A method for detecting weak faults in rotating machinery based on signal reconstruction using variable-parameter time-frequency manifolds

Abstract

The present invention discloses a method for detecting weak faults of rotating machinery based on variable parameter time-frequency manifold signal reconstruction, comprising: step (1), extracting components within the fault band; step (2), constructing a high-dimensional TFD matrix; step (3), extracting TFM features; step (4), denoising TFM features; step (5), reconstructing fault component TFD; step (6), reconstructing the fault component time domain signal. The technical method has at least the following advantages: good universality for different mechanical vibration signals, high computational efficiency, high signal-to-noise ratio, and the ability to quantitatively analyze the severity of the fault.

Description

Technical Field

[0001] The present invention relates to the field of fault detection, and in particular to a method for detecting weak faults of rotating machinery based on variable parameter time-frequency manifold signal reconstruction. Background Art

[0002] Vibration monitoring technology is often used in rotating machinery fault diagnosis. By analyzing the vibration signal, the operating status of the rotating machinery can be monitored and faults can be discovered in time, which is conducive to reducing downtime and ensuring efficient production. When a local fault occurs in the rotating machinery at a constant speed, a periodic transient pulse component will appear in the vibration signal. However, due to the complex mechanical structure and working environment, the transient pulse is often submerged in a large amount of noise, making the fault information appear very weak in the vibration signal, which brings challenges to mechanical fault detection. Therefore, extracting the fault transient pulse component from the vibration signal is the key to realizing the detection of weak faults in rotating machinery.

[0003] Time-frequency transformation methods, such as short-time Fourier transform (STFT), can transform a one-dimensional time-domain vibration signal into a time-frequency distribution (TFD) in the time-frequency domain. They are often used to extract the time-frequency pattern of fault transient pulses, that is, the pulse area that appears periodically in a specific frequency band. However, noise is also distributed in the TFD of the vibration signal, interfering with the identification of fault transient pulses. Traditional denoising methods only have the effect of bandpass filtering, that is, they can only remove noise outside the frequency band where the fault is located, but cannot remove noise within the fault frequency band (i.e., in-band noise).

[0004] Time-frequency manifold (TFM) is a method that aims to remove the noise in the fault band in the time-frequency domain. It extracts the intrinsic manifold structure embedded in the high-dimensional TFD through manifold learning nonlinearity, and reduces the noise in the fault band while retaining the fault transient pulse. The specific steps of TFM are: 1) using phase space reconstruction (PSR) to convert the one-dimensional vibration signal to a high-dimensional space; 2) using time-frequency transformation to process each dimensional signal in the high-dimensional space to obtain a high-dimensional TFD; 3) performing manifold learning on the high-dimensional TFD to obtain low-dimensional TFM features.

[0005] Traditional technologies have the following technical problems:

[0006] TFM can reduce the noise in the fault band in the signal time-frequency distribution to a certain extent, but it still has the following disadvantages: 1) The parameters of the time-frequency transformation are selected manually and are not universal for different signals; 2) The amount of high-dimensional TFD data is large, and the computational burden of manifold learning is heavy; 3) The neighbor point parameters of manifold learning need to be optimized, which increases the computational burden; 4) A small amount of fault band noise still remains in the TFM feature; 5) The TFM feature only reflects the intrinsic time-frequency manifold structure of the fault transient pulse, and its amplitude is far from the TFD of the signal, so it cannot be used for quantitative analysis of the severity of mechanical faults. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a method for detecting weak faults of rotating machinery based on reconstruction of variable parameter time-frequency manifold signals. This method aims at the problems existing in TFM, proposes a variable parameter time-frequency transformation method with universal applicability to signals to construct a high-dimensional TFD, and performs adaptive threshold denoising on the obtained two-dimensional TFM features, and finally reconstructs the time domain signal using the denoised TFM features as the time-frequency feature basis. The method of the present invention can greatly remove the noise in the vibration signal, and while extracting the transient pulse manifold structure of the fault, it can also restore its amplitude, thereby realizing the accurate detection of weak faults of rotating machinery.

[0008] In order to solve the above technical problems, the present invention provides a method for detecting weak faults of rotating machinery based on variable parameter time-frequency manifold signal reconstruction, comprising:

[0009] Step (1), extracting components within the fault band: extracting components within the frequency band where the fault is located from the signal using traditional signal processing methods;

[0010] Step (2), high-dimensional TFD matrix construction: perform variable-parameter time-frequency transformation on the components within the fault band, downsample each obtained TFD and convert it into a column vector for constructing a high-dimensional TFD matrix;

[0011] Step (3), TFM feature extraction: perform manifold learning on the high-dimensional TFD matrix to obtain two-dimensional TFM features;

[0012] Step (4), TFM feature denoising: determine a threshold value according to the joint amplitude distribution of the two-dimensional TFM feature, set the amplitude of the two-dimensional TFM feature below the threshold value to zero, and obtain the denoised two-dimensional TFM feature;

[0013] Step (5), fault component TFD reconstruction: perform weighted summation on the denoised two-dimensional TFM features according to given weights, inversely transform the obtained one-dimensional vector into a two-dimensional matrix, and obtain the reconstructed fault component TFD;

[0014] Step (6), reconstruction of the fault component time domain signal: up-sample the reconstructed fault component TFD to the original size, borrow the phase information corresponding to the component TFD in the fault band, and reconstruct the fault component time domain signal through inverse time-frequency transformation;

[0015] Among them, in step (1), the traditional signal processing method can determine the frequency band position where the fault information is located and extract the signal components therein, and filter out the noise outside the fault frequency band.

[0016] In one of the embodiments, in step (1), the traditional signal processing methods include kurtosis spectrum, sparse spectrum, information spectrum, empirical mode decomposition, variational mode decomposition, empirical wavelet transform, wavelet transform and wavelet packet transform.

[0017] In one of the embodiments, in step (2), the variable parameter time-frequency transform uses different parameters to perform time-frequency transform on the signal, and the variable parameter time-frequency transform method includes short-time Fourier transform, wavelet transform and Wigner-Wiley distribution; the downsampling is a method that can reduce the amount of data, including downsampling, two-dimensional discrete wavelet transform, two-dimensional discrete cosine transform, two-dimensional empirical mode decomposition and two-dimensional variational mode decomposition.

[0018] In one of the embodiments, in step (3), the manifold learning is a method with dimensionality reduction function, and the manifold learning includes a local tangent space arrangement algorithm, an isometric mapping algorithm, a local linear embedding algorithm, a Laplace eigenmapping algorithm and a local preserving projection algorithm.

[0019] In one of the embodiments, in step (4), the joint amplitude distribution of the two-dimensional TFM features refers to taking the first-dimensional TFM features corresponding to each data point as the horizontal coordinate and the second-dimensional TFM features as the vertical coordinate. The graph is in the shape of a check mark, and the points in the lower left part of the check mark are noise points, and the points in the upper right part are fault pulse points. The threshold is the critical point for distinguishing noise points from fault pulse points.

[0020] In one embodiment, in step (5), the given weight can make the amplitude of the reconstructed fault component TFD reach the amplitude level of the fault band component TFD in step (2), and the weight acquisition method includes using the mean of all dimensional data or one dimensional data in the high-dimensional TFD matrix, and performing mathematical operations on the denoised two-dimensional TFM features respectively, and the mathematical operation method includes calculating the Euclidean distance, inner product and cosine similarity; the size of the reconstructed fault component TFD is the same as the size of the fault band component TFD after downsampling.

[0021] In one embodiment, in step (6), the upsampling is the inverse transform of the downsampling in step (2); the original size is the size of the fault band component TFD in step (2); the phase information corresponding to the fault band component TFD is the phase of all time-frequency points obtained by time-frequency transforming the fault band component under one of the parameters in step (2); the inverse time-frequency transform is the inverse transform of the time-frequency transform in step (2); wherein, envelope spectrum analysis is performed on the reconstructed signal, and whether a fault exists is detected by identifying the characteristic frequency of the rotating machinery fault.

[0022] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of any one of the methods are implemented when the processor executes the program.

[0023] A computer-readable storage medium stores a computer program, which implements the steps of any one of the methods when executed by a processor.

[0024] A processor is used to run a program, wherein the program executes any one of the methods when running.

[0025] Beneficial effects of the present invention:

[0026] Compared with the prior art, the present invention discloses a method for detecting weak faults of rotating machinery based on reconstruction of time-frequency manifold signals with variable parameters; the present invention adopts a time-frequency transformation method under different parameters to construct a high-dimensional TFD, which can improve the universality of different signals; downsampling the high-dimensional TFD and then performing manifold learning can reduce the computational burden of manifold learning; adaptively selecting a threshold to denoise the TFM feature can more thoroughly remove residual noise, and is robust to the neighboring point parameters of manifold learning, without the need to optimize the neighboring point parameters, further improving the computational efficiency of the method; the reconstructed signal restores the amplitude of the transient pulse component of the fault, which can be used to quantitatively analyze the severity of the fault. The technical method has at least the following advantages: good universality for different mechanical vibration signals, high computational efficiency, high signal-to-noise ratio, and the ability to quantitatively analyze the severity of the fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The present invention discloses a flow chart of a method for detecting weak faults of rotating machinery based on variable parameter time-frequency manifold signal reconstruction.

[0028] Figure 2 (a) time domain waveform diagram, (b) power spectrum diagram and (c) STFT-based TFD of the vibration signal of a wear fault gearbox provided in an embodiment of the present invention.

[0029] Figure 3 To use VMD from Figure 2 Three variable parameters TFD of the fault band components extracted from the signal.

[0030] Figure 4 To downsample Figure 3 The three variable parameter TFDs after downsampling the longitudinal data of the TFD.

[0031] Figure 5 For Figure 4 The two-dimensional TFD features are obtained after manifold learning of the high-dimensional TFD.

[0032] Figure 6 for Figure 5 Joint amplitude distribution graph of the two-dimensional TFD features.

[0033] Figure 7 For Figure 5 The two-dimensional TFD features are two-dimensional TFD features after threshold denoising.

[0034] Figure 8 To use Figure 7 The denoised two-dimensional TFD feature reconstructs (a) the fault component TFD and (b) the time domain signal and (c) the envelope spectrum. DETAILED DESCRIPTION

[0035] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0036] See also Figure 1 , a method for detecting weak faults in rotating machinery based on variable parameter time-frequency manifold signal reconstruction, the technology specifically includes:

[0037] Step 101: Extracting components within the fault band: using traditional signal processing methods to extract components within the frequency band where the fault is located from the signal.

[0038] Traditional signal processing methods are commonly used signal denoising methods, which can determine the frequency band location of the fault information and extract the signal components therein, and filter out the noise outside the fault frequency band. These methods include but are not limited to Kurtogram, Sparsogram, Infogram, Empirical Mode Decomposition (EMD), Variational Mode Decomposition (VMD), Empirical Wavelet Transform (EWT), Wavelet Transform (WT), Wavelet Packet Transform (WPT), etc.

[0039] Step 102: high-dimensional TFD matrix construction: perform variable-parameter time-frequency transformation on the components within the fault band, downsample each obtained TFD and convert it into a column vector for constructing a high-dimensional TFD matrix.

[0040] Variable parameter time-frequency transform uses different parameters to transform the signal in time-frequency, and the variable parameter time-frequency transform method includes but is not limited to short-time Fourier transform (STFT), wavelet transform (WT), Wigner-Wiley distribution (WVD), etc. Downsampling is a method that can reduce the amount of data, including but not limited to downsampling, two-dimensional discrete wavelet transform (2D-DWT), two-dimensional discrete cosine transform (2D-DCT), two-dimensional empirical mode decomposition (BEMD), two-dimensional variational mode decomposition (BVMD), etc.

[0041] Step 103: TFM feature extraction: perform manifold learning on the high-dimensional TFD matrix to obtain two-dimensional TFM features.

[0042] Manifold learning is a method with dimensionality reduction function, including but not limited to local tangent space arrangement algorithm (LSTA), isometric mapping algorithm (Isomap), local linear embedding algorithm (LLE), Laplace eigenmap algorithm (LE), local preserving projection algorithm (LPP), etc.

[0043] Step 104: TFM feature denoising: determine a threshold value according to the joint amplitude distribution of the two-dimensional TFM feature, set the amplitudes of the two-dimensional TFM feature below the threshold value to zero, and obtain the denoised two-dimensional TFM feature.

[0044] The joint amplitude distribution of the two-dimensional TFM features refers to the first-dimensional TFM features corresponding to each data point as the horizontal axis and the second-dimensional TFM features as the vertical axis. The graph is in the shape of a check mark (√). The points in the lower left part of the check mark are noise points, and the points in the upper right part are fault pulse points. The threshold is the critical point for distinguishing noise points from fault pulse points.

[0045] Step 105: Fault component TFD reconstruction: perform weighted summation on the denoised two-dimensional TFM features according to given weights, inversely transform the obtained one-dimensional vector into a two-dimensional matrix, and obtain the reconstructed fault component TFD.

[0046] The given weight can make the amplitude of the reconstructed fault component TFD reach the amplitude level of the fault band component TFD in step 102. The weight acquisition method includes but is not limited to using the mean of all dimensional data or one dimensional data in the high-dimensional TFD matrix, and performing mathematical operations with the two-dimensional TFM features after denoising. The mathematical operation method includes but is not limited to calculating Euclidean distance, inner product, cosine similarity, etc. The size of the reconstructed fault component TFD is the same as the size of the fault band component TFD after downsampling.

[0047] Step 106: Reconstruction of the time domain signal of the fault component: upsample the reconstructed fault component TFD to the original size, borrow the phase information corresponding to the component TFD in the fault band, and reconstruct the time domain signal of the fault component through inverse time-frequency transformation.

[0048] Upsampling is the inverse transformation of downsampling in step 102. The original size is the size of the fault band component TFD in step 102. The phase information corresponding to the fault band component TFD is the phase of all time-frequency points obtained by time-frequency transformation of the fault band component under one parameter in step 102. Inverse time-frequency transformation is the inverse transformation of the time-frequency transformation in step 102. Envelope spectrum analysis is performed on the reconstructed signal to detect whether there is a fault by identifying the characteristic frequency of the rotating machinery fault.

[0049] A specific application scenario of the present invention is given below:

[0050] In order to more clearly understand the technical solution and effects of the present invention, a detailed description is given below in conjunction with a specific embodiment.

[0051] Take the gearbox wear fault detection as an example. The gearbox is an automotive transmission gearbox with 5 forward gears and 1 reverse gear. The gearbox is engaged in the third forward gear, and its meshing frequency is 500Hz. The rotation frequencies of the tested driving gear and driven gear are 20Hz and 18.5Hz respectively. The accelerometer is installed on the gearbox housing and collects vibration signals with a sampling frequency of 3kHz. The wear fault occurs on the driving gear, so the fault characteristic frequency of the gear is f d =20Hz.

[0052] Reference Figure 2 , Figure 2 It is a time domain waveform diagram, power spectrum diagram and STFT-based TFD of a wear fault gearbox provided by an embodiment of the present invention. The parameters of STFT are window length = 65, window moving step length = 5, and the number of data points for calculating Fourier transform = 128. In the waveform diagram, the transient pulse of the gear fault is submerged in the noise. In the power spectrum diagram, the frequency amplitude is the highest at 500Hz, but it has not been modulated. There is modulation in the range of 200Hz to 360Hz, which is the fault frequency band, but there is noise in the band. In the TFD, periodic fault transient pulse areas are observed in the range of 200Hz to 360Hz, but the noise is also large, and some noise will be mistakenly identified as fault pulses.

[0053] Using the technology disclosed in the present invention Figure 2 The fault characteristic component extraction method is VMD, the variable parameter time-frequency transform method is STFT, and the window length selection range is [50, 80]. 11 values ​​are evenly selected from them to process the fault characteristic components, and 11 TFDs are obtained, among which the TFDs with window lengths of 50, 65 and 80 are as follows Figure 3 As shown. It can be seen that although VMD can remove most of the noise outside the fault frequency band, it cannot remove the noise within the band. In addition, the amplitude of TFD under different window lengths is different, and the time-frequency resolution is also different (reflected in the difference in the shape of the fault pulse area). The longitudinal data of 11 TFDs are downsampled by a ratio of 1:4 using the downsampling method. Figure 3 The TFD shown in the figure is downsampled. Figure 4 As shown, the fault pulse and noise are difficult to distinguish.

[0054] After the 11 downsampled TFDs are converted into column vectors, an 11-dimensional TFD matrix can be constructed. LTSA is used to perform manifold learning, with 15 neighbor points, to obtain a 2-dimensional TFM feature. Each dimensional TFM feature is converted into a time-frequency diagram as shown below: Figure 5 The TFM feature retains the fault pulse, but there is still noise in the band, and the out-of-band noise is added to the second-dimensional TFM feature. The amplitude of the two-dimensional TFM feature is reduced by two orders of magnitude compared to the amplitude of the TFD.

[0055] The joint amplitude distribution of the two-dimensional TFM features is shown in Figure 6 As shown in the figure. The vertical coordinate of the leftmost data point is the threshold of the second-dimensional TFM feature. Use this threshold to draw a horizontal line in the figure. The horizontal coordinate of the second intersection with the check mark is the threshold of the first-dimensional TFM feature. The two-dimensional TFM feature after threshold denoising is shown in the figure. Figure 7 As shown, all the noise is removed and the fault pulse is retained.

[0056] In order to restore its amplitude, the two-dimensional TFM features after denoising are weighted and summed. The weighting coefficients are the mean of the 11-dimensional TFD and the Euclidean distance of the two TFM features. The longitudinal data of the obtained time-frequency diagram is upsampled with a ratio of 4:1 to obtain the fault component TFD, as shown in Figure 8 As shown in (a), its amplitude has recovered to Figure 3 The same order of magnitude as the TFD shown, with the fault pulse retained and all noise removed.

[0057] use Figure 3 The phase information of the TFD in the middle is shown, and the fault component TFD is subjected to short-time inverse Fourier transform to obtain the reconstructed fault component time domain signal, as shown in Figure 8 (b) shows that its amplitude is Figure 2 The time domain waveforms shown are comparable, indicating that the method of the present invention can accurately restore the amplitude of the fault component. Figure 8 (c) is the envelope spectrum of the reconstructed fault signal, which can identify the fault characteristic frequency f d and its second and third harmonics, thereby accurately detecting the slight fault of the third gear driving gear of the gearbox.

[0058] In summary, the universality of different signals can be improved by performing variable parameter time-frequency transformation on the gearbox vibration signal, the computational burden can be reduced by downsampling the high-dimensional TFD, the threshold is determined according to the joint amplitude distribution of the two-dimensional TFM feature, the adaptability of the threshold selection and the robustness of the number of neighboring points for different manifold learning can be improved, and the TFD and time domain signals of the fault components can be reconstructed by weighted summing of the two-dimensional TFM features and performing inverse time-frequency transformation, and the amplitude of the extracted fault components can be restored, thereby effectively detecting the weak fault of the gearbox. The method of the present invention overcomes the problems of the existing TFM technology that the time-frequency transformation parameters need to be manually selected and the parameters of the neighboring points for manifold learning are optimized, the computational efficiency of manifold learning is low, and the TFM features have residual noise and amplitude distortion. It has the advantages of good universality for different mechanical vibration signals, high computational efficiency, high signal-to-noise ratio, and the ability to quantitatively analyze the severity of the fault, and is of great significance for the effective detection of weak faults in rotating machinery.

[0059] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for detecting weak faults in rotating machinery based on variable parameter time-frequency manifold signal reconstruction, characterized in that: include: Step (1), extracting components within the fault band: using traditional signal processing methods to extract components within the frequency band where the fault is located from the signal; Step (2), high-dimensional TFD matrix construction: perform variable-parameter time-frequency transformation on the components within the fault band, downsample each obtained TFD and convert it into a column vector for constructing a high-dimensional TFD matrix; Step (3), TFM feature extraction: perform manifold learning on the high-dimensional TFD matrix to obtain two-dimensional TFM features; Step (4), TFM feature denoising: determine a threshold value based on the joint amplitude distribution of the two-dimensional TFM feature, set the amplitude of the two-dimensional TFM feature below the threshold value to zero, and obtain the denoised two-dimensional TFM feature; Step (5), fault component TFD reconstruction: perform weighted summation on the denoised two-dimensional TFM features according to the given weights, inversely transform the obtained one-dimensional vector into a two-dimensional matrix, and obtain the reconstructed fault component TFD; Step (6), reconstruction of the fault component time domain signal: upsample the reconstructed fault component TFD to the original size, borrow the phase information corresponding to the component TFD in the fault band, and reconstruct the fault component time domain signal through inverse time-frequency transform; Wherein, in step (1), the conventional signal processing method can determine the frequency band location of the fault information and extract the signal components therein, and filter out the noise outside the fault frequency band; In the step (3), the manifold learning is a method with a dimensionality reduction function, and the manifold learning includes a local tangent space arrangement algorithm, an isometric mapping algorithm, a local linear embedding algorithm, a Laplace eigenmapping algorithm, and a local preservation projection algorithm; In step (4), the joint amplitude distribution of the two-dimensional TFM features refers to taking the first-dimensional TFM features corresponding to each data point as the horizontal coordinate and the second-dimensional TFM features as the vertical coordinate. The graph is in the shape of a check mark, and the points in the lower left part of the check mark are noise points, and the points in the upper right part are fault pulse points. The threshold is the critical point for distinguishing noise points from fault pulse points. In step (5), the given weight can make the amplitude of the reconstructed fault component TFD reach the amplitude level of the fault band component TFD in step (2), and the weight acquisition method includes using the mean of all dimensional data or one dimensional data in the high-dimensional TFD matrix, and performing mathematical operations on the two-dimensional TFM features after denoising, respectively, and the mathematical operation method includes calculating the Euclidean distance, the inner product and the cosine similarity; the size of the reconstructed fault component TFD is the same as the size of the fault band component TFD after downsampling; In step (6), the upsampling is the inverse transform of the downsampling in step (2); the original size is the size of the fault band component TFD in step (2); the phase information corresponding to the fault band component TFD is the phase of all time-frequency points obtained by time-frequency transforming the fault band component under one of the parameters in step (2); the inverse time-frequency transform is the inverse transform of the time-frequency transform in step (2); wherein, an envelope spectrum analysis is performed on the reconstructed signal, and whether a fault exists is detected by identifying the characteristic frequency of the rotating machinery fault.

2. The method for detecting weak faults of rotating machinery based on variable parameter time-frequency manifold signal reconstruction according to claim 1, characterized in that: In the step (1), the traditional signal processing methods include kurtosis spectrum, sparse spectrum, information spectrum, empirical mode decomposition, variational mode decomposition, empirical wavelet transform, wavelet transform and wavelet packet transform.

3. The method for detecting weak faults of rotating machinery based on variable parameter time-frequency manifold signal reconstruction according to claim 1, characterized in that: In the step (2), the variable parameter time-frequency transform uses different parameters to perform time-frequency transform on the signal, and the variable parameter time-frequency transform method includes short-time Fourier transform, wavelet transform and Wigner-Wiley distribution; the downsampling is a method that can reduce the amount of data, including downsampling, two-dimensional discrete wavelet transform, two-dimensional discrete cosine transform, two-dimensional empirical mode decomposition and two-dimensional variational mode decomposition.

Images