CN-122019995-A - Time-frequency double-robust filtering method for demagnetization diagnosis of permanent magnet motor

CN122019995ACN 122019995 ACN122019995 ACN 122019995ACN-122019995-A

Abstract

The invention relates to a time-frequency double-robust filtering method for permanent magnet motor demagnetization diagnosis, which belongs to the field of permanent magnet motor fault diagnosis and comprises the steps of loading a permanent magnet motor original signal, carrying out time-domain filtering through a self-adaptive Hampel filter, carrying out framing windowing processing, converting the time-domain filtered signal into a frequency-domain signal, training a Gaussian process regression model by using a normal sample, calculating a reference noise spectrum, carrying out frequency-domain deep noise reduction by adopting weighted spectral subtraction, and reconstructing the filtered signal. The method can effectively remove sudden abnormal values and pulse interference, and can carry out frequency domain deep noise reduction through weight spectrum subtraction, deep suppression in a strong noise frequency band and weak suppression or retention in a weak noise frequency band and a frequency band possibly containing fault characteristics. The invention can keep and even strengthen weak characteristic components related to demagnetizing faults while suppressing background noise to the maximum extent, thereby fundamentally avoiding the loss of diagnostic information and improving the reliability of diagnosis.

Inventors

YU JUNTAO
NI SHENG
WANG LI
WANG YINAN
ZHENG LINZHI
Zhang Yuanziman

Assignees

山东大学
山东大学威海工业技术研究院

Dates

Publication Date: 20260512
Application Date: 20260121

Claims (10)

1. The time-frequency double-robust filtering method for the demagnetization diagnosis of the permanent magnet motor is characterized by comprising the following steps of: s1, loading a permanent magnet motor original signal, and performing time domain filtering on the permanent magnet motor original signal through a self-adaptive Hampel filter; S2, loading time-domain filtered data, carrying out framing and windowing processing, and converting the time-domain filtered signals into frequency domain signals; S3, training a Gaussian process regression model by using a normal sample; S4, calculating a reference noise spectrum; s5, generating frequency weights based on a Gaussian process regression model, and carrying out frequency domain depth noise reduction by adopting weighted spectral subtraction; s6, reconstructing the filtered signal.
2. The time-frequency double-robust filtering method for demagnetization diagnosis of permanent magnet motor according to claim 1, wherein in step S1, firstly, original signals of the permanent magnet motor are extracted, and the process is as follows: the method comprises the steps of recording a data matrix stored by original signals of a permanent magnet motor as D, wherein the dimension is n multiplied by m, n is 2048 rows and represents signal data, m is columns, each column represents an independent signal, 2049 rows are additionally arranged at the bottom of the 2048 rows and are set as label rows, and a data matrix D' is formed; The permanent magnet motor is a 6-slot 4-pole surface-mounted permanent magnet motor, and the number of the tags is 6, namely, tag 0 indicates that the motor does not demagnetize, tag 1 indicates that the motor demagnetizes a single magnetic pole, tag 2 indicates that two adjacent magnetic poles of the motor demagnetize, tag 3 indicates that two magnetic poles at opposite positions demagnetize, tag 4 indicates that three magnetic poles demagnetize, and tag 5 indicates that four magnetic poles of the motor demagnetize; Decomposing the data matrix D ' into signal data X and tag data L, wherein x=d ' (1:2048:) represents all columns of rows 1 to 2048 of the data matrix D ', and tag data l=d ' (2049:) represents all columns of row 2049 of the data matrix D '; an all-zero matrix Y of the same dimension as the signal data X is initialized as a processed data storage matrix in a subsequent process.
3. The method of time-frequency dual-robust filtering for demagnetization diagnosis of permanent magnet motor according to claim 2, wherein in step S1, for each column of independent signal X in signal data X, sequentially calling improved adaptive weight Hampel filtering function to obtain filtered signal Y, storing each column of filtered signal Y in corresponding position of matrix Y, finally adding original tag data L to end of matrix Y of filtered signal to combine into new matrix And (3) saving the result as a new data file to complete the whole filtering and data reconstruction process.
4. A time-frequency dual-robust filtering method for permanent magnet machine demagnetization diagnosis according to claim 3, characterized in that the filtering process of the improved adaptive weight Hampel filtering function comprises: S12, traversing each point i in the independent signal x, calculating the left boundary l=max (1, i-h) of the window, and simultaneously calculating the right boundary r=min (n, i+h), h being half-width, h=0.5 # -1), Representing the size of a sliding window, extracting window data And the actual window length n=r-l+1; s13, extracting the median of the window data Calculating the absolute deviation of each data window point from the median : Wherein the method comprises the steps of Is a window of data The first of (3) An element; S14, based on absolute deviation Calculating weight, and introducing a weight attenuation coefficient alpha, wherein alpha is a positive constant; Wherein, the Is the first Weights of the individual points; S15, matching window data Sorting according to the ascending order of the values to obtain sorted values For all of Normalization processing is carried out to obtain normalized weights ; Calculating an accumulated weight: Wherein, the Is in front of Cumulative weight of the ranking points, followed by finding the median position The method comprises the following steps: Wherein, the The cumulative weight value representing the median position, ; Let weighted median , Representing the first of the ordered data sequences A value of a point; S16, calculating absolute deviation based on the weighted median: Wherein the method comprises the steps of Is the first Absolute deviation of the individual points from the weighted median; For a pair of Performing ascending order to obtain Corresponding to the normalized weight after sequencing The cumulative weight is then calculated: finding the final weighted median position The method comprises the following steps: Representing the cumulative weight value when the cumulative weight reaches or exceeds all weights and half in the ordered absolute deviation sequence; Let the final weighted median mad weighted be: ; S17, judging the abnormal condition, when the abnormal condition is satisfied When in use, order Otherwise keep Wherein, the method comprises the steps of, For the ith point in x, y i is the ith point of y, Is the MAD threshold coefficient.
5. The method of time-frequency dual-robust filtering for permanent magnet motor demagnetization diagnosis of claim 4, characterized in that in step S2, a result after time-domain filtering is performed Taking the= [ Y; L ] as input, letting D filtered = Y, storing the time-domain filtered data in a matrix D filtered , wherein the dimension is N multiplied by M, and initializing an enhanced data matrix Y enhanced as a 0 matrix; Dividing the data column of D filtered according to the six types of labels to obtain the third Data column index set corresponding to class label ; Defining a hamming window function to reduce spectral leakage caused by framing: Wherein the method comprises the steps of Is the value of the hamming window function at the discrete time index point n, L 1 represents the frame length; and carrying out FFT calculation on the frames subjected to windowing and framing, and converting the time domain signals into frequency domain complex sequences.
6. The time-frequency double-robust filtering method for demagnetization diagnosis of permanent magnet motor according to claim 5, wherein the windowing and framing and FFT calculation process is as follows: First, the columns of D filtered are taken as signal frames For signal frames Windowing: Wherein the method comprises the steps of Is a windowed signal; The FFT transform formula is: Wherein, the Is the first in the frequency domain The complex representation of the frequency components, i.e. the complex spectrum, j is the imaginary unit.
7. The time-frequency double-robust filtering method for demagnetization diagnosis of permanent magnet motor according to claim 6, wherein the specific process of step S3 is: S31, collecting the power spectrum characteristics of the normal sample data corresponding to the non-demagnetizing prototype, and marking as The method specifically comprises the following steps: To the front Framing each sample into frames Frame, T= (M-1-L 1 )/H+1 frame, H represents frame shift, T frame signal is recorded as And (3) windowing: Representing the multiplication of the t frame signal with the hamming window element by element; Calculating a power spectrum: Wherein, the Is the FFT result of the t-th frame signal, Power spectrum for the t frame; calculating an average power spectrum: Wherein, the Average power spectrum for the kth sample; s32, constructing a training data set Normalized frequency points: Wherein, the Representing normalized frequency points, j representing an index; Constructing an input feature matrix: Wherein the method comprises the steps of Represents the Kronecker product of the equation, Representing a row vector consisting of K1 s, and outputting the row vector as ; S33, training a Gaussian process regression model, wherein the formula is as follows: wherein GPR represents an optimal Gaussian process regression model found by an optimization process, And theta is a hyper-parameter, which is a square index kernel function.
8. The method of time-frequency dual-robust filtering for permanent magnet machine demagnetization diagnosis according to claim 7, characterized in that in step S4, for all normal samples, a reference noise spectrum is calculated as a global estimate of the noise power spectrum: 。
9. The time-frequency double-robust filtering method for demagnetization diagnosis of permanent magnet motor according to claim 8, wherein the implementation process of step S5 is as follows: S51, regarding the current sample Wherein ct is 0,1,2,3,4 or 5, and the frame is windowed Frame represents a framing function, and then short-term spectrum is calculated Sum power spectrum Wherein F () represents a discrete Fourier transform and F c () represents (Ω, t) represents coordinates on a frequency-time plane for indexing the time and frequency dimensions of the short-time spectrum; s52, predicting frequency weight: The frequency point omega is taken as input and substituted into a trained Gaussian process regression model to obtain a predicted initial weight value of the frequency point, and normalized to obtain normalized frequency weight : Min (W (ω)) represents the minimum value of W (ω), and max (W (ω)) represents the maximum value of W (ω); S53, performing weighted spectral subtraction using the following formula: Wherein the method comprises the steps of Is a noise suppression parameter for controlling the pre-noise suppression amplitude; realizing frequency self-adaption, wherein beta represents a minimum power threshold value; Representing the enhanced signal power spectrum after weighted spectral subtraction.
10. The time-frequency double-robust filtering method for permanent magnet motor demagnetization diagnosis according to claim 9, wherein the process of reconstructing the signal is: the enhanced spectrum is obtained using: e represents the sign of the natural base; the spectrum will then be enhanced Performing inverse Fourier transform to obtain enhanced time domain frame signal The complete time domain signal is synthesized by overlapadd, overlap-add function: h represents frame shift, M represents output signal length; And finally reconstructing a complete data matrix: 。

Description

Time-frequency double-robust filtering method for demagnetization diagnosis of permanent magnet motor Technical Field The invention relates to a time-frequency double-robust filtering method for permanent magnet motor demagnetization diagnosis, and belongs to the technical field of permanent magnet motor fault diagnosis. Background The permanent magnet synchronous motor is widely applied in the fields of new energy automobiles, industrial driving and the like due to high power density, high efficiency and excellent control performance. However, the permanent magnet synchronous motor may have an irreversible demagnetization fault during operation, which is one of the most dominant fault modes. The demagnetization of the permanent magnet can lead to weakening of the magnetic flux of the motor, increase of torque fluctuation and reduction of efficiency, and even cause the motor to run away when serious, so that huge economic loss and safety accidents are caused. Therefore, the method has extremely important engineering value and economic significance for early, accurate and reliable demagnetization fault diagnosis of the permanent magnet synchronous motor. At present, fault diagnosis based on signal processing is one of main technical means of demagnetization detection of a permanent magnet motor. The basic principle is that demagnetizing faults change the air gap field of the motor, so that specific fault characteristic frequency components are introduced into signals such as stator current, vibration, noise and the like. By collecting and analyzing the signals, diagnosis and identification of faults can be realized. However, in practical engineering applications, using the collected motor noise signal for demagnetization fault diagnosis mainly faces the following two challenges: First, there is a strong background noise and interference, and motors typically operate in complex industrial environments where the monitoring signals are extremely vulnerable to switching noise of the power electronics, coupled vibration of other mechanical devices, transmission line interference, and contamination by noise of the sensor itself. These strong background noise can overwhelm the weak early fault features, particularly the specific frequency components associated with the fault, resulting in difficult feature extraction and reduced diagnostic accuracy. And secondly, pulse abnormal disturbance, namely, during the running process of the motor, the events such as abrupt change of load, poor instantaneous contact of a circuit or external impact and the like can introduce non-Gaussian and pulse abnormal values into a monitoring signal. These outliers are not periodic, but are of high energy, which severely distorts the statistical properties of the signal, interfering with the determination of the true characteristics of the signal. To address the above challenges, the prior art mainly includes spectral subtraction and various filtering techniques, but they all have significant limitations. Traditional spectral subtraction achieves noise reduction by subtracting an estimated noise power spectrum from the power spectrum of the noisy signal. Although simple and effective, it has a core disadvantage in that a fixed spectral subtraction factor needs to be set manually. Under the actual complex motor working condition, the noise is not stable and unchanged, the fixed coefficient can cause the problem of signal distortion caused by residual or over-suppression of the environmental noise, and the optimal noise reduction effect is difficult to obtain in the global. Traditional filtering techniques, such as wiener filtering or kalman filtering, either rely on accurate system models that are difficult to build in practice, or assume that the noise follows a specific distribution like a gaussian distribution and thus are less robust to impulsive anomalies. In addition, standard median filtering suppresses impulse noise, but excessively smoothes the signal, losing a great deal of useful fault detail information. While the model-based optimal filter can degrade dramatically when it encounters non-gaussian pulse disturbances. In view of the above, a significant problem in the prior art is the lack of a comprehensive solution capable of cooperatively handling the burst noise and the strong background noise. Therefore, a new technical scheme is needed to effectively solve the problem of demagnetization fault diagnosis of the permanent magnet motor in the coexistence environment of strong noise and pulse interference. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a time-frequency double-robust filtering method for demagnetization diagnosis of a permanent magnet motor, which adopts a signal double-robust enhancement strategy of serial synergy of a time domain and a frequency domain, firstly provides a self-adaptive Hampel filter to perform time domain pretreatment on an original signal so as to effec