CN110737867A

CN110737867A - Transformer based on deep learning and mechanical vibration charged acquisition processing device

Info

Publication number: CN110737867A
Application number: CN201810794354.6A
Authority: CN
Inventors: 范盛荣
Original assignee: Wuhan Sheng Jetta Power Technology Co Ltd
Current assignee: Wuhan Sheng Jetta Power Technology Co Ltd
Priority date: 2018-07-19
Filing date: 2018-07-19
Publication date: 2020-01-31

Abstract

The invention discloses devices for judging the mechanical vibration states of a transformer core, a winding and accessories through deep learning, wherein the vibration of the transformer winding, the core and the accessories is extremely complex, different manufacturers and structures are different, even manufacturers and transformers in different batches have different vibration characteristics, the states of the transformer core, the winding and the accessories are directly related to the safe operation of the transformer, the characteristic that a large-scale transformer cannot be powered off for detection is considered, how to acquire specific parameters and obtain the internal structure state under the electrified state is very necessary, a big data analysis method with the deep learning function is utilized, the characteristic quantity in the vibration of the transformer in the operation process is extracted, the device and the algorithm for detecting the vibration states of the transformer core and the accessories through comparison of early-stage and offline results and general test verification are provided, and the reliability can reach more than 95 percent.

Description

Transformer based on deep learning and mechanical vibration charged acquisition processing device

Technical Field

The invention belongs to the field of detection of states of power equipment and mechanical equipment, and particularly relates to a method for detecting and processing vibration signals generated in the motion of a rotary or reciprocating machine with the states of a winding, an iron core and accessories of a transformer with the voltage level of 110KV and above and the rotating speed of 12000 rpm in an electrified way.

Background

During the operation of the equipment, various complex vibration signals are generated, the vibration of the surrounding environment and accessories is added, and the natural frequency of the hydraulic machine or the pipeline of the propagation path is added, so that the vibration signals collected on the outer surface are various superposed highly mixed signals. Therefore, how to extract the real signals representing the operation of the equipment from the mixed signals is a problem which must be faced by all vibration detection.

At present, the vibration signal decomposition is commonly used by EMD, EEMD, FFT, Hilbert-Huang transform and BBS blind source separation, the processing methods are different in thousands of autumn and different in occasions and application environments, and cannot be treated in .

Disclosure of Invention

The method can be compatible with the processing of most of the vibration signals of transformers and high-speed electromechanical equipment, and the processing methods have detailed requirements on the positions of the acquired signals. For the vibration data acquired by the acquisition part, various interference signals can be accurately eliminated through a comprehensive processing algorithm of system integration, and the vibration data representing the equipment operation information is acquired. By corresponding to the specific equipment, the state of the equipment can be accurately acquired.

1. The concept of instantaneous frequency, the instantaneous energy of the signal and the instantaneous envelope, has been widely accepted by , while the concept of instantaneous frequency is at odds with the two fundamental difficulties of accepting instantaneous frequency , which are primarily influenced by the deep foundation of the fourier transform, in the conventional fourier transform, the frequency defines a sine or cosine function with constant amplitude over the entire data length, as an extension of this definition, the concept of instantaneous frequency must also be related to a sine or cosine function, therefore, at least full cycles of sine and cosine fluctuations are required to define local frequency values.

For any time series x (t), the Hilbert transform y (t) is defined as:

（1）

where P represents the Cauchy principal value, and the transformation holds true for all LPs, according to this definition of , when X (t) and Y (t) form complex conjugates, analytic signals Z (t) are obtained:

（2）

wherein

Thus, the Hilbert transform provides unique amplitude and phase defining functions, equation (2) defines the Hilbert transform as a convolution of X (t) with 1/t, and thus emphasizes the local nature of X (t). in equation (1), the polar expression further shows its local nature, which is the best local approximation of amplitude and phase varying trigonometric functions x (t). even with the Hilbert transform, there is considerable controversy in defining instantaneous frequency using the following equation.

For example, the real part of the Fourier transform must have only positive frequencies, this constraint can be mathematically proven, but is still a global definition.

x(t)=sin(t) （3）

Its Hilbert transform is cos (t) its phase diagram as shown in fig. 1 is unit circles on the x-y plane, the phase as a function of time is straight line diagrams 2, and the instantaneous frequency diagram 3 is constants if the mean of x (t) is changed, for example, plus constants a:

x(t)=a+sin(t)

its x-y plane phase diagram is still simple circles, but the center is shifted a long distance, if a <1, the center is still in the circle, under which condition x (t) has violated constraints because its fourier transform has dc terms, but its average zero-crossing frequency is still the same as that of two 0, but its phase function and instantaneous frequency will differ from that of two 0, if a >1, the center is not in the circle, and x (t) no longer satisfies the constraints, it can be seen that both the phase function and instantaneous frequency have meaningless negative values.

This example illustrates that for simple signals, such as sinusoidal signals, the instantaneous frequency is meaningful only if local symmetry to zero mean is satisfied, for generic signal data, any superimposed signal waveform will be equivalent to a >1 case, any asymmetric signal waveform will be equivalent to a <1 case.

2. Since most signals or data are not natural mode functions, at any instant the data may contain multiple oscillation modes, which explains why a simple Hilbert transform does not give a complete description of the frequency content of common signals.

(1) The number of extreme points and the number of zero-crossing points must be equal or at most different by not more than in the whole data sequence;

(2) at any time point, the local maxima and minima of the signal define an envelope average value of zero.

The th constraint is very obvious and it approximates the definition of the traditional smooth Gaussian process for narrow bands the second condition is new ideas that change the traditional global constraint to a local one.

The provision of the natural modal functions allows the instantaneous frequencies defined by the Hilbert transform to have practical physical significance, while the advent of empirical modal decomposition methods that provide the natural modal function components allows the instantaneous frequencies to be used for analysis of complex, non-stationary signals.

3. Empirical Mode Decomposition (EMD) brand-new non-stationary signal processing methods, which starts from the separation of inherent mode functions from complex signals, considering that any complex signal is composed of FM-AM components, for non-stationary signals, if X is used_k(t) represents both amplitude and frequency variations, and x (t) can be expressed in the form:

（4）

for any signal s (t), the empirical mode decomposition algorithm can be described as follows, and assuming that the signal variable s (t) has an initial value of m0(t), let i = 1 perform the following operation on it:

(1) obtaining all maximum and minimum points of s (t), and interpolating between the maximum points by cubic spline function to obtain upper and lower envelope lines e_max(t) and e_min(t)；

(2) Calculating the average value m (t) = (e) of the upper and lower envelopes_max(t)+e_min(t))/2；

(3) Then the extraction details are denoted as d (t), i.e. d (t) = s (t) -m (t);

(4) judging whether d (t) meets two conditions of the industrial MF, if so, making d_i(t) = d (t), considering the IMF components as IMF components, transferring to the step (5), if the IMF components do not meet the requirement, making s (t) = d (t), and repeating the step(1)-(4)。

(5) If mi (t) = s (t) -di (t), determine if mi (t) is monotonic functions or if | mi (t) | is very small, and consider as the measurement error, if yes, the loop ends, otherwise let s (t) = mi (t) go to step (1).

When the program cycle is finished, n work scores satisfying the conditions are obtained. C_imfi(t)=d_i(t) let r_n=m_n(t), referred to as residual components. This results in empirical mode decomposition:

（5）

i.e. the original signal may be represented as the sum of a number of natural mode function components and residual terms.

4. Hilbert spectrum and marginal spectrum: performing a hilbert transform on the natural mode function components:

（6）

thus constructing an analytic signal:

（7）

wherein

Step obtains the instantaneous frequency:

the hilbert spectrum of the signal thus obtained is:

（8）

after integrating the above equation t, the marginal spectrum of the signal can be obtained as follows:

reflecting the amplitude of the signal as a function of frequency over the entire frequency band. The Hilbert marginal spectrum obtained by the Hilbert-Huang transform gets rid of the limitation of the traditional Fourier transform, does not generate energy leakage, can accurately describe the amplitude of each frequency component, and has high resolution.

5. Adaptive threshold denoising of the natural mode function: according to the empirical mode decomposition principle, the obtained natural mode functions are filtered layer by layer from high frequency to low frequency, noise signals are mainly concentrated in the first few layers of natural mode functions, and therefore the self-adaptive selection method of the threshold processing layer number of the natural mode functions is provided, and noise interference can be well inhibited after the processing by the method.

Random noise interference is , which is the most common interference in the field, and because the characteristic information generated at the early stage of the mechanical failure of the transformer is weak, the random noise interference is often submerged in the random noise and is difficult to distinguish, and the random noise interference is based on gross error theory

The criteria solve the problem well. All inherent modal function components obtained after the random noise is subjected to empirical mode decomposition still meet the random distribution rule, but the vibration signal does not have the rule. Therefore, the original signal can be regarded as coarse difference, and the signal-noise separation can be realized by combining a closed-value processing method. For random errors that follow a normal distribution.

The mathematical statistics show that:

wherein the content of the first and second substances,

for the mathematical expectation of the random variable,

is the standard deviation. Obviously, if the threshold is chosen

Then 99.74% of the noise can be suppressed. Most of the value falls on the random noise

The following threshold values may therefore be selected:

as can be known from the analysis in the previous section, the IMF obtained through EMD shows layer-by-layer filtering from high frequency to low frequency, and noise signals are mainly concentrated in the first few layers of IMFs, so that an IMF adaptive threshold denoising method is provided. The algorithm comprises the following steps:

(1) setting the number of layers for threshold processing as m according to the IMF obtained by EMD decomposition, wherein the value of m is the total number of layers from 1 to IMF;

(2) performing closed-value processing on IMF (1) to IMF (m), the threshold value being based on gross error detection

Criteria determination

，Is the standard deviation; c = 3.0-4.0, and simultaneously, in order to keep the vibration signal characteristics, a hard threshold processing method is selected;

(3) reconstructing all IMFs to obtain a denoised vibration signal, and calculating the mean square error of the denoised vibration signal and the original vibration signal;

(4) and (3) checking whether the end condition is met, if the mean square deviation value obtained currently meets the preset requirement, ending the circulation, taking m corresponding to the mean square deviation value at the moment as the optimal IMF threshold processing layer number, otherwise, returning to the step (1), and continuing the circulation.

6. Core vibration signal analysis, namely, testing the body vibration of a three-phase oil-immersed transformer with power stations, the rated capacity of 500kVA, &tttttranslation = one "&ttt/t &tttsecondary winding (high-voltage side) voltage of l0kV, the secondary winding (low-voltage side) voltage of 0.4kV and the winding connection group of Y/D11 under no-load conditions, wherein the vibration signals of the transformer body are transmitted through more links, so that the measured vibration signals contain a large amount of noise, and fig. 4 and 5 are respectively the time domain waveforms of the vibration acceleration signals of the transformer body measured under normal conditions and loose iron core fastening bolts (taking the high-voltage side A phase as an example), so that the amplitude of the vibration acceleration signals of the transformer body measured under the loose iron core bolts is changed from the normal conditions, but the vibration acceleration signals cannot form effective criteria for diagnosis, therefore, the vibration signals need to be analyzed so as to extract effective characteristic information.

According to the Hilbert yellow transformation principle, firstly, EMD decomposition is carried out on a vibration signal to obtain time domain waveforms of I-M IF 1-M IF 7, as shown in FIG. 6, then threshold processing is carried out on the I-M IF, including self-adaptive closed value selection and self-adaptive inherent mode function threshold processing layer number selection, the I-M IF after threshold processing is shown in FIG. 7, a threshold parameter c =3.27 is obtained, so that an I-M component after interference removal is obtained, and finally Hilbert transformation is carried out on each IMF after noise removal, so that a transformer vibration signal Hilbert energy spectrum (subjected to returning processing) and a marginal spectrum are obtained, as shown in FIG. 8.

As can be seen from FIGS. 9 and 10, the energy of the vibration signal after the core bolt is loosened has a significant change, the energy of the 300-400Hz frequency band has a significant increase, especially at 300Hz, the energy of the frequency bands near 104Hz and 700-900Hz has a small increase, the energy near 200Hz and 600Hz has a small decrease, and the energy change near 500Hz is not significant. The more the core bolt loosens, the greater the energy variation. The marginal spectrum of the vibration signals in fig. 9 and 10 can clearly observe the position and the size of the spectral peak of the core fault, the spectral peak appears at the vibration fundamental frequency of 100Hz and the frequency multiplication thereof, and the energy of the core vibration is mainly concentrated near 200-400 Hz. By analyzing the energy distribution characteristics of the Hilbert spectrum and the marginal spectrum, effective and simple state fingerprints can be extracted to judge the condition of the transformer core.

7. Analyzing a winding vibration signal: the transformer body vibration under the load condition was tested. During the experiment, the low voltage side was short circuited and a voltage was applied to the high voltage side. Because the low-voltage side is directly short-circuited without any impedance, the impedance of the equivalent circuit of the whole transformer is very small, and the applied voltage is very low. The applied voltage when the short circuit current reaches the rated value is referred to as the impedance voltage which is only 5% -10% of the rated voltage. The magnetic flux in the core is small and the vibration of the core due to magnetostriction is negligible, and the vibration signal measured from the tank surface is essentially the vibration signal of the transformer winding.

The method comprises the steps of changing the thickness of a gasket between windings, and measuring vibration acceleration signals of the windings in different loosening and tightening states, wherein the vibration acceleration signals of the windings in different loosening and tightening states are respectively shown in FIGS. 11, 12 and 13, which are vibration signal time domain graphs of a high-voltage side phase A winding measured under the conditions of normal and loosening faults of a transformer winding (a fault graph 12 is more serious than a fault graph 11), and the vibration signals of the windings in FIGS. 11, 12 and 13 are different from the vibration signals of the windings in normal operation under the fault condition, but effective fault characteristic information cannot be obtained, so that whether the windings are in fault and the fault degree are judged, therefore, the vibration signals need to be further decomposed to extract characteristic information capable of representing the faults of the windings.

According to the Hilbert yellow transformation principle, firstly, EMD decomposition is carried out on the vibration signals shown in the figures 12 and 13 to obtain 6 IMF components, as shown in figure 14 (taking a fault 1 vibration signal as an example), then, adaptive threshold denoising is carried out on the obtained IMF components, the IMF components after denoising processing are shown in figure 15, and as can be seen, the method adaptively selects the first 5 layers of IMFs for threshold processing, and the threshold parameters

=3.427, which provides more accurate IMF component for subsequent hilbert transform, so as to extract characteristic information of winding fault more accurately; finally, after denoisingThe IMFs of (a) are subjected to hilbert transform to obtain hilbert energy spectra that reflect intrinsic characteristics of the signals, and the signals are subjected to quantization , which is shown in fig. 16.

The method comprises the steps of measuring the vibration signal of a transformer at a high voltage side, and obtaining vibration signal time domain diagrams of the transformer under the condition that a fault 1 is measured under the conditions of 60% rated current and 80% rated current IN, carrying out Hilbert yellow transformation and IMF adaptive threshold value denoising on the vibration signals shown IN the figures 17 and 18 to obtain energy diagrams of each frequency band of the vibration signals shown IN the figure 19, comparing the figures 19 and 16 to obtain the result that the energy of the vibration signals of the winding is mainly concentrated at 100Hz, when the winding has a fault, the energy at 100Hz is obviously reduced, the energy at 200Hz is slightly increased, the energy at 300 Hz-500 Hz is not greatly changed, the fault is more serious, the energy change is larger, and meanwhile, the energy of the vibration signals of the winding is increased along with the increase of the load current.

Drawings

Fig. 1 is a phase diagram showing unit circle instantaneous frequencies in the x-y plane.

Fig. 2 is a graph of phase as a function of time for straight instantaneous frequencies.

Fig. 3 is an instantaneous frequency of constants.

Fig. 4 is a time domain diagram of a transformer vibration signal in a normal case.

Fig. 5 is a time domain diagram of a core fastening bolt loosening transformer vibration signal.

Fig. 6 shows respective MF components of the core fault vibration signal.

FIG. 7 shows the respective artificial MF components after adaptive threshold denoising.

Fig. 8 is a transformer vibration signal hilbert energy spectrum.

Fig. 9 is a transformer vibration signal margin spectrum for normal conditions.

Fig. 10 is a core fastening bolt loosening transformer vibration signal margin spectrum.

FIG. 11 is a time domain diagram of a winding vibration signal under normal conditions.

FIG. 12 is a time domain diagram of an abnormal condition winding vibration signal.

FIG. 13 is a time domain plot of a winding vibration signal in a more severe case than FIG. 12.

Fig. 14 is a diagram showing the IMF components of the fault winding vibration signal.

FIG. 15 is a graph of IMF components of a fault winding vibration signal after thresholding.

Fig. 16 is a graph of energy in each frequency band of a vibration signal.

FIG. 17 is a time domain plot of the vibration signal at 60% load current.

FIG. 18 is a time domain plot of the vibration signal at 80% load current.

FIG. 19 is a graph of energy in each frequency band of a vibration signal at different load currents.

Fig. 20 is a schematic diagram of the system architecture.

Detailed Description

(1) The position of the sensor placement needs to meet the following requirements: transformer height 3/4, 1/2.

(2) For the arc position, the sensor cannot be placed.

(3) The lowest position of the bottom is required to be more than 30 cm.

(4) For high-speed machines, it is necessary to determine the position of the bearing and the position of the gear directly above, right and front.

Claims

1. Transformer based on degree of depth study, the electrified processing apparatus that gathers of mechanical vibration, its characterized in that: the sensor is arranged at a designated position of a transformer and a high-speed machine through a magnetic seat.

2. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the sensor is connected with the acquisition system through a shielding armored cable.

3. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the sensor is connected with the acquisition system through DB25 through a shielded armored cable.

4. The deep learning-based transformer and high-speed mechanical vibration charged acquisition and processing device as claimed in claim 1, wherein: the sensor is connected with the acquisition system through a radio frequency terminal through a shielding armored cable.

5. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the sensor passes through the shielding armoured cable and is connected through BNC with collection system.

6. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the processing system is an industrial-grade computer.

7. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the processing system supports filtering of ambient noise, ground vibration, cooling system signals.

8. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the system provides a time domain graph and a frequency domain graph of the monitored object after processing according to the acquired vibration signal.

9. The deep learning-based transformer and mechanical vibration charged acquisition and processing device according to claim 1, wherein: the system provides a diagnostic result of the object by processing the acquired vibration signal.