibration, which is also a major reason why rubbing will produce resonance.

[0018] The eccentric wear of the guide bearing is a kind of local friction. When the eccentric wear occurs, it will cause radial and tangential impact forces, generate impact response, and cause instantaneous local resonance signals. Generally speaking, this kind of friction is a transient periodic weak friction with transient characteristics. The local deformation and rebound movement caused by the impact force of this friction make the friction vibration signal contain a large number of non-synchronous harmonic components, usually accompanied by resonance signals. Friction and impact make local friction very nonlinear. For a long time, the main basis for people to diagnose friction faults is the characteristic spectrum, but it often has similar spectral characteristics to vibration faults caused by other reasons. At the same time, the transient impact characteristics of the time domain where the friction is located are diluted after Fourier transform, which is equivalent to averaging the local characteristic information in the entire analysis frequency domain, so that the analysis result has a large error. Therefore, it is very difficult to accurately diagnose friction faults based on the spectrum alone.

[0019] In addition, from the perspective of signal analysis, the resonance signal caused by instantaneous friction usually appears in the high frequency band, and the signal is often modulated by the periodic 1X vibration signal. The signal amplitude is not very strong, so the signal is often submerged in other periodic new vibration signals and other noise signals.

[0020] Therefore, the peak-to-peak value, effective value and other time domain characteristic parameters cannot identify the collision fault between the components and the rotating components, and it is also difficult to separate and identify the impact signal from the frequency domain using conventional FFT changes. Figure 1This is the swing signal waveform after a local weak rubbing occurs in a water-guided bearing of a turbine in a power station (rubbing characteristics have appeared in the signal). In the time domain waveform, the swing change caused by weak rubbing is very small and is submerged in the 1X signal of the swing, and has no effect on the peak-to-peak value of the signal. Therefore, it is difficult to identify the local weak rubbing fault using conventional peak-to-peak values.

[0021] In addition, a feature that is significantly different from other weak friction is that the occurrence of eccentric wear of the guide bearing pad is accompanied by an increase in the temperature of the bearing pad, which is also an important feature of eccentric wear and other weak friction.

[0022] In summary, the characteristics of bearing eccentric wear fault in the signal are as follows:

[0023] (1) Eccentric wear will cause short-term high-frequency impact resonance signals;

[0024] (2) Stable eccentric wear will occur periodically, and the impact resonance signal caused by it will appear periodically. The eccentric wear will occur once or more during one rotation of the unit.

[0025] (3) Eccentric wear will cause a short-term high-frequency resonance signal that will be modulated by the speed frequency;

[0026] (4) The resonance signal can be detected by the swing and radial vibration measurement points, but it is often masked by other vibration signals and noise.

[0027] (5) Eccentric wear failure is usually accompanied by an increase in the corresponding bearing temperature.

[0028] The guide bearings of the hydroelectric generator set (such as upper guide bearings, lower guide bearings, and water guide bearings) are the supporting parts of the rotating parts of the hydroelectric generator set (including rotors, main shafts, runners, etc.). Their function is to support and position the rotating parts of the turbine, and they have good load-bearing capacity and sufficient strength. However, due to defects such as the tortuous axis of the unit, improper bearing clearance, and non-concentric guide bearings, eccentric wear problems will occur between the main shaft of the unit and the bearing surface. Long-term operation will not only cause bearing wear, damage, heat, and even burning of bearings, but will also cause adverse effects on the operation of the unit, including increased vibration and reduced stability. Therefore, it is particularly important to conduct online detection of the eccentric wear of the unit's guide bearings and identify the health status of the bearings for the safe operation of the unit. Summary of the invention

[0029] The purpose of the present invention is to provide an early warning method for eccentric wear of guide bearings of hydropower units in response to the problems existing in the prior art. According to characteristic signals such as vibration swing, vibration, bearing temperature, etc. collected on the hydropower generator unit, wavelet packet filtering and Hilbert transform methods are used for envelope demodulation to identify impact characteristics caused by eccentric wear. In combination with the SVR machine learning method, an abnormal bearing temperature is identified. By comparing the impact orientation and the bearing temperature changes at the corresponding orientation, the identification and early warning detection of eccentric wear of the guide bearing of the unit is achieved, thereby solving the above-mentioned problems.

[0030] The technical solution of the present invention is as follows:

[0031] A method for early warning of eccentric wear of guide bearing bushes of a hydropower unit, comprising:

[0032] Step S1: training based on historical health data samples to obtain a prediction model for the temperature change of each guide bearing bush and a maximum allowable bush temperature deviation;

[0033] Step S2: Based on wavelet packet bandpass filtering, the impact characteristic index of the guide bearing swing signal and vibration signal collected in real time is extracted through the impact envelope demodulation algorithm;

[0034] Step S3: Based on wavelet packet bandpass filtering, through the impact envelope demodulation algorithm, select multiple historical health data to accumulate the envelope average peak value in the historical sample data , which correspond to the upper guide X-direction swing signal, the upper guide Y-direction swing signal, the upper frame X-direction vibration signal, and the upper frame X-direction vibration signal respectively;

[0035] Step S4: Perform early warning detection of eccentric wear of the guide bearing bushing of the unit.

[0036] Furthermore, the guide bearing shafts of each part include: an upper guide bearing, a lower guide bearing, and a water guide bearing.

[0037] Furthermore, the step S1 comprises:

[0038] Step S11: extracting characteristic index data in a steady state after grid connection from the stored historical data records;

[0039] Step S12: Preprocess the characteristic index data extracted in step S11, remove abnormal points, and obtain the time series of new characteristic index data and the time series of the predicted features ;in, , , is the total number of feature indicators involved in the prediction, is the length of the time series of each characteristic indicator;

[0040] Step S13: Normalization processing: map all time series of feature indicators to the range of [-1, 1] to form a new time series ;

[0041] Step S14: For the new time series and the time series of the predicted features Randomly shuffle the order, and then extract the first 70% of the data as the training set time series and , the remaining 30% of the data is used as the test set time series and ;

[0042] Step S15: Call the training set time series and ,Bundle As independent variable, As the dependent variable, the SVR prediction model is input for training, and the prediction model vector data is obtained after training. ; The parameters of the SVR prediction model are set as follows: the penalty coefficient C=1, and the kernel function is selected as the Gaussian kernel function;

[0043] Step S16: The prediction model vector data obtained by training in step S15 Based on the test set time series and , bring it into the SVR prediction model, calculate the prediction error and the penalty coefficient;

[0044] Step S17: The prediction model vector data obtained by training in step S15 Based on and Bring in the SVR prediction model and calculate the error point by point;

[0045] Step S18: Repeat steps S11 to S17, traverse all upper guide bearing pad temperatures, lower guide bearing pad temperatures, and water guide bearing pad temperatures, obtain the SVR prediction model vector data and maximum allowable deviation of each pad temperature, and record them. is the vector data of a certain watt temperature SVR prediction model, is the maximum allowable deviation of the watt temperature, , It is the sum of the upper, lower and water-conducting tile temperatures.

[0046] Furthermore, the preprocessing in step S12 includes:

[0047] The time series of each characteristic indicator was preprocessed using Gaussian filtering with a 99.7% confidence interval;

[0048] The step S13 comprises:

[0049]

[0050] in:

[0051] For the The maximum value of the characteristic index time series, For the The minimum value of the time series of the characteristic indicators.

[0052] Furthermore, the step S16 comprises:

[0053]

[0054] In the above formula, is the actual measured value in the training set, is the average value of the actual measured values ​​in the training set, represents the predicted value, is the length of the test set time series, The closer the value is to 1, the better the fitting effect is and the smaller the model error is. If it is less than 0.8, it is necessary to readjust the training set, adjust the SVR training parameters, and re-execute step S11, otherwise, execute the next step.

[0055] Furthermore, the step S17 includes:

[0056]

[0057] In the above formula, is the actual measured value in the time series of the entire sample set, represents the predicted value, is the error value;

[0058] Then according to 1.25 times of the distribution is taken as the maximum allowable temperature deviation, that is:

[0059]

[0060] In the above formula, Indicates that from m Find it in the data Width range, record This is the allowable deviation of the bearing temperature.

[0061] Furthermore, the step S2 comprises:

[0062] Step S21: Obtain the collected real-time swing and vibration original time domain waveform signals , Key phase marks are recorded in the

[0063] Step S22: Select Morlet as the wavelet basis function, set the decomposition level to 4, Perform wavelet packet transform to extract wavelet packet coefficients of each scale , ;

[0064] Step S23: bandpass filtering based on wavelet packets to filter out low-frequency and high-frequency noise;

[0065] Step S24: Based on the wavelet packet coefficients after filtering Perform inverse wavelet packet transform to obtain the reconstructed signal and form new vibration and swing signals ;

[0066] Step S25: , the envelope waveform signal of the impulse signal is obtained by using the Hilbert transform digital envelope demodulation technology ;

[0067] Step S26: Extracting impact features.

[0068] Furthermore, the step S23 includes:

[0069] set up is the wavelet packet coefficient after filtering, then Calculated by the following formula:

[0070]

[0071] In the above formula, for The corresponding center frequency of the passband frequency;

[0072] Then, in order to reduce the noise impact of signals in the range of 20X to 300Hz, the following method is used to perform noise filtering:

[0073]

[0074] is the adaptive threshold after wavelet transformation, is an estimate of the noise level, calculated as:

[0075]

[0076] in, is the number of data points in the data window used in wavelet transform, The function represents the median value. is the soft threshold function:

[0077]

[0078] in, is the threshold parameter, is the coefficient after transformation, The input value is , The input value is ; The function is a symbolic function and its values ​​are as follows:

[0079]

[0080] Indicates finding the maximum value.

[0081] Furthermore, the step S25 includes:

[0082] set up is the analytical signal of the impact signal envelope signal, then:

[0083]

[0084] in, is an imaginary unit, Represents the Hilbert transform:

[0085]

[0086] According to the impulse signal envelope waveform, use fast Fourier transform to perform envelope spectrum recognition and extract envelope signal ,in, is the complex conjugate of the analytic signal:

[0087]

[0088] The step S26 comprises:

[0089] In obtaining Afterwards, according to the key phase signal in the waveform, the data of each cycle is intercepted, and the envelope peak value and azimuth of each cycle are calculated according to the data of each cycle;

[0090] from The waveform signal between two adjacent key phases is intercepted to generate the first Cycle Signal;

[0091] right Sort the data, remove the largest 50% of the data, and find the maximum value from the remaining 50% of the data as the noise baseline value in the signal ;

[0092] Constructing new time domain waveform ,right The signal is pulse peak detected to calculate the pulse amplitude caused by the impact, and the phase corresponding to the pulse peak is synchronously calculated according to the key phase signal;

[0093] Repeated calculation Record the number of shock pulses, pulse amplitude and pulse orientation in all cycles; is the total number of pulses, For the amplitude of each pulse, is the pulse direction, find:

[0094]

[0095] For envelope waveform signal Perform full-cycle FFT transformation to form an envelope spectrum, extract the main frequency, and record the maximum amplitude main frequency as .

[0096] Furthermore, the step S4 comprises:

[0097] Step S41: Based on the extracted envelope peak and orientation of each cycle, calculate the average pulse amplitude and basic pulse amplitude growth coefficient of each guide bearing swing and vibration ;

[0098] Step S42: All the impact signals of the guide bearing swing and vibration are judged as follows:

[0099] Condition 1: Judgment of swing impact index: , and the main frequency for The frequency multiplication, where is an integer, and , is the number of pads of the corresponding guide bearing;

[0100] Condition 2: Vibration and shock index judgment:

[0101] For the vibration and shock characteristic indicators of displacement output: , and the main frequency for Frequency doubling;

[0102] Vibration shock characteristic indicators for speed output: , and the main frequency for Frequency doubling;

[0103] Vibration and shock characteristic indicators for acceleration output: , and the main frequency for Frequency doubling;

[0104] If all the above conditions are not met, there is no eccentric wear of the guide bearing bush, and the process returns. Otherwise, proceed to the next step.

[0105] Step S43: Detection of abnormal bearing bush temperature:

[0106] Read the phase of the corresponding impact pulse from the vibration or swing signal that meets the above conditions ,according to Calculate the bearing number of the corresponding position and set the bearing number to ;

[0107] Real-time collection of bearings Temperature, setting For real-time bearing The measured value of temperature; at the same time, other input parameters used in the bearing temperature prediction model are collected; the prediction model vector data obtained by training in step S1 Based on the above input parameters, the SVR prediction model is used to calculate the predicted temperature and record it as ,remember The bearing obtained by training in step S1 The allowable deviation of the watt temperature is:

[0108] if , then the bearing Eccentric wear warning, bearing number is ;

[0109] Otherwise the bearing If eccentric wear is not established, step S43 is repeated until the impact pulse features in all directions are detected.

[0110] Compared with the prior art, the present invention has the following beneficial effects:

[0111] This method uses wavelet packet filtering and Hilbert transform methods to perform envelope demodulation to identify the impact characteristics caused by eccentric wear based on the characteristic signals of vibration swing, vibration, bearing temperature, etc. installed on the hydro-turbine generator set. It combines the SVR machine learning method to identify bearing temperature anomalies. By comparing the impact orientation and the bearing temperature changes at the corresponding orientation, the identification and early warning detection of eccentric wear of the guide bearing bearing of the unit can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 This is the swing signal waveform diagram of a water-guided bearing of a turbine in a power station after local weak friction occurs;

[0113] Figure 2 It is a schematic diagram of the SVR model;

[0114] Figure 3is a detailed flow chart of step S1;

[0115] Figure 4 A comparison chart of the temperature data predicted by the prediction model obtained after training using this method and the temperature of the tile obtained by actual measurement;

[0116] Figure 5 It is the error distribution diagram between the predicted temperature and the actual temperature;

[0117] Figure 6 is a detailed flow chart of step S2;

[0118] Figure 7 is the original top cover vibration signal of a unit;

[0119] Figure 8 is the signal processed by wavelet packet filtering;

[0120] Fig. 9 is the envelope signal of the impulse signal;

[0121] Fig.10 is a periodic envelope signal;

[0122] Fig.11 is the envelope spectrum;

[0123] Fig.12 is a detailed flow chart of step S4. DETAILED DESCRIPTION

[0124] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0125] First of all, it should be noted that the method of identifying rubbing faults mainly starts from finding the original vibration signal that meets the above characteristics. Based on the above characteristics, this method proposes and adopts the impact vibration detection method based on wavelet packet filtering, combined with the bearing temperature abnormality identification based on SVR machine learning method to carry out early warning identification of bearing eccentric wear faults.

[0126] 1. Basic SVR method and wavelet packet filtering method

[0127] 1.1 Support Vector Regression (SVR) Machine Learning Algorithm

[0128] First, the support vector regression (SVR) machine learning algorithm is used. Support vector regression (SVR) is a machine learning method for regression based on support vector machine (SVM). Compared with the commonly used support vector classification machine, SVR predicts the output value corresponding to the newly observed input variable based on the relationship between the input variable and the output variable.

[0129] like Figure 2 The figure shows a schematic diagram of the SVR model. Different from the traditional regression method, SVR allows a certain amount of deviation between the model prediction and the true reference during the fitting process, which is denoted as Based on the fitting ,by As the center, construct a width The deviation interval band is a set of sample data that falls into this interval band. There is no need to calculate the loss, and the prediction is considered correct. However, when the data is outside the interval band, the prediction result is considered to be wrong and the loss needs to be calculated.

[0130] When there is When there are 1 input variable, a nonlinear estimation model is assumed:

[0131]

[0132] Where: represents the deviation, is the weight vector, Transpose it, It is a nonlinear function in high-dimensional feature space; is the input vector.

[0133] SVR optimizes the model by maximizing the interval width and minimizing the total loss, that is, solving In and Certain constraints need to be met, namely:

[0134]

[0135] is the output vector. In practical applications, If the value is too small, it cannot guarantee that all sample points are in the interval band. If the value is too large, it cannot guarantee that all sample points are in the interval band. The regression hyperplane will be biased by some abnormal points. Therefore, SVR allows each sample Adding slack variables In order to avoid the infeasibility of solving the optimization problem, the SVR objective function is expressed as:

[0136]

[0137] In the formula, , respectively represent the interval upper and lower bound slack variables; It is the penalty coefficient, which is used to balance the model complexity and loss. If the value is too large, the interval band will be too narrow. If the value is too small or close to 0, the interval band will be too wide, which will affect the fitting effect. .

[0138] When the solution cannot be found in the sample space, the kernel function is introduced The solution form is as follows:

[0139]

[0140] In the formula and The Lagrange method is used to solve the unknown coefficients. It is a symmetric positive real function. Commonly available kernel functions include linear kernel function, polynomial kernel function, radial basis function kernel, Gaussian kernel function, etc.

[0141] Then, wavelet packet filtering

[0142] 1.2 Wavelet Packet Filtering

[0143] Traditional vibration signal analysis and processing methods generally use Fourier transform. The basic idea of ​​Fourier transform is to decompose the signal into a series of cosine functions, while wavelet transform decomposes the signal into multi-scale wavelet functions (basic wavelet). Fourier analysis is an analysis method with a fixed window function, which cannot reflect the non-stationary, short-duration, time domain and frequency domain localization characteristics of the signal. Wavelet analysis is a time-frequency localization analysis method with a fixed window area but a changeable shape, that is, both the time and frequency windows can be changed.

[0144] Wavelet Packet Decomposition (WPD) is a further development of discrete wavelet transform. In wavelet transform, the signal is decomposed into a series of high-frequency and low-frequency components, but only the low-frequency part is iteratively decomposed. In contrast, WPD iteratively decomposes both high-frequency and low-frequency components at each level of decomposition. This means that it can analyze the frequency content of the signal in more detail.

[0145] Compared with traditional wavelet decomposition, the key advantage of wavelet packet decomposition is its full-band decomposition capability for signals. Traditional DWT ignores decomposition in high-frequency bands, which may lose important signal details in some applications. This full-band analysis capability of WPD is particularly suitable for applications where signal characteristics are not limited to the low-frequency range. Due to its fine-grained analysis characteristics, wavelet packet decomposition shows its unique advantages in various applications.

[0146] Wavelet packet filtering is a denoising filtering method implemented using wavelet packet transform, which is generally completed in three steps: wavelet packet transform decomposition, nonlinear processing of wavelet coefficients to filter out noise, and inverse wavelet packet transform (signal reconstruction).

[0147] (1) Wavelet transform: decompose the signal into wavelet coefficients of different scales and frequency bands ( ). Wavelet packet transform provides local information of the signal in the time domain and frequency domain, so it has better time-frequency resolution.

[0148] (2) Threshold processing: wavelet coefficients Thresholding is performed to suppress noise by setting smaller coefficients to zero. The threshold is usually calculated based on the energy or statistical properties of the coefficients.

[0149] (3) Inverse wavelet packet transform: The processed wavelet coefficients are transformed back to the original signal. The inverse transform produces a denoised signal.

[0150] The advantages of wavelet / wavelet packet filtering include strong adaptability to non-stationary signals and better ability to handle transient signals and local features. While Fourier filtering focuses more on global frequency domain characteristics, wavelet / wavelet packet filtering focuses more on local time-frequency characteristics, and is particularly suitable for nonlinear signals with large transient characteristic changes, while impulse signals are typical time-varying nonlinear signals. In addition, compared with wavelet analysis, wavelet packet analysis has the ability to decompose the signal in the full frequency band and can retain more high-frequency signal details. Therefore, this method uses wavelet packet filtering to achieve bandpass filtering and noise filtering.

[0151] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0152] Embodiment 1

[0153] A method for early warning of eccentric wear of guide bearings of hydropower generator sets is based on vibration, swing, temperature and other signals collected by an online monitoring system or an offline measurement system installed on the hydropower generator set, and realizes eccentric wear detection of bearings by extracting impact signal features and detecting temperature anomalies. These signals specifically include:

[0154] Table 1 Guide bearing eccentric wear warning signal table

[0155]

[0156] In addition to the above signals, in order to identify the temperature changes in the thermal stabilization process under cold startup conditions, it is also necessary to introduce "time after grid connection" as an auxiliary characteristic signal.

[0157] In this embodiment, a method for early warning of eccentric wear of guide bearing bushes of a hydropower unit is provided, which specifically comprises the following steps:

[0158] Step S1: Train based on historical health data samples to obtain the prediction model of the bearing temperature change of each guide bearing and the maximum allowable bearing temperature deviation; the actual flow chart is as follows Figure 3 As shown:

[0159] The specific sub-steps are as follows:

[0160] Furthermore, the step S1 comprises:

[0161] Step S11: extracting characteristic index data in a steady state after grid connection from the stored historical data records; these characteristic index data include:

[0162] Table 2 Prediction characteristic index of guide bearing bush temperature

[0163]

[0164] That is, the time series data of the original input feature index data is set to ( , ),in Represents the Input feature indicators, is the total number of characteristic indicators involved in the prediction, that is, the total number of input variables. Take the above guide bearing temperature prediction as an example, Represents the time series of the peak-to-peak value of the upper X-direction swing, Represents the time series of the peak-to-peak value of the upward Y-direction swing, Represents the time series of the peak-to-peak value of the upper frame's X-axis vibration; Represents the time series of a certain characteristic parameter data, corresponding to the true value of the characteristic indicator at a certain moment, represents the length of the time series of the characteristic parameter obtained, that is, the total number of real measurement data. The corresponding time series of the predicted characteristic index is , that is, the temperature time series of a certain bearing of a guide bearing.

[0165] Step S12: Preprocess the characteristic index data extracted in step S11, remove abnormal points, and obtain the time series of new characteristic index data and the time series of the predicted features ;in, , , is the total number of feature indicators involved in the prediction, is the length of the time series of each characteristic indicator; among them, the preprocessing includes:

[0166] The time series of each characteristic indicator was preprocessed using Gaussian filtering with a 99.7% confidence interval.

[0167] Step S13: Normalization processing: map all time series of feature indicators to the range of [-1, 1] to form a new time series ; That is, for each input feature index ,from Find the eigenvalue sequence The maximum value of and minimum value , then define:

[0168]

[0169] in:

[0170] For the The maximum value of the characteristic index time series, For the The minimum value of the time series of the characteristic indicators.

[0171] Step S14: For the new time series and the time series of the predicted features Randomly shuffle the order, and then extract the first 70% of the data as the training set time series and , the remaining 30% of the data is used as the test set time series and .

[0172] Step S15: Call the training set time series and ,Bundle As independent variable, As the dependent variable, the SVR prediction model is input for training, and the prediction model vector data is obtained after training. ; The parameters of the SVR prediction model are set as follows: penalty coefficient C=1, and the kernel function is selected as the Gaussian kernel function.

[0173] Step S16: The prediction model vector data obtained by training in step S15 Based on the test set time series and , bring it into the SVR prediction model, calculate the prediction error and calculate the penalty coefficient , also called the coefficient of determination:

[0174]

[0175] In the above formula, is the actual measured value in the training set, is the average value of the actual measured values ​​in the training set, represents the predicted value, is the length of the test set time series, The closer the value is to 1, the better the fitting effect is and the smaller the model error is. If it is less than 0.8, it is necessary to readjust the training set, adjust the SVR training parameters (including penalty coefficient and kernel function selection, etc.), and re-execute step S11, otherwise proceed to the next step;

[0176] like Figure 4 The figure shows the comparison of the temperature data predicted by the prediction model obtained after training with this method and the actual measured temperature. The predicted temperature tracks the actual measured temperature well, as shown in Figure 2. Figure 5 The error distribution between the predicted temperature and the actual temperature is shown, with the maximum being approximately 0.2%.

[0177] Step S17: The prediction model vector data obtained by training in step S15 Based on and Bring in the SVR prediction model and calculate the error point by point:

[0178]

[0179] In the above formula, is the actual measured value in the time series of the entire sample set, represents the predicted value, is the error value;

[0180] Then according to 1.25 times of the distribution is taken as the maximum allowable temperature deviation, that is:

[0181]

[0182] In the above formula, Indicates that from m Find it in the data Width range, record This is the allowable deviation of the bearing temperature.

[0183] Step S18: Repeat steps S11 to S17, traverse all upper guide bearing pad temperatures, lower guide bearing pad temperatures, and water guide bearing pad temperatures, obtain the SVR prediction model vector data and maximum allowable deviation of each pad temperature, and record them. is the vector data of a certain watt temperature SVR prediction model, is the maximum allowable deviation of the watt temperature, , It is the sum of the upper, lower and water-conducting tile temperatures.

[0184] Step S2: Based on wavelet packet bandpass filtering, the impact envelope demodulation algorithm is used to extract the impact characteristic index of the guide bearing swing signal and vibration signal collected in real time; the actual process is as follows Figure 6 As shown, the specific sub-steps are as follows:

[0185] Step S21: Obtain the collected real-time swing and vibration original time domain waveform signals , The key phase mark is recorded in The obtained unit guide bearing swing and vibration time domain continuous waveform signal has key phase marks recorded in the signal. . Taking the above guide as an example, is the time domain signal of the upward X-direction swing, is the time domain signal of the upward Y-direction swing, is the X-axis vibration time domain signal of the upper frame, It is the X-axis vibration time domain signal of the upper frame, and the lower guide and water guide are similar. The sampling frequency of the time domain waveform is , in the usual vibration and swing signal collection, Select 1024Hz; Set for The number of sampling points is not less than 4096 points, and The sampled waveform contains a signal with no less than 8 complete rotation cycles, and the number of cycles is recorded as .

[0186] Step S22: Select Morlet as the wavelet basis function, set the decomposition level to 4, Perform wavelet packet transform to extract wavelet packet coefficients of each scale , ;

[0187] According to the principle of wavelet packet decomposition, after 4 layers of decomposition, It is decomposed into a signal composed of 16 different wavelet packets, and the intensity of each wavelet packet is the wavelet coefficient. , and each wavelet packet corresponds to a bandpass filter. =1024Hz, then the frequency ranges corresponding to these 16 wavelet packets are:

[0188] Table 3 Corresponding frequency bands of each wavelet packet after 4-layer wavelet packet decomposition

[0189]

[0190] Step S23: bandpass filtering based on wavelet packets to filter out low-frequency and high-frequency noise;

[0191] In order to remove the influence of the rotation frequency and its main multiple frequencies on the impact signal identification, the signals below 20X (20 times the rotation frequency) are filtered out. At the same time, considering that the measurement noise is concentrated in the high frequency band, it is necessary to filter out the signals above 300Hz, and only retain the signals within the range of 20X to 300Hz. In other words, the impact signal caused by the eccentric wear of the bearing is mainly detected and identified through the signals of vibration and swing within this frequency range;

[0192] This method is based on the wavelet packet decomposition result of step S22, and is designed according to the bandpass filter. Set different filter coefficients to achieve bandpass filtering, as follows:

[0193] set up is the wavelet packet coefficient after filtering, then Calculated by the following formula:

[0194]

[0195] In the above formula, for The corresponding center frequency of the passband frequency; assuming that the rated speed of a unit is 75r / min, then the rated speed frequency of the unit is 1.25Hz, and the corresponding frequency of 20X is 25Hz. According to the above formula, , , , , , , , All are 0, others ;

[0196] Then, in order to reduce the noise impact of signals in the range of 20X to 300Hz, the following method is used to perform noise filtering:

[0197]

[0198] is the adaptive threshold after wavelet transformation, is an estimate of the noise level, calculated as:

[0199]

[0200] in, is the number of data points in the data window used in wavelet transform, The function represents the median value. is the soft threshold function:

[0201]

[0202] in, is the threshold parameter, is the coefficient after transformation, The input value is , The input value is ; The function is a symbolic function and its values ​​are as follows:

[0203]

[0204] Indicates finding the maximum value.

[0205] Step S24: Based on the wavelet packet coefficients after filtering Perform inverse wavelet packet transform to obtain the reconstructed signal and form new vibration and swing signals ; Since the signal filters out low-frequency signals and high-frequency interference signals, as well as noise filtering, the instantaneous impact signal caused by local weak friction will be highlighted. Figure 7 is the original top cover vibration signal of a unit; Figure 8 This is the signal of the above vibration signal processed by wavelet packet filtering, and a clear impact signal can be observed.

[0206] Step S25: , the envelope waveform signal of the impulse signal is obtained by using the Hilbert transform digital envelope demodulation technology ;

[0207] The step S25 comprises:

[0208] set up is the analytical signal of the impact signal envelope signal, then:

[0209]

[0210] in, is an imaginary unit, Represents the Hilbert transform:

[0211]

[0212] According to the impulse signal envelope waveform, use fast Fourier transform to perform envelope spectrum recognition and extract envelope signal ,in, is the complex conjugate of the analytic signal:

[0213]

[0214] like Fig. 9 is the envelope signal of the impulse signal.

[0215] Step S26: impact feature extraction;

[0216] Specifically, the step S26 includes:

[0217] In obtaining Afterwards, according to the key phase signal in the waveform, the data of each cycle is intercepted, and the envelope peak value and azimuth of each cycle are calculated according to the data of each cycle;

[0218] like Fig.10 A periodic envelope signal, from Intercept two adjacent bond phases (such as Fig.10 The waveform signal between the key phase signal 1 and the key phase signal 2) in the first Cycle Signal, ;

[0219] right Sort the data, remove the largest 50% of the data, and find the maximum value from the remaining 50% of the data as the noise baseline value in the signal ;

[0220] Constructing new time domain waveform ,right The signal is pulse peak detected to calculate the pulse amplitude caused by the impact, and the phase corresponding to the pulse peak is synchronously calculated according to the key phase signal;

[0221] Repeated calculation Record the number of shock pulses, pulse amplitude and pulse orientation in all cycles; is the total number of pulses, For the amplitude of each pulse, is the pulse direction, find:

[0222]

[0223] For envelope waveform signal Perform full-cycle FFT transformation to form an envelope spectrum (as shown below Fig.11), and extract the main frequency, and record the maximum amplitude main frequency as .

[0224] Step S3: Based on wavelet packet bandpass filtering, through the impact envelope demodulation algorithm, select multiple historical health data to accumulate the envelope average peak value in the historical sample data , which respectively correspond to the upper guide X-direction swing signal, the upper guide Y-direction swing signal, the upper frame X-direction vibration signal, and the upper frame X-direction vibration signal; the method is the same as step S2, and will not be repeated here.

[0225] Step S4: Perform early warning detection of eccentric wear of the guide bearing bushing of the unit.

[0226] like Fig.12 As shown, the specific steps are as follows, the step S4 comprises:

[0227] Step S41: Based on the extracted envelope peak and orientation of each cycle, calculate the average pulse amplitude and basic pulse amplitude growth coefficient of each guide bearing swing and vibration ;

[0228] Step S42: All the impact signals of the guide bearing swing and vibration are judged as follows:

[0229] Condition 1: Judgment of swing impact index: , and the main frequency for The frequency multiplication, where is an integer, and , is the number of pads of the corresponding guide bearing;

[0230] Condition 2: Vibration and shock index judgment:

[0231] For the vibration and shock characteristic indicators of displacement output: , and the main frequency for Frequency doubling;

[0232] Vibration shock characteristic indicators for speed output: , and the main frequency for Frequency doubling;

[0233] Vibration and shock characteristic indicators for acceleration output: , and the main frequency for Frequency doubling;

[0234] If all the above conditions are not met, there is no eccentric wear of the guide bearing bush, and the process returns. Otherwise, proceed to the next step.

[0235] Step S43: Detection of abnormal bearing bush temperature:

[0236] Read the phase of the corresponding impact pulse from the vibration or swing signal that meets the above conditions ,according to Calculate the bearing number of the corresponding position and set the bearing number to ;

[0237] Real-time collection of bearings Temperature, setting For real-time bearing The measured value of temperature; at the same time, other input parameters used in the bearing temperature prediction model are collected (such as the peak-to-peak value of the X / Y direction swing, the peak-to-peak value of the upper frame X / Y direction vibration, the temperature of two adjacent bearings, the temperature of the opposite bearing, active power, reactive power, ambient temperature, time after grid connection, etc.); the prediction model vector data obtained by training in step S1 Based on the above input parameters, the SVR prediction model is used to calculate the predicted temperature and record it as ,remember The bearing obtained by training in step S1 The allowable deviation of the watt temperature is:

[0238] if , then the bearing Eccentric wear warning, bearing number is ;

[0239] Otherwise the bearing If eccentric wear is not established, step S43 is repeated until the impact pulse features in all directions are detected.

[0240] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

[0241] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A method for early warning of eccentric wear of guide bearing bushes of a hydropower unit, characterized in that: include: Step S1: training based on historical health data samples to obtain a prediction model for the temperature change of each guide bearing bush and a maximum allowable bush temperature deviation; Step S2: Based on wavelet packet bandpass filtering, the impact characteristic index of the guide bearing swing signal and vibration signal collected in real time is extracted through the impact envelope demodulation algorithm; Step S3: Based on wavelet packet bandpass filtering, through the impact envelope demodulation algorithm, select multiple historical health data to accumulate the envelope average peak value in the historical sample data , which correspond to the upper guide X-direction swing signal, the upper guide Y-direction swing signal, the upper frame X-direction vibration signal, and the upper frame X-direction vibration signal respectively; Step S4: Perform early warning detection of eccentric wear of guide bearing bush of the unit; The step S2 comprises: Step S21: Obtain the collected real-time swing and vibration original time domain waveform signals , Key phase marks are recorded in the Step S22: Select Morlet as the wavelet basis function, set the decomposition level to 4, Perform wavelet packet transform to extract wavelet packet coefficients of each scale , ; Step S23: bandpass filtering based on wavelet packets to filter out low-frequency and high-frequency noise; Step S24: Based on the wavelet packet coefficients after filtering Perform inverse wavelet packet transform to obtain the reconstructed signal and form new vibration and swing signals ; Step S25: , the envelope waveform signal of the impulse signal is obtained by using the Hilbert transform digital envelope demodulation technology ; Step S26: impact feature extraction; The step S23 comprises: set up is the wavelet packet coefficient after filtering, then Calculated by the following formula: In the above formula, for The corresponding center frequency of the passband frequency; Then, in order to reduce the noise impact of signals in the range of 20X to 300Hz, the following method is used to perform noise filtering: is the adaptive threshold after wavelet transformation, is an estimate of the noise level, calculated as: in, is the number of data points in the data window used in wavelet transform, The function represents the median value. is the soft threshold function: in, is the threshold parameter, is the coefficient after transformation, The input value is , The input value is ; The function is a symbolic function and its values ​​are as follows: Indicates finding the maximum value; The step S25 comprises: set up is the analytical signal of the impact signal envelope signal, then: in, is an imaginary unit, Represents the Hilbert transform: According to the impulse signal envelope waveform, use fast Fourier transform to perform envelope spectrum recognition and extract envelope signal ,in, is the complex conjugate of the analytic signal: The step S26 comprises: In obtaining Afterwards, according to the key phase signal in the waveform, the data of each cycle is intercepted, and the envelope peak value and azimuth of each cycle are calculated according to the data of each cycle; from The waveform signal between two adjacent key phases is intercepted to generate the first Cycle Signal; right Sort the data, remove the largest 50% of the data, and find the maximum value from the remaining 50% of the data as the noise baseline value in the signal ; Constructing new time domain waveform ,right The signal is pulse peak detected to calculate the pulse amplitude caused by the impact, and the phase corresponding to the pulse peak is synchronously calculated according to the key phase signal; Repeated calculation Record the number of shock pulses, pulse amplitude and pulse orientation in all cycles; is the total number of pulses, For the amplitude of each pulse, is the pulse direction, find: For envelope waveform signal Perform full-cycle FFT transformation to form an envelope spectrum, extract the main frequency, and record the maximum amplitude main frequency as ; The step S4 comprises: Step S41: Based on the extracted envelope peak and orientation of each cycle, calculate the average pulse amplitude and basic pulse amplitude growth coefficient of each guide bearing swing and vibration ; Step S42: All the impact signals of the guide bearing swing and vibration are judged as follows: Condition 1: Judgment of swing impact index: , and the main frequency for The frequency multiplication, where is an integer, and , is the number of pads of the corresponding guide bearing; Condition 2: Vibration and shock index judgment: For the vibration and shock characteristic indicators of displacement output: , and the main frequency for Frequency doubling; Vibration shock characteristic indicators for speed output: , and the main frequency for Frequency doubling; Vibration and shock characteristic indicators for acceleration output: , and the main frequency for Frequency doubling; If all the above conditions are not met, there is no eccentric wear of the guide bearing bush, and the process returns. Otherwise, proceed to the next step. Step S43: Detection of abnormal bearing bush temperature: Read the phase of the corresponding impact pulse from the vibration or swing signal that meets the above conditions ,according to Calculate the bearing number of the corresponding position and set the bearing number to ; Real-time collection of bearings Temperature, setting For real-time bearing The measured value of temperature; at the same time, other input parameters used in the bearing temperature prediction model are collected; the prediction model vector data obtained by training in step S1 Based on the above input parameters, the SVR prediction model is used to calculate the predicted temperature and record it as ,remember The bearing obtained by training in step S1 The allowable deviation of the watt temperature is: if , then the bearing Eccentric wear warning, bearing number is ; Otherwise the bearing If eccentric wear is not established, step S43 is repeated until the impact pulse features in all directions are detected.

2. A method for early warning of eccentric wear of guide bearings of hydropower generator sets according to claim 1, characterized in that: The guide bearing shafts of each part include: an upper guide bearing, a lower guide bearing, and a water guide bearing.

3. A method for early warning of eccentric wear of guide bearings of hydropower generator sets according to claim 2, characterized in that: The step S1 comprises: Step S11: extracting characteristic index data in a steady state after grid connection from the stored historical data records; Step S12: Preprocess the characteristic index data extracted in step S11, remove abnormal points, and obtain the time series of new characteristic index data and the time series of the predicted features ;in, , , is the total number of feature indicators involved in the prediction, is the length of the time series of each characteristic indicator; Step S13: Normalization processing: map all time series of feature indicators to the range of [-1, 1] to form a new time series ; Step S14: For the new time series and the time series of the predicted features Randomly shuffle the order, and then extract the first 70% of the data as the training set time series and , the remaining 30% of the data is used as the test set time series and ; Step S15: Call the training set time series and ,Bundle As independent variable, As the dependent variable, the SVR prediction model is input for training, and the prediction model vector data is obtained after training. ; The parameters of the SVR prediction model are set as follows: the penalty coefficient C=1, and the kernel function is selected as the Gaussian kernel function; Step S16: The prediction model vector data obtained by training in step S15 Based on the test set time series and , bring it into the SVR prediction model, calculate the prediction error and the penalty coefficient; Step S17: The prediction model vector data obtained by training in step S15 Based on and Bring in the SVR prediction model and calculate the error point by point; Step S18: Repeat steps S11 to S17, traverse all upper guide bearing pad temperatures, lower guide bearing pad temperatures, and water guide bearing pad temperatures, obtain the SVR prediction model vector data and maximum allowable deviation of each pad temperature, and record them. is the vector data of a certain watt temperature SVR prediction model, is the maximum allowable deviation of the watt temperature, , It is the sum of the upper, lower and water-conducting tile temperatures.

4. A method for early warning of eccentric wear of guide bearings of hydropower generator sets according to claim 3, characterized in that: The preprocessing in step S12 includes: The time series of each characteristic indicator was preprocessed using Gaussian filtering with a 99.7% confidence interval; The step S13 comprises: in: For the The maximum value of the characteristic index time series, For the The minimum value of the time series of the characteristic indicators.

5. A method for early warning of eccentric wear of guide bearings of hydropower generator sets according to claim 4, characterized in that: The step S16 comprises: In the above formula, is the actual measured value in the training set, is the average value of the actual measured values ​​in the training set, represents the predicted value, is the length of the test set time series, The closer the value is to 1, the better the fitting effect is and the smaller the model error is. If it is less than 0.8, it is necessary to readjust the training set, adjust the SVR training parameters, and re-execute step S11, otherwise, execute the next step.

6. A method for early warning of eccentric wear of guide bearings of hydropower generator sets according to claim 5, characterized in that: The step S17 comprises: In the above formula, is the actual measured value in the time series of the entire sample set, represents the predicted value, is the error value; Then according to 1.25 times of the distribution is taken as the maximum allowable temperature deviation, that is: In the above formula, Indicates that from m Find it in the data Width range, record This is the allowable deviation of the bearing temperature.

Images