CN115166514A - Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising - Google Patents
Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising Download PDFInfo
- Publication number
- CN115166514A CN115166514A CN202210722448.9A CN202210722448A CN115166514A CN 115166514 A CN115166514 A CN 115166514A CN 202210722448 A CN202210722448 A CN 202210722448A CN 115166514 A CN115166514 A CN 115166514A
- Authority
- CN
- China
- Prior art keywords
- motor
- spectrum
- signal
- function
- fault
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001228 spectrum Methods 0.000 title claims abstract description 78
- 238000000034 method Methods 0.000 title claims abstract description 57
- 230000011218 segmentation Effects 0.000 title claims abstract description 22
- 238000013507 mapping Methods 0.000 claims abstract description 31
- 230000008569 process Effects 0.000 claims abstract description 24
- 230000009467 reduction Effects 0.000 claims abstract description 22
- 230000002087 whitening effect Effects 0.000 claims abstract description 15
- 238000003745 diagnosis Methods 0.000 claims abstract description 12
- 238000005070 sampling Methods 0.000 claims abstract description 9
- 238000005192 partition Methods 0.000 claims abstract description 3
- 230000006870 function Effects 0.000 claims description 79
- 230000003044 adaptive effect Effects 0.000 claims description 21
- 230000004913 activation Effects 0.000 claims description 13
- 210000002569 neuron Anatomy 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 7
- 230000001133 acceleration Effects 0.000 claims description 5
- 230000007704 transition Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000004590 computer program Methods 0.000 claims description 4
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000003595 spectral effect Effects 0.000 claims description 3
- 238000012549 training Methods 0.000 description 25
- 238000013135 deep learning Methods 0.000 description 8
- 230000006835 compression Effects 0.000 description 6
- 238000007906 compression Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 239000006185 dispersion Substances 0.000 description 4
- 238000007781 pre-processing Methods 0.000 description 4
- 238000000513 principal component analysis Methods 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 230000006698 induction Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000005299 abrasion Methods 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 230000008602 contraction Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000014509 gene expression Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000013021 overheating Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 150000003839 salts Chemical class 0.000 description 1
- 230000005236 sound signal Effects 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/34—Testing dynamo-electric machines
- G01R31/343—Testing dynamo-electric machines in operation
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R15/00—Details of measuring arrangements of the types provided for in groups G01R17/00 - G01R29/00, G01R33/00 - G01R33/26 or G01R35/00
- G01R15/14—Adaptations providing voltage or current isolation, e.g. for high-voltage or high-current networks
- G01R15/18—Adaptations providing voltage or current isolation, e.g. for high-voltage or high-current networks using inductive devices, e.g. transformers
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R19/00—Arrangements for measuring currents or voltages or for indicating presence or sign thereof
- G01R19/0092—Arrangements for measuring currents or voltages or for indicating presence or sign thereof measuring current only
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R19/00—Arrangements for measuring currents or voltages or for indicating presence or sign thereof
- G01R19/25—Arrangements for measuring currents or voltages or for indicating presence or sign thereof using digital measurement techniques
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/34—Testing dynamo-electric machines
- G01R31/346—Testing of armature or field windings
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Power Engineering (AREA)
- Control Of Electric Motors In General (AREA)
Abstract
The invention discloses a motor fault identification method and system based on self-adaptive spectrum segmentation and denoising. The method comprises the following steps: collecting vibration signals and stator current signals of the motor in a normal state and a fault state under the condition that the motor is in no-load stable operation; carrying out Fourier transform on an original motor signal to obtain a signal frequency spectrum, determining a self-adaptive partition coefficient and a boundary of spectrum division according to the frequency spectrum and sampling information, and dividing the frequency spectrum into different parts; establishing a wavelet base, and decomposing the spectrum signals of each interval by using empirical wavelets; calculating a base line passing rate and a correlation coefficient, removing low-frequency signals and high frequencies with small correlation, and denoising by adopting a semi-soft threshold function; reconstructing the denoised signal, whitening the reconstructed signal, sending the reconstructed signal into a sparse self-encoder for dimensionality reduction, establishing a mapping relation between the dimensionality reduced characteristics and the motor faults, and identifying the faults in the real-time working process of the motor based on the mapping relation. The invention improves the accuracy and efficiency of fault diagnosis.
Description
Technical Field
The invention relates to the technical field of motor fault diagnosis, in particular to a motor fault identification method and system based on self-adaptive spectrum segmentation and denoising.
Background
The motor is an important electromechanical device, can convert electric energy and mechanical energy into each other, and has a very important position in various industrial fields such as electric transmission, transportation, servo control and the like. In many motor-related industrial application scenarios, the operating conditions of the motor or the power transmission system are often harsh, and various types of faults of the motor or the power transmission system may be caused by vibration, moisture, mold, salt fog, aging, abrasion, overheating of the equipment, and other factors in the industrial environment. The motor is a complex system, so when a certain specific fault occurs, the current signal, the vibration signal, the sound signal and the temperature signal of the stator of the motor also change, how to select one or more signals, the characteristic signals capable of representing the fault type are extracted by a signal processing method, the internal rules of the signals changing can be found, and the diagnosis of the early fault of the motor by using the characteristics is an important direction in the field of motor fault monitoring. The traditional fault diagnosis methods at present have some defects, and the requirements of the fault monitoring of the motor at present cannot be met.
Disclosure of Invention
The invention aims to: aiming at the problems in the prior art, the invention provides a motor fault identification method and system based on self-adaptive frequency spectrum segmentation and denoising, and the fault identification precision and efficiency are improved.
The technical scheme is as follows: a motor fault identification method based on self-adaptive spectrum segmentation and denoising comprises the following steps:
(1) Acquiring vibration signals and stator current signals of the motor in a normal state and a fault state under the condition of no-load stable operation of the motor;
(2) Fourier transform is carried out on the original motor signal to obtain a signal frequency spectrum X (f), and an adaptive division coefficient f is determined according to the frequency spectrum and sampling information g The spectrum is divided into shares such that each share contains f g Dividing points, determining a boundary of spectrum division according to an extreme value of each spectrum, and establishing a corresponding filter bank;
(3) Defining a scale function and an empirical wavelet function, and decomposing the spectrum signals of each interval by using the empirical wavelet;
(4) For the decomposed signals, calculating a baseline passing rate and a correlation coefficient based on a given baseline, removing low-frequency signals and high frequencies with insufficient correlation, denoising by adopting a semi-soft threshold function, and reconstructing the denoised signals;
(5) After whitening pretreatment is carried out on the reconstructed signal, the reconstructed signal is sent into a sparse self-encoder for dimension reduction, and a mapping relation between the features after dimension reduction and motor faults is established;
(6) And identifying the fault in the real-time working process of the motor based on the mapping relation.
A motor fault identification system based on adaptive spectrum division denoising comprises:
the signal acquisition system comprises an acceleration sensor and a pincerlike current transformer and is used for respectively acquiring a vibration signal and a stator current signal of the motor in a normal state and a fault state; and
signal processing apparatus comprising a processor, a memory and a computer program, wherein the computer program is stored in the memory and configured to be executed by the processor, the program when executed by the processor implementing the steps of:
fourier transform is carried out on an original motor signal to obtain a signal frequency spectrum X (f), and an adaptive division coefficient f is determined according to the frequency spectrum and sampling information g Dividing the frequency spectrum intoSeveral portions such that each portion contains f g Dividing points, determining a boundary of spectrum division according to an extreme value of each spectrum, and establishing a corresponding filter bank;
defining a scale function and an empirical wavelet function, and decomposing the spectrum signals of each interval by using the empirical wavelet;
for the decomposed signals, calculating a baseline passing rate and a correlation coefficient based on a given baseline, removing low-frequency signals and high frequencies with insufficient correlation, denoising by adopting a semi-soft threshold function, and reconstructing the denoised signals;
after whitening pretreatment is carried out on the reconstructed signal, the reconstructed signal is sent into a sparse self-encoder for dimension reduction, and a mapping relation between the features after dimension reduction and motor faults is established;
and identifying the fault in the real-time working process of the motor based on the mapping relation.
Has the advantages that: on one hand, in the empirical wavelet transform of the self-adaptive segmentation coefficient and the threshold, the filtering and denoising of motor fault vibration signals are completed, the drifting low-frequency noise signals with the passing rate lower than a given value are removed by utilizing the base line passing rate, the high-frequency noise signals with small correlation are removed by utilizing the calculation correlation coefficient, and finally, the processing of the residual signals is completed by utilizing the semi-soft threshold wavelet method, so that useful fault signals are effectively recovered from original signals containing noise. On the other hand, in the deep learning sparse self-encoder identification framework, through the preprocessing with the data whitening process, the characteristic data are not related to each other, the redundancy among the data is reduced, through sparse dimension reduction, on the basis of ensuring the system uncertainty and the measurement noise to have fault tolerance, the most sensitive information to the motor fault is obtained, and through establishing a model of the characteristic and the motor fault, the robust mapping relation between the characteristic signal and the motor fault is ensured to be obtained.
Drawings
FIG. 1 is a schematic diagram of an asynchronous motor fault signal acquisition system;
FIG. 2 is a graphical representation of the division of the Fourier spectrum axis;
FIG. 3 is an empirical wavelet denoising process;
FIG. 4 is a signal diagram of the working state of the motor before and after de-noising;
FIG. 5 is a deep learning sparse structure self-encoder framework;
FIG. 6 is a data whitening process;
FIG. 7 is a general block diagram of sparse coding;
FIG. 8 is a hierarchical pre-training process for a deep sparse autoencoder.
Detailed Description
For better explanation of the present invention to facilitate understanding, the technical solutions of the present invention are described in detail below. The following examples are illustrative of the present invention, and the present invention is not limited to the following examples.
Aiming at the problem of motor faults commonly existing in manufacturing production equipment, the invention provides a motor fault identification method based on an adaptive spectrum segmentation denoising and deep learning self-encoder. The motor vibration signals collected from the motor unit usually contain noise, the noise mainly comes from sensor, circuit components and environmental noise, the signal waveform is disordered and has obvious burrs, fault characteristics are difficult to characterize, the final fault diagnosis is greatly influenced, signal processing is needed, and the purpose is to recover useful signal waveforms from original signals containing noise and extract characteristic quantities capable of distinguishing different faults. In the invention, firstly, the acquisition of motor fault signals is completed, and a fault diagnosis database is established; and then denoising the motor working state signal by using empirical wavelet transform (FG-EWT for short) based on the adaptive segmentation coefficient and a threshold value. Carrying out Fourier transform on an original signal x (t), normalizing a Fourier spectrum, dividing a spectrum interval by using a method based on an adaptive partition coefficient and a threshold value, decomposing the spectrum interval by using FG-EWT empirical wavelet to establish a wavelet base, calculating a baseline passing rate and a correlation coefficient to remove a low-frequency signal and a high-frequency EMF with small correlation, denoising a residual signal by using a semi-soft threshold wavelet method, and reconstructing the signal by using FG-EWT. The invention also provides a deep learning sparse self-encoder for motor fault diagnosis and identification, which firstly carries out preprocessing with a data whitening process, then reduces the dimension of an original input vector as much as possible on the premise of keeping necessary information, and finally establishes a mapping relation between the feature after dimension compression and the motor fault.
When the motor runs at a high speed, rotor body faults often occur, wherein, the common eccentric faults of the rotor can generate unbalanced magnetic pull force, so as to cause vibration, when the vibration is aggravated, the stator and the rotor are collided and rubbed, and finally the motor is damaged, in addition, the common fracture faults of the rotor conducting bars can cause the asymmetry of three-phase currents of the stator and the rotor, the torque of the motor is unbalanced, so that the starting time of the motor is prolonged, the effective torque is reduced, the slip is increased, the vibration and the noise of the motor are enhanced, the current of the stator fluctuates, the local temperature rise of the motor is caused, and the faults are fault types which need special attention in the running of the motor. The method and the device aim at processing the normal state, the rotor conducting bar breakage fault and the rotor eccentric fault state of the asynchronous motor to realize the automatic fault diagnosis of the motor.
Fig. 1 shows a motor fault signal acquisition system in an embodiment of the present invention, which is composed of a three-phase asynchronous motor, an acceleration sensor, a pincer-shaped current transformer, an oscilloscope, a multi-channel data acquisition instrument, and a computer. The experiment is carried out under the condition that the motor operates stably in no-load, mainly vibration signals and stator current signals of the motor in a normal state and a fault state are collected, and finally the collected signals are sent to a computer for denoising and fault diagnosis processing.
With reference to fig. 1, a motor fault identification method based on adaptive spectrum segmentation and denoising includes the following steps:
in the embodiment of the invention, when facing a Y801-4 asynchronous motor, the power supply frequency is 50HZ, the slip ratio is S =0.05, and the motor runs in an idle state. Respectively carrying out signal acquisition on the normal state of the motor, the breakage fault of the rotor conducting bar and the eccentric fault state of the rotor, and respectively acquiring 40 groups of data for storage and analysis in each state.
The three-phase asynchronous motor parameters are as in table 1.
TABLE 1 three asynchronous motor parameters
In a specific implementation, in order to obtain a complete and reliable vibration signal of the motor, the piezoelectric acceleration sensor selects 3 points to detect the vibration signal of the motor, wherein the vibration signal is respectively in the motor shaft direction, the vertical direction and the horizontal direction. A current transformer in the form of a clamp is used to clamp one phase of a three-phase power supply and measure the stator current flowing through that phase of the motor. The rated current of the motor is 1.6A, and the measuring range of the pincerlike current transformer is adjusted to 10A.
Collecting and primarily analyzing the working state signals of the motor:
(1) Time domain analysis of vibration signals
Firstly, the vibration signal of the working state of the motor is obtained, and a magnitude domain parameter value method is adopted to carry out time domain analysis and judgment on the vibration signal. The method comprises the following steps: dimensionless parameters (peak index, waveform index, pulse index, margin index, kurtosis index). A group of time domain indexes of the motor in the normal state and the fault state are shown in a table 2.
TABLE 2 time domain index of motor in normal and fault states
(2) Rotor conducting bar fracture fault signal acquisition and characteristic frequency analysis
The rotor conducting bar breakage fault is subjected to spectrum analysis, and the characteristic frequency and amplitude of the motor in normal operation and rotor conducting bar breakage can be obtained, and are shown in table 3.
TABLE 3 characteristic frequency and amplitude values at break of rotor conducting bars
(3) Rotor eccentric fault signal acquisition and characteristic frequency analysis
The amplitude values at the characteristic frequencies of normal operation of the motor and eccentric faults of the rotor are obtained by using the acquisition system, and are shown in table 4.
TABLE 4 characteristic frequency and amplitude at break of rotor conducting bars
And 2, carrying out Fourier transform on the original motor signal to obtain a signal frequency spectrum, dividing the frequency spectrum into a plurality of parts, and establishing a corresponding filter bank.
The invention utilizes empirical wavelet transform (FG-EWT) based on adaptive segmentation coefficient and threshold to filter and denoise motor fault vibration signals. Firstly, the frequency spectrum interval is divided based on the self-adaptive division coefficient and the threshold value. The method comprises the steps of carrying out Fourier transform on an original signal x (t), normalizing a Fourier spectrum, dividing the Fourier spectrum into an infinite number of intervals by a spectrum interval division method, and establishing a wavelet basis on the basis. In the process of dividing the frequency spectrum interval, the self-adaptive division coefficient and the threshold value are set to realize the division.
And (2) obtaining a frequency spectrum after Fourier transform. The original signal is set as X (t), the original signal comprises a normal signal and a fault signal, and the frequency spectrum after Fourier transform is set as X (f), namely:
X(f)＝FFT[x(t)] (1)
step 2 (b), determining the adaptive segmentation coefficient f according to the step 2 g ：
f d ＝y in *g z (2)
Wherein, y in The value range is as follows according to the number of the self-adaptive change of the specific situation: 2. 2.2, 2.4, 2.6 and 2.8, which are used for controlling and selecting a moderate frequency band and avoiding that the selected local extremum is positioned between two sidebands which take the fault frequency as an interval, so as to divide redundant segments. g z For a predetermined motor failure frequency, f d Is the segment frequency, n is the number of sampling points, f s Is the sampling frequency.
The division coefficient is used for taking an extreme value of the frequency spectrum in a segmented manner, so that the effect of simply enveloping the amplitude spectrum of the fault signal is achieved, the principle is simple, the operation is convenient, and the fault distribution mechanism of the vibration signal is met.
And step 2 (c), obtaining a spectrum division boundary line and dividing a spectrum interval.
With f g For dividing the number of points, X (f) is divided into m shares, i.e. each share includes f g Dividing the points, and calculating the maximum value MAX of each point i I =1,2, \8230m, m, sorting maximum points in turn according to amplitude values, and searching minimum values MIN in adjacent maximum points j And setting a threshold value y z The adjustment of the minimum value is accomplished according to the following equation (3):
finally, the MIN is used j And dividing the spectrum into different parts as a boundary of spectrum division, and establishing a corresponding filter bank.
The local minimum value of the amplitude spectrum envelope screened by the boundary factor is used as a segmentation boundary, and the method is not influenced by signal background noise due to the fact that a threshold value is set. Consider [0, π]The inner Fourier spectrum is divided into N successive segments, each segment being defined as
At intervals of each section
To indicate in order to
As a center, define a width of 2T n Transition phase T of n 。
The division of the fourier spectrum axis is shown in fig. 2.
And 3, defining a scale function and an empirical wavelet function, and decomposing the spectrum signals of each interval by using the empirical wavelet.
Step 3.1, using a band pass filter to set the scaling function and empirical wavelet function of FG-EWT, defined by (4) and (5), respectively:
n is the number of the frequency spectrum interval,
is the frequency of the nth spectral interval, T n The transition phase is specifically defined later. The function β (x) is defined as follows:
β(x)＝x 4 (35-85x+α 1 x 4 -α 2 x 3 ) (6)
wherein alpha is 1 ∈[65,75],α 2 ∈[15,25]. The function is obtained by fitting a polynomial through empirical data and is verified by implementation.
Step 3.2, simplifying the scale function and the empirical wavelet function, and aiming at the parameter T n Etc. for further implementation.
According to and
to select T n I.e. by
0 < gamma < 1, thus, for any
(4) And (5) can be simplified to (7) and (8):
the parameter γ can ensure that there is no overlap between two consecutive transition regions, so the parameter γ is set to conform to the following equation:
and 3.3, performing FG-EWT empirical wavelet decomposition.
Decomposing the signal with FG-EWT empirical wavelet to extract Empirical Mode Function (EMF), EWT definition is similar to wavelet transform, and its coefficient
Consists of the inner product of the following empirical wavelets:
finally, the approximation coefficients are represented by the inner product of the scaling function as follows:
and 4, calculating a base line passing rate and a correlation coefficient, removing the low-frequency signal and the high-frequency signal with insufficient correlation, denoising by using a soft threshold, and reconstructing a denoised signal.
And 4.1, removing low-frequency signals representing baseline drift, of which the baseline pass rate is lower than a given value, by calculating the baseline pass rate. Base line passage rate J t Is calculated as shown in equation (12):
wherein EMF n Is the nth empirical mode function, N is the length of the mode function, J x Given a baseline. The base line is a standard set for a low-frequency noise signal representing a drift of a motor signal (a signal containing a rotor lead break fault or a rotor eccentricity fault), and is at a base line J x The signal whose fluctuation degree is less than the set value is regarded as low-frequency noise and is removed, and the base line can be set according to the concrete practice. To limit the final accumulation result to a small range, a 1/2 multiplication term is added to the right side in equation (12), and since the data in the absolute value is 0 or 2, the multiplication term may not be added, or may be set to another value.
Step 4.2, calculating a correlation coefficient between the residual EMF after the low-frequency signal is removed and the original signal, and removing the high-frequency EMF with small correlation by using the correlation coefficient, wherein a correlation coefficient calculation formula is as follows:
where x (t) is the original signal including noise, M is the number of sample points of the original signal,
respectively, the raw signal and the average of the empirical mode function.
And 4.3, denoising by using the semi-soft threshold wavelet.
Wavelet threshold cancellationThe noise method is simple, small in calculation amount and wide in application in practice. Generally, the wavelet coefficient amplitude of a real signal is larger than the wavelet coefficient amplitude of noise, that is, the wavelet coefficient corresponding to a valid signal is large, and the wavelet coefficient corresponding to noise is small, and the wavelet coefficient is used for evaluation through selection of a threshold, and for selection of the threshold, there are two methods, namely a soft threshold function and a hard threshold function, wherein in the hard threshold function, the size of the wavelet coefficient absolute value is compared with a given threshold λ, if the wavelet coefficient absolute value is smaller than the threshold λ, the wavelet coefficient is set to 0, otherwise, the wavelet coefficient is not changed. In the soft threshold function, the magnitude of the wavelet coefficient absolute value is compared with a given threshold λ, if the wavelet coefficient absolute value is smaller than λ, the wavelet coefficient is set to 0, otherwise, a contraction is made toward the direction of reducing the coefficient amplitude. The invention uses a semi-soft threshold function to carry out denoising, and controls FG-EWT wavelet coefficient by adjusting parameter beta (beta is more than 0 and less than 1)
And the semi-soft threshold function is defined in formula (15) and is between the conventional hard threshold and the conventional soft threshold, so that the coefficient is closer to the original coefficient, and the noise reduction effect is ensured to the maximum extent.
Wherein sgn is a sign function, see formula (13) above; lambda is a threshold value, beta is an adjustment parameter,
are the FG-EWT empirical wavelet decomposition coefficients.
And 4.4, reconstructing the signal by using FG-EWT empirical wavelets.
And finally, reconstructing the denoised signal by using FG-EWT. See in particular the following formula:
f 0 (t) and f k (t) are the 0 th and k-th components of the reconstructed empirical mode, respectively.
The reconstructed signal is then:
the FG-EWT empirical wavelet denoising process is shown in figure 3. The working state signals of the motor before and after denoising are shown in FIG. 4.
And 5, carrying out whitening pretreatment on the reconstructed signal, sending the signal into a sparse self-encoder for dimensionality reduction, and establishing a mapping relation between the dimensionality reduced characteristic and the motor fault.
The deep learning sparse structure self-encoder framework proposed by the invention is shown in fig. 5.
And 5.1, carrying out whitening data preprocessing.
The denoised data is again pre-processed pre-diagnostically using a data whitening process. This is a linear transformation for transforming random variables with a known covariance matrix into a new set of variables with unity covariance, the purpose of data whitening is to make the input data redundancy low, the data are uncorrelated with each other, and all features have the same characteristic variance. This process is referred to herein as "whitening" because the input vector is converted to a white noise vector.
The orthogonal matrix U is obtained by Principal Component Analysis (PCA) of the raw input data, with U making the input features uncorrelated, as in equation (18).
x i And denoising the reconstructed signal for the ith motor fault.
To make each input feature have a unit variance, it is then scaled as follows:
in the formula, λ j Is the eigenvalue corresponding to the jth eigenvector obtained from PCA,
is the processed jth whitened data sample. The data whitening process is completed based on principal component analysis, so that decorrelation and sphericization processing of data is realized, a preprocessing data set with low redundancy is provided, and then verification, training and testing of a subsequent network are completed.
In the process of the invention, and is less than 1e -12 The feature value of (2) is discarded corresponding to the whitened data, and the rest is the data set which is stored in the original information to the maximum extent. The data whitening process is as in fig. 6.
And 5.2, establishing a sparse autoencoder to perform sparse dimension reduction.
The sparse dimension reduction process is used for compressing the dimension of the motor signal characteristics, so that the information most sensitive to the motor fault is obtained, and meanwhile, the fault tolerance is realized on the uncertainty and the measurement noise of the system.
The invention provides a deep neural network based on a sparse self-encoder, which is used in dimension compression application. The first hidden layer is used for executing the fusion of characteristics such as the working frequency of the motor, and the hidden layers from the second layer to the k-th layer are used for executing the characteristic compression. The overall structure of sparse coding is shown in fig. 7.
In the implementation process, expressions of the sparse self-encoder, the sparse activation function and the target function for dimension reduction are as follows:
(1) Calculating an average activation function
Let the activation function of the jth hidden layer unit of the network be h j (x i ) Wherein x is i Defining an average activation function for the ith input
Wherein m is the number of samples,
is the average activation function (average over the training period) of the jth hidden layer unit. Then the mandatory constraints are set as follows:
where ρ is a sparsity parameter, ρ =0.05 for sigmoid activation functions. In the present invention, the average activation function of each hidden neuron is close to 0, and therefore, the activation function of the hidden layer unit is also substantially close to 0, which is experimentally set by applying the verification data set.
(2) Adding penalty conditions
The self-encoder mainly has the main function of performing dimensionality reduction learning on high-dimensional data, in a network structure, if the number of nodes in a hidden layer is more than that of nodes in an input layer, an algorithm loses the automatic learning capacity, in order to avoid the problem, sparsity constraint is utilized to enable neurons in the hidden layer to be in a suppression state most of time, a penalty condition CF is specifically added to measure the similarity between the average activation output and the sparsity of the nodes in the hidden layer, in order to ensure that the neurons in the hidden layer are in lower liveness, the smaller the CF dispersion is, the better the smaller the CF dispersion is, the smaller the CF dispersion represents the similarity between the average activation output and the sparsity of the nodes in the hidden layer, and the smaller the CF dispersion is, the smaller the more the probability is
The difference between the motor fault characteristic and the rho is smaller, so that the self-learning capability of the motor fault characteristic is stronger.
Adding penalty condition to optimized objective function
For punishing
Course of significant deviation of rhoDegree, defined as follows:
wherein r is the number of hidden layer neurons, r 1 ，r 2 Is a random variable, j is the serial number of the jth hidden layer unit,
represented as a sparse regularization term.
(3) Defining a sparse target cost function
The means for constraining the potential characterization information is typically to make it sparse or low dimensional. In the invention, a loss function is reconstructed by introducing a sparse induction term and a weight attenuation function, and a target cost function is defined as follows:
wherein,
is a reconstruction loss function;
is the weight decay function (L2 regularization of all weights);
is a sparse penalty equation, see formula (23), for denoising; alpha and gamma are regularization parameters that balance reconstruction accuracy and application constraints. The optimal parameters are obtained by validating the data set. Weight attenuation term
For avoidingOverfitting, defined as follows:
wherein,
represents the weight W l Each element in (1), S l Indicates the number of cells in the l-th layer.
In the present invention, the sparse self-encoders defined above are stacked together to form an in-depth architecture for learning the mapping relationship between the input signal and the output signal. And the first hidden layer performs motor fault signal characteristic fusion, so that nonlinear dimension reduction is realized. The second hidden layer compresses the low-dimensional features learned from the first hidden layer, the next hidden layers respectively further compress the low-dimensional features from the previous hidden layer, and finally, a robust feature space from the last hidden layer is obtained, and the space can retain useful information reflecting the measured mapping relation.
In addition, the hidden layer in sparse nonlinear dimension reduction is not necessarily smaller than the input layer. The method introduces sparsity constraint into an objective function, and in dimension reduction, a sparse activation function is used for selecting the optimal number of nodes of an implicit layer in a training process.
The motor failure frequency and its corresponding characteristics are input to the self-encoder, defined as follows:
wherein
Is the motor status signal frequency after the i-th (i =1, \ 8230;, n) de-noising reconstruction in the r-th sample, i.e. the motor operation signal including the fault and normal signals,
a series of consecutive high-dimensional features, which are used as input to a sparse self-encoder.
Specifically, the loss function of the p-th layer of the sparse self-encoder is reconstructed as follows:
where p = {1, \8230;, k }, where k is the last layer in the dimensionality reduction, Q is the number of samples involved in the training, g (·), f (·) are the decoder and encoder functions, respectively,
representing the low-dimensional features established at the p-1 level for the r-th sample, wherein,
encoder function f p Set to ReLU, which supports sparse representation of the input signal, decoder function g p And setting the function as purelin to reconstruct the true value of the input.
The objective function in the formula is defined as formula (24). For training sparse autoencoders. Obtaining potential tokens from the last hidden layer in the dimension reduction
I.e., k-th layer, and then fed back to the map learning module.
Step 5.3, establishing a mapping relation between the characteristics after dimension compression and motor faults
(1) Pre-training learning of mapping relationships
In the process of establishing the mapping relation between the feature after dimension compression and the motor fault, firstly, pre-training learning of the mapping relation is carried out, and the main purpose of the learning of the mapping relation is to learn the feature after dimension reduction
And motor fault parametersThe deep learning sparse structure self-encoder with pre-training is used for training the nonlinear mapping relation. The "tanh" function is chosen as the excitation function, and the cost function for each layer is defined as follows:
also with the weight attenuation function described in equation (25), m hidden layers are defined, and the reconstructed layer loss functions are defined as follows:
wherein q = { k +1, \8230, k + m } is a parameter in the mth layer in the mapping relation learning module; g (-) and f (-) are decoder and encoder functions, respectively,
representing the low-dimensional features built at the level q-1 by the r-th sample,
is the tagged output vector for the r-th sample.
The invention defines different layers to carry out effective mapping relation learning on the global nonlinearity, realizes optimization by stacking different layers, and further reduces errors in the following layers. And (3) pre-training all layers by adopting a full sample gradient BP algorithm, and once the optimal parameters are obtained, fine-tuning the whole network again to optimize all the layers.
(2) Hierarchical network training and fine-tuning
After the pre-training is performed, the whole network is subjected to layered training and fine tuning.
And combining sparse dimension reduction and relationship learning into a deep neural network. The training process implements a layered training scheme, and the process of layered training learning for deep sparse autoencoder networks is given in fig. 8. The first two hidden layers for coding are pre-trained to execute nonlinear dimension reduction, the last three layers are trained to learn the mapping relation between the compressed dimension characteristics and the motor fault parameters, and in this way, the deep sparse self-encoder network reserves required information to establish the mapping relation between the learned robust characteristics and the motor fault.
The sparse autoencoder model and the layered training proposed by the present invention are specifically shown in fig. 8. Therefore, the hidden layers can be trained one by one, and therefore a more efficient and accurate training process is obtained.
After pre-training, the whole deep network is fine-tuned to optimize all layers with the objective function at the same time, as follows:
wherein,
is the estimated motor fault parameter output vector for the r-th sample,
is the labeled output vector of the r-th sample, g (-) and f (-) are the decoder and encoder functions, respectively. The fine adjustment and the joint optimization of the objective function are completed through a formula (30), and the whole network is ensured to learn characteristic parameters
And the mapping relation between the motor fault and the motor fault is finely adjusted, so that better and accurate motor fault diagnosis and identification are realized.
In the specific implementation, the sparse constraint is only applied in the dimension reduction, and in the mapping relationship learning, it is crucial to obtain a corresponding nonlinear relationship between the feature to be dimension reduced and the output signal, for this process, the sparse constraint or the sparse activation function will not perform well, and an effective mapping relationship cannot be obtained in the training, so the sparse induction term is not used in the mapping relationship learning, as shown in equations (28) and (29). Furthermore, pre-training is performed and the entire network is fine-tuned using equations (30) and (31), except for those sparse terms for which the lower layers have been pre-trained (the output of most hidden nodes will be 0 after pre-training), the fine-tuning mainly affects the higher layers (the mapping learning part), thereby obtaining a mapping between the features after establishing the compression dimension and the motor fault.
In the embodiment of the invention, the stator current signals acquired by the experiment are used as research objects of the deep learning self-encoder, and each group of data extracts the characteristic frequency of the stator current: the amplitude values corresponding to the characteristic frequencies of the collected stator current signals are normalized and then used as the input quantity of the deep learning self-encoder, so that the number of neuron nodes of an input layer is 4. The number of neuron nodes of the output layer is determined by the motor state type, and the motor state includes a normal state, a rotor lead breakage fault and a rotor eccentricity fault, which are respectively expressed by (1 0), (0) and (0) 0, so that the number of neuron nodes of the output layer is 3. The test results are shown in Table 5.
TABLE 5 test results
Therefore, the method is used for motor fault diagnosis, the network training is completed through 80 steps of iteration, the error precision is 0.00036, and the network training time is 6.287s.
Claims (10)
1. A motor fault identification method based on self-adaptive spectrum segmentation and denoising is characterized by comprising the following steps:
(1) Collecting vibration signals and stator current signals of the motor in a normal state and a fault state under the condition that the motor is in no-load stable operation;
(2) Fourier transform is carried out on the original motor signal to obtain a signal frequency spectrum X (f), and the signal frequency spectrum X (f) is obtained according to the Fourier transformDetermination of adaptive partition coefficients f from spectral and sampling information g The spectrum is divided into shares such that each share contains f g Dividing points, determining a boundary of spectrum division according to an extreme value of each spectrum, and establishing a corresponding filter bank;
(3) Defining a scale function and an empirical wavelet function, and decomposing the spectrum signals of each interval by using the empirical wavelet;
(4) For the decomposed signals, calculating a baseline passing rate and a correlation coefficient based on a given baseline, removing low-frequency signals and high frequencies with insufficient correlation, denoising by adopting a semi-soft threshold function, and reconstructing the denoised signals;
(5) After whitening pretreatment is carried out on the reconstructed signal, the reconstructed signal is sent into a sparse self-encoder for dimension reduction, and a mapping relation between the features after dimension reduction and motor faults is established;
(6) And identifying the fault in the real-time working process of the motor based on the mapping relation.
2. The method for identifying motor faults based on adaptive frequency spectrum division and denoising as claimed in claim 1, wherein in the step (1), 3 points are selected by an acceleration sensor to detect vibration signals of the motor, which are respectively in a motor shaft direction, a vertical direction and a horizontal direction, one phase of a three-phase power supply of a three-phase asynchronous motor is buckled by a pincer-shaped current transformer, and stator currents of the phase flowing through the motor are measured; and respectively acquiring and storing data of the normal state of the motor, the breakage fault of the rotor conducting bar and the eccentric fault state of the rotor, and establishing a fault diagnosis database.
3. The method for motor fault recognition based on adaptive spectrum segmentation and denoising as claimed in claim 1, wherein in step (2), the adaptive segmentation coefficient f is g The calculation formula of (c) is as follows:
f d ＝y in *g z
wherein, y in Is an adaptively changing number within the range of 2, 2.2, 2.4, 2.6 and 2.8, f d Is the fractional frequency, g z For a predetermined motor failure frequency, n is the number of sampling points, f s Is the sampling frequency.
4. The method for identifying motor faults based on adaptive spectrum segmentation and denoising as claimed in claim 1, wherein the step (2) of determining the boundary of the spectrum division according to the extreme value of each spectrum comprises:
for the divided frequency spectrum, the maximum value MAX of each share is obtained i I =1,2, \ 8230, m is the number of frequency spectrum parts, maximum points are sequentially sorted according to the amplitude value, and the minimum value MIN in adjacent maximum points is searched j And setting a threshold value y z The adjustment of the minimum value is completed according to the following formula:
finally, MIN is used j As the dividing line of the spectrum.
5. The method for motor fault recognition based on adaptive spectrum segmentation denoising as claimed in claim 1, wherein in the step (3), defining the scale function and the empirical wavelet function comprises:
setting a scaling function using a band-pass filter
And empirical wavelet function
Wherein n is the number of the frequency spectrum interval,
is the frequency of the nth spectral interval, T n For the transition phase, the function β (x) is defined as follows: β (x) = x 4 (35-85x+α 1 x 4 -α 2 x 3 ) In which α is 1 ∈[65,75],α 2 ∈[15,25]；
According to and
to select T n Is provided with
0 < gamma < 1 for arbitrary
Scale function
And empirical wavelet function
The method is simplified as follows:
the parameter γ is used to ensure that there is no overlap between two consecutive transition regions, and the parameter γ follows the following equation:
6. the method for identifying motor faults based on adaptive spectrum segmentation and denoising as claimed in claim 1, wherein in the step (4), the baseline pass rate is calculated according to the following formula:
wherein EMF n Is the nth empirical mode function, N is the length of the empirical mode function, J x For a given baseline, sgn is a sign function,
7. the method for identifying motor faults based on adaptive spectrum segmentation and denoising as claimed in claim 1, wherein in the step (4), the correlation coefficient is calculated according to the following formula:
where x (t) is the original signal including noise, M is the number of sample points of the original signal,
the average of the original signal and the empirical mode function, respectively.
8. The method for identifying motor faults based on adaptive spectrum segmentation and denoising as claimed in claim 1, wherein in the step (4), the semi-soft threshold function is as follows:
wherein, lambda is a threshold value, beta is an adjusting parameter,
for empirical wavelet decomposition coefficients, sgn is a sign function,
9. the method for identifying motor faults based on adaptive spectrum division denoising as claimed in claim 1, wherein in the step (5), when a sparse self-encoder is used, penalty conditions are added to the optimization objective function of the sparse self-encoder
For punishing
The degree of significant deviation ρ is defined as follows:
wherein r is the number of neurons in the hidden layer, r 1 ，r 2 Is a random variable, j is the serial number of the jth hidden layer unit,
representing a sparse regularization term, p is a sparse parameter,
is the average activation function value of the jth hidden layer element.
10. Motor fault identification system based on self-adaptation spectrum segmentation is denoised, its characterized in that includes:
the signal acquisition system comprises an acceleration sensor and a pincerlike current transformer and is used for respectively acquiring a vibration signal and a stator current signal of the motor in a normal state and a fault state; and
signal processing apparatus comprising a processor, a memory and a computer program, wherein the computer program is stored in the memory and configured to be executed by the processor, the program when executed by the processor implementing the steps of:
fourier transform is carried out on an original motor signal to obtain a signal frequency spectrum X (f), and an adaptive division coefficient f is determined according to the frequency spectrum and sampling information g The spectrum is divided into shares such that each share contains f g Dividing points, determining a boundary of spectrum division according to an extreme value of each spectrum, and establishing a corresponding filter bank;
defining a scale function and an empirical wavelet function, and decomposing the spectrum signals of each interval by using the empirical wavelet;
for the decomposed signals, calculating a baseline passing rate and a correlation coefficient based on a given baseline, removing low-frequency signals and high frequencies with insufficient correlation, denoising by adopting a semi-soft threshold function, and reconstructing the denoised signals;
after whitening pretreatment is carried out on the reconstructed signal, the reconstructed signal is sent into a sparse self-encoder for dimension reduction, and a mapping relation between the features after dimension reduction and motor faults is established;
and identifying the fault in the real-time working process of the motor based on the mapping relation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210722448.9A CN115166514A (en) | 2022-06-24 | 2022-06-24 | Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210722448.9A CN115166514A (en) | 2022-06-24 | 2022-06-24 | Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115166514A true CN115166514A (en) | 2022-10-11 |
Family
ID=83487757
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210722448.9A Pending CN115166514A (en) | 2022-06-24 | 2022-06-24 | Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115166514A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116955995A (en) * | 2023-09-20 | 2023-10-27 | 深圳市嘉友锦磁科技有限公司 | Three-phase direct current brushless motor inverter fault diagnosis method |
| CN117630554A (en) * | 2023-12-07 | 2024-03-01 | 深圳市森瑞普电子有限公司 | Testing device and testing method for conductive slip ring |
-
2022
- 2022-06-24 CN CN202210722448.9A patent/CN115166514A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116955995A (en) * | 2023-09-20 | 2023-10-27 | 深圳市嘉友锦磁科技有限公司 | Three-phase direct current brushless motor inverter fault diagnosis method |
| CN116955995B (en) * | 2023-09-20 | 2024-01-05 | 深圳市嘉友锦磁科技有限公司 | Three-phase direct current brushless motor inverter fault diagnosis method |
| CN117630554A (en) * | 2023-12-07 | 2024-03-01 | 深圳市森瑞普电子有限公司 | Testing device and testing method for conductive slip ring |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112785016B (en) | New energy automobile maintenance and fault monitoring and diagnosis method based on machine learning | |
| CN107702922B (en) | Rolling bearing fault diagnosis method based on LCD and stacking automatic encoder | |
| CN115166514A (en) | Motor fault identification method and system based on self-adaptive spectrum segmentation and denoising | |
| CN112326210A (en) | Large motor fault diagnosis method combining sound vibration signals with 1D-CNN | |
| CN109145706A (en) | A kind of sensitive features selection and dimension reduction method for analysis of vibration signal | |
| CN109884419B (en) | Smart power grid power quality online fault diagnosis method | |
| CN111413089A (en) | Gear fault diagnosis method based on combination of VMD entropy method and VPMCD | |
| CN105206270A (en) | Isolated digit speech recognition classification system and method combining principal component analysis (PCA) with restricted Boltzmann machine (RBM) | |
| CN113673346A (en) | Motor vibration data processing and state recognition method based on multi-scale SE-Resnet | |
| Jiang et al. | A fault diagnostic method for induction motors based on feature incremental broad learning and singular value decomposition | |
| CN114386452B (en) | Nuclear power circulating water pump sun gear fault detection method | |
| CN115840120A (en) | High-voltage cable partial discharge abnormity monitoring and early warning method | |
| CN118296470B (en) | Intelligent classification method for multi-classifier fusion | |
| CN116758922A (en) | Voiceprint monitoring and diagnosing method for transformer | |
| CN112747921A (en) | Multi-sensor mechanical fault diagnosis method based on NA-MEMD | |
| CN110458071B (en) | DWT-DFPA-GBDT-based optical fiber vibration signal feature extraction and classification method | |
| CN114201993B (en) | Three-branch attention feature fusion method and system for detecting ultrasonic defects | |
| Zhao et al. | Bearing fault diagnosis based on mel frequency cepstrum coefficient and deformable space-frequency attention network | |
| CN117219124B (en) | Switch cabinet voiceprint fault detection method based on deep neural network | |
| CN112560674B (en) | Method and system for detecting sound signal quality | |
| CN114036977A (en) | Fan power quality disturbance monitoring method based on two-dimensional image coding features | |
| CN118051839A (en) | Power cable insulation aging identification method based on grounding current and neutral current | |
| CN110222390B (en) | Gear crack identification method based on wavelet neural network | |
| CN116559728A (en) | Power transmission line fault diagnosis method based on wavelet transformation-multilayer perceptron | |
| CN116625678A (en) | Fault diagnosis method and system based on average kurtosis deconvolution network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |