CN116256158A - Rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation - Google Patents
Rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation Download PDFInfo
- Publication number
- CN116256158A CN116256158A CN202211740653.4A CN202211740653A CN116256158A CN 116256158 A CN116256158 A CN 116256158A CN 202211740653 A CN202211740653 A CN 202211740653A CN 116256158 A CN116256158 A CN 116256158A
- Authority
- CN
- China
- Prior art keywords
- frequency
- time
- signal
- variable
- depth
- 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
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/10—Noise analysis or noise optimisation
Abstract
The invention discloses a rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation, which comprises the following steps: 1) Establishing a depth binary masking signal separation model; 2) Training a depth binary masking signal separation model by utilizing a variable speed signal time-frequency representation and a frequency-conversion fundamental harmonic time-frequency binary mask in a training template signal; 3) Monitoring a current rotating machinery vibration signal, and performing low-pass filtering to obtain a rotating machinery low-pass filtering vibration signal; 4) And according to the instantaneous phase information, angular domain resampling and Fourier transformation are carried out on the rotating machinery vibration signal to obtain an order spectrum, and rotating machinery fault diagnosis is carried out under the working condition of variable rotating speed. The amplitude of the fault order of the rotating machine extracted by the trained depth binary masking signal separation model is obvious, the error compared with the theoretical fault order of the rotating machine is within 1%, the calculation accuracy of the fault order is high, and the fault diagnosis effect is good under the working condition of changing the rotating speed of the rotating machine.
Description
Technical Field
The invention relates to the field of fault diagnosis of rotary machines, in particular to a rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation.
Background
Rotary machines have found widespread use in aeroengines, wind turbines and traction motors, which play a vital role in these heavy equipment. However, the rotating machinery often works in severe environments such as speed change, heavy load, high temperature and the like, and is easy to cause damage to the rotating machinery, so that early fault monitoring and diagnosis of the rotating machinery plays a vital role in intelligent operation and maintenance of important equipment. Methods based on vibration signal monitoring are an important means for state monitoring and fault diagnosis of rotating machinery today. Because the rotating machinery is often in a non-stable working condition with a variable rotating speed, the collected vibration signals under the working condition with the variable rotating speed are easy to generate frequency modulation-amplitude modulation. How to extract the rotating frequency harmonic component from the non-stationary vibration signal rapidly and accurately plays an important role in the extraction of the instantaneous phase of the rotating machinery, and is worthy of intensive study.
The Order Tracking (OT) can convert a non-stationary time domain signal into an angular domain stationary signal, and is widely applied to the field of rotary machine fault diagnosis under a variable rotating speed working condition. The method specifically comprises the steps of converting a time domain non-stationary vibration signal into an angular domain stationary signal, and performing Fourier transformation to obtain an order spectrum for rotary machinery fault diagnosis. The current OT method is mainly divided into two types, one is to install a phase sensor such as a tachometer, a shaft encoder or an optical-based sensor. The other is to extract the Instantaneous Frequency (IF) by using the diagnostic signal itself, then obtain the instantaneous phase curve by integration, and then perform the OT operation by angular domain resampling by using the instantaneous phase information. The rotation speed of the rotating shaft is measured by using a sensor such as a tachometer, and the like, so that the rotating machine has the characteristics of high accuracy, high reliability and the like, but the rotating machine working in a severe environment cannot be provided with the tachometer to measure the rotation speed, for example, an aeroengine is in a high-temperature and heavy-load environment, and too many sensors cannot be arranged. Therefore, the conventional OT technology is severely limited in practical industrial applications.
In light of the above problems, a keyless phase order tracking algorithm (TLOT) has developed. Many methods are improved for use in TLOT, such as generalized demodulation, iterative generalized demodulation, and frequency modulation. However, the above methods are limited in popularization and application, for example, in the process of selecting generalized demodulation and iterative generalized demodulation methods, proper demodulation phase functions need to be determined in advance, but proper phase functions cannot be accurately selected in practical industrial application. In practical industrial application, the rotating machine is in a non-stable working condition, the working environment of the rotating machine cannot be predicted, so that the slope of an IF track cannot be kept consistent, the frequency modulation conversion cannot obtain more accurate time-frequency representation, the harmonic IF of the rotating machine cannot be accurately obtained, and accurate instantaneous phase information cannot be provided for the subsequent OT.
In summary, the TLOT method is affected by the difficulty in adaptive tracking of the time-frequency ridge line and extracting harmonic components, so that the instantaneous phase information of the rotating machinery is difficult to estimate accurately, and the application capability of the traditional TLOT method in fault diagnosis of the rotating machinery under a variable speed condition is restricted.
Disclosure of Invention
The invention aims to provide a rotary mechanical instantaneous phase self-adaptive extraction method based on depth signal separation, which comprises the following steps of:
1) And simulating the fault state of the rotary machine under the variable speed working condition, and acquiring fault signals, so as to generate simulation template signals with different instantaneous frequencies.
Dividing the simulation template signal into a training template signal and a test template signal;
2) Adding Gaussian white noise into the simulation template signal, and performing short-time Fourier transform on the simulation template signal to obtain a time-frequency representation of the variable-speed signal;
3) Establishing a depth binary masking signal separation model, wherein an objective function of the model is an optimal time-frequency binary mask of a frequency conversion basic harmonic component;
4) Training the depth binary masking signal separation model by utilizing the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask in the training template signal until the mean square error of the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask is minimum;
testing the depth binary masking signal separation model by using a test template signal, outputting a nonlinear mapping relation between the time-frequency representation of the frequency conversion basic harmonic component and the time-frequency representation of the speed conversion signal if the test is passed, otherwise, returning to the step 1);
5) Establishing a mean square error loss function between the variable speed signal time-frequency representation and the variable frequency fundamental harmonic time-frequency binary mask according to the variable frequency fundamental harmonic component time-frequency representation and the variable speed signal time-frequency representation nonlinear mapping relation, and storing the network parameters of the trained depth binary mask signal separation model;
6) Monitoring a current rotating machinery vibration signal, and performing low-pass filtering to obtain a rotating machinery low-pass filtering vibration signal;
performing short-time Fourier transform on the rotary mechanical low-pass filtering vibration signal to obtain a vibration signal represented by a time-frequency domain, and inputting the vibration signal represented by the time-frequency domain into a depth binary masking signal separation model to obtain a time-frequency representation of a frequency conversion basic harmonic component;
7) Performing inverse short-time Fourier transform on the time-frequency representation of the frequency-conversion fundamental harmonic component to obtain a frequency-conversion fundamental harmonic component time domain signal, and performing Hilbert transform on the frequency-conversion fundamental harmonic component time domain signal to obtain instantaneous phase information of the frequency-conversion fundamental harmonic component;
8) And (3) carrying out angular domain resampling and Fourier transformation on the rotating machinery vibration signal in the step (6) according to the instantaneous phase information to obtain an order spectrum, and carrying out rotating machinery fault diagnosis under the working condition of variable rotating speed according to the order spectrum.
Further, the test template signal comprises a signal of speed rise under a variable speed condition;
the training template signal comprises a variable speed signal with the instantaneous frequency firstly increased and then decreased, the instantaneous frequency is arbitrarily changed, the instantaneous frequency is firstly increased and then fixed and then decreased, namely:
Wherein f i (u) is the instantaneous frequency; i is the instantaneous frequency order; i=1, 2,3; n (t) is Gaussian white noise; x (t) is a variable speed signal; u is a time variable;
wherein the instantaneous frequency f i (u) is as follows:
f 2 (u)＝64sin(u1.5)+64(3)
where u is a time variable.
Further, in step 1), the test template signal and the training template signal are further expanded by using the truncated simulation template signal.
Further, in step 2), the simulated template signal S (t, f) of the time domain characterization is as follows:
wherein T is a time variable, f is a frequency variable, h (τ -T) is a window function, and T is a time period; x (τ) is the shift signal at τ time.
Further, the objective function BM (t, f) of the depth binary masking signal separation model is as follows:
wherein the time-frequency representation S (t, f 1 ) Time-frequency representation S (t, f) of harmonic components other than the fundamental harmonic of the conversion i ) The following are respectively shown:
wherein t is a time variable, f 1 、f i H (τ -t) is a window function, which is a frequency variable; t is the time period. x (τ) is the shift signal. n is the number of frequency bins.
Further, the depth binary masking signal separation model comprises an input layer, a full connection layer, an activation function layer, a batch normalization layer, a dropout layer and an output layer;
Wherein the full connection layer outputs
The following is shown:
in the method, in the process of the invention,
input for the full connection layer; />
And->
Respectively the weight and the bias of the full connection layer;
the activation function σ (x) of the fully connected layer is as follows:
wherein x is an independent variable;
output of batch normalization layer
The following is shown:
in the method, in the process of the invention,
outputting data for full connection layer>
Is a variance of (2); />
Outputting data for full connection layer>
Is the average value of (2); />
And->
Respectively representing the scale and displacement parameters; />
Is an intermediate parameter;
output of dropout layer
The following are provided:
in the formula, w| p The subset obtained by sampling the output data of the upper layer according to the probability P.
Further, the time-frequency representation S of the frequency conversion fundamental harmonic component optimal (t，f 1 ) The following is shown:
S optimal (t,f 1 )＝argmin(L) (14)
wherein the mean square error L of the loss function is as follows:
in the method, in the process of the invention,
signals output by the model are separated for the depth binary masking signals; n is the number of samples.
Further, the frequency-converted fundamental harmonic component time domain signal x 1 (t) is as follows:
instantaneous phase information of frequency conversion fundamental harmonic component
The following is shown:
wherein H (x) 1 (t) is x 1 The Hilbert transform of (t); unwrap [. Times. ]]Is a phase unwrapping operator.
Further, the step of performing angular domain resampling and fourier transformation on the rotating mechanical vibration signal of step 6) to obtain an order spectrum includes:
6.1 Setting an angular interval for equal angle resampling
Maximum analysis order O max And the total length N of the resampled data, i.e.:
wherein T is the total sampling time, f 1 And (t) is a converted harmonic component.
6.2 Calculating a resampling time scale T in the angular domain 1 The method comprises the following steps:
wherein a, b, c, d are equation coefficients, T 0 Starting time for time domain sampling;
6.3 Acquiring intra-angular resampling time scale T 1 Corresponding vibration amplitude, and interpolating the vibration amplitude by utilizing the Langerhans formula to obtain a time mark T 1 Amplitude in the angular region x (T 1 ) The method comprises the following steps:
x(T 1 )＝x(t i )+(x(t i+1 )-x(t i ))(t i+1 -t i )(T 1 -t i ),t i ≤T 1 ≤t i+1 (21)
wherein t is i 、t i+1 Is a time variable;
6.4 For amplitude x (T) 1 ) Fourier transform is performed to obtain an order spectrum a (k), namely:
where k is the length of the resampled data.
Further, the method for diagnosing the fault of the rotary machine under the working condition of variable rotating speed according to the order spectrum comprises the following steps:
the highest order of the amplitude value in the order spectrum A (k) is positioned, the highest order of the amplitude value is compared with the inherent order of the rotating machine under the working condition of changing the rotation speed, and whether the rotating machine has faults under the working condition of changing the rotation speed is determined.
The technical effect of the invention is undoubtedly that the invention has the following beneficial effects:
1) The depth binary masking signal separation model provided by the invention solves the problems that the traditional keyless phase order tracking method is difficult to adaptively track a time-frequency ridge line and accurately extract harmonic components under the variable rotation speed condition.
2) The depth binary masking signal separation model provided by the invention can adaptively separate the frequency conversion basic harmonic component with definite physical meaning from the acquired time-frequency representation of the rotating mechanical vibration signal.
3) The depth binary masking signal separation model provided by the invention does not need to select the starting point of a certain time-frequency ridge line or determine the harmonic relation like the traditional time-frequency ridge line tracking algorithm, and can directly map the complex time-frequency representation of an original signal into the time-frequency representation with a single fundamental harmonic component, thereby utilizing the fundamental harmonic component to carry out instantaneous phase information estimation.
4) The trained depth binary masking signal separation model provided by the invention has been successfully popularized to self-adaptive separation application of the frequency conversion fundamental harmonic component of the civil aircraft engine vibration signal under the working conditions of rotating machinery fault test bed and variable rotation speed.
5) The amplitude of the fault order of the rotating machine extracted by the trained depth binary masking signal separation model is obvious, the error compared with the theoretical fault order of the rotating machine is within 1%, the calculation accuracy of the fault order is high, and the fault diagnosis effect is good under the working condition of changing the rotating speed of the rotating machine.
6) The invention constructs a depth nonlinear mapping relation between the original multi-harmonic signal and the frequency conversion fundamental harmonic for instantaneous phase estimation. According to the invention, the rotating speed basic harmonic wave of the rotating machinery is adaptively separated by constructing the depth signal separation model, so that the complexity of TLOT calculation is greatly reduced, the frequency ridge line tracking operation and single harmonic decomposition are not needed, and a new thought is provided for accurately estimating the instantaneous phase and providing reliable TLOT rotating speed information.
Drawings
Fig. 1 is a schematic flow chart of a method for adaptively extracting instantaneous phase of a rotary machine based on depth signal separation according to the present invention;
FIGS. 2 (a) -2 (c) are time-frequency representations of the frequency-converted fundamental harmonic components of training template signals constructed in embodiments of the present invention;
FIGS. 3 (a) -3 (c) are respectively a time-frequency representation of a test sample according to an embodiment of the present invention, an optimal time-frequency binary mask obtained by a depth binary mask signal separation model of the test sample, and a frequency conversion fundamental harmonic component adaptively output by the test sample through the depth model;
FIGS. 4 (a) -4 (c) are respectively time-frequency representations of the simulation signals constructed in the embodiment of the present invention, the optimal time-frequency binary mask obtained by the separation model of the constructed simulation signals through the depth binary masking signal, and the frequency conversion fundamental harmonic components of the constructed simulation signals adaptively output through the depth model;
FIGS. 5 (a) -5 (b) are respectively a transient phase information and an order spectrum of a frequency conversion fundamental harmonic, which are obtained by adaptively separating a simulation signal constructed in an embodiment of the present invention through a depth binary masking signal separation model;
FIGS. 6 (a) -6 (b) are frequency conversion fundamental harmonic components obtained by time-frequency representation after low-pass filtering and adaptive separation of a depth binary masking signal separation model of vibration signals of an accessory gearbox of an aeroengine, which are acquired in an embodiment of the invention;
7 (a) -7 (b) are transient phase information and order spectrums of a frequency conversion basic harmonic component obtained by a depth binary masking signal separation model of a vibration signal of an accessory gearbox of an aeroengine, which is acquired in the embodiment of the invention;
fig. 8 (a) -8 (b) are frequency conversion fundamental harmonic components obtained by time-frequency characterization after low-pass filtering and adaptive separation of a depth binary masking signal separation model of a vibration signal of a rolling bearing test bed acquired in an embodiment of the invention;
fig. 9 (a) -9 (b) are transient phase information and order spectra of a harmonic component of a frequency conversion base obtained by a depth binary masking signal separation model of a vibration signal of a rolling bearing test bed acquired in the embodiment of the invention.
Detailed Description
The present invention is further described below with reference to examples, but it should not be construed that the scope of the above subject matter of the present invention is limited to the following examples. Various substitutions and alterations are made according to the ordinary skill and familiar means of the art without departing from the technical spirit of the invention, and all such substitutions and alterations are intended to be included in the scope of the invention.
Example 1:
referring to fig. 1 to 9, a method for adaptively extracting instantaneous phase of a rotary machine based on depth signal separation includes the steps of:
1) Simulating fault signals of the rotary machine under the working condition of variable rotation speed, so as to generate simulated template signals with different instantaneous frequencies;
dividing the simulation template signal into a training template signal and a test template signal;
2) Adding Gaussian white noise into the simulation template signal, and performing short-time Fourier transform on the simulation template signal to obtain a time-frequency representation of the variable-speed signal;
3) Establishing a depth binary masking signal separation model, wherein an objective function of the model is an optimal time-frequency binary mask of a frequency conversion basic harmonic component;
4) Training the depth binary masking signal separation model by utilizing the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask in the training template signal until the mean square error of the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask is minimum;
testing the depth binary masking signal separation model by using the variable speed signal time-frequency representation and the variable frequency fundamental harmonic time-frequency binary mask in the test template signal, outputting the nonlinear mapping relation of the variable speed signal time-frequency representation and the variable frequency fundamental harmonic component time-frequency representation if the test passes, otherwise returning to the step 1);
The standard for passing the test is as follows: the mean square error of the time-frequency representation of the current variable rotation speed signal and the time-frequency binary mask of the frequency conversion fundamental harmonic is minimum.
5) Establishing a mean square error loss function between the variable speed signal time-frequency representation and the variable frequency fundamental harmonic time-frequency binary mask according to the variable frequency fundamental harmonic component time-frequency representation and the variable speed signal time-frequency representation nonlinear mapping relation, and storing the network parameters of the trained depth binary mask signal separation model;
6) Monitoring a current rotating machinery vibration signal, and performing low-pass filtering to obtain a rotating machinery low-pass filtering vibration signal;
performing short-time Fourier transform on the rotary mechanical low-pass filtering vibration signal to obtain a vibration signal represented by a time-frequency domain, and inputting the vibration signal represented by the time-frequency domain into a depth binary masking signal separation model to obtain a time-frequency representation of a frequency conversion basic harmonic component;
7) Performing inverse short-time Fourier transform on the time-frequency representation of the frequency-conversion fundamental harmonic component to obtain a frequency-conversion fundamental harmonic component time domain signal, and performing Hilbert transform on the frequency-conversion fundamental harmonic component time domain signal to obtain instantaneous phase information of the frequency-conversion fundamental harmonic component;
8) And (3) carrying out angular domain resampling and Fourier transformation on the rotating machinery vibration signal in the step (6) according to the instantaneous phase information to obtain an order spectrum, and carrying out rotating machinery fault diagnosis under the working condition of variable rotating speed according to the order spectrum.
The test template signal comprises a signal of speed rise under a variable speed condition;
the training template signal comprises a variable speed signal with the instantaneous frequency firstly increased and then decreased, the instantaneous frequency is arbitrarily changed, the instantaneous frequency is firstly increased and then fixed and then decreased, namely:
wherein f i (u) is the instantaneous frequency; i is the instantaneous frequency order; i=1, 2,3; n (t) is Gaussian white noise; x (t) is a variable speed signal;
wherein the instantaneous frequency f i (u) is as follows:
f 2 (u)＝64sin(u1.5)+64(3)
wherein u is.
In step 1), the test template signal and the training template signal are further expanded by using the truncated simulation template signal.
In step 2), the simulated template signal S (t, f) of the time domain characterization is as follows:
wherein t is a time variable, f is a frequency variable, and h (τ) is a window function; t is the time period.
The objective function BM (t, f) of the depth binary masking signal separation model is as follows:
wherein the time-frequency representation S (t, f 1 ) Time-frequency representation S (t, f) of harmonic components other than the fundamental harmonic of the conversion i ) The following are respectively shown:
wherein t is a time variable, f is a frequency variable, and h (τ) is a window function; t is the time period. x (τ) is the shift signal.
The depth binary masking signal separation model comprises an input layer, a full connection layer, an activation function layer, a batch normalization layer, a dropout layer and an output layer;
Wherein the full connection layer outputs
The following is shown:
in the method, in the process of the invention,
input for the full connection layer; />
And->
Respectively the weight and the bias of the full connection layer;
the activation function σ (x) of the fully connected layer is as follows:
output of batch normalization layer
The following is shown:
in the method, in the process of the invention,
outputting data for full connection layer>
Is a variance of (2); />
Outputting data for full connection layer>
Is the average value of (2); />
And->
Respectively representing the scale and displacement parameters; />
Is an intermediate parameter;
output of dropout layer
The following are provided:
in the formula, w| p The subset obtained by sampling the output data of the upper layer according to the probability P.
Time-frequency representation S of frequency conversion fundamental harmonic component optimal (t，f 1 ) The following is shown:
S optimal (t,f 1 )＝argmin(L) (14)
wherein the mean square error L of the loss function is as follows:
in the method, in the process of the invention,
signals output by the model are separated for depth binary masking signals.
Frequency conversion fundamental harmonic component time domain signal x 1 (t) is as follows:
instantaneous phase information of frequency conversion fundamental harmonic component
The following is shown:
wherein H (x) 1 (t) is x 1 The Hilbert transform of (t); unwrap [. Times. ]]Is a phase unwrapping operator.
The step of performing angular domain resampling and fourier transformation on the rotating mechanical vibration signal of step 6) to obtain an order spectrum comprises the steps of:
6.1 Setting an angular interval for equal angle resampling
Maximum analysis order O max And the total length N of the resampled data, i.e.:
wherein T is the total sampling time, f 1 And (t) is a converted harmonic component.
6.2 Calculating a resampling time scale T in the angular domain 1 The method comprises the following steps:
wherein a, b, c, d are equation coefficients, T 0 Starting time for time domain sampling;
6.3 Acquiring intra-angular resampling time scale T 1 Corresponding vibration amplitude, and interpolating the vibration amplitude by utilizing the Langerhans formula to obtain a time mark T 1 Amplitude in the angular region x (T 1 ) The method comprises the following steps:
x(T 1 )＝x(t i )+(x(t i+1 )-x(t i ))(t i+1 -t i )(T 1 -t i ),t i ≤T 1 ≤t i+1 (21)
6.4 For amplitude x (T) 1 ) Fourier transform is performed to obtain an order spectrum a (k), namely:
where k is the length of the resampled data.
The method for diagnosing the fault of the rotary machine under the working condition of variable rotating speed according to the order spectrum comprises the following steps:
the highest order of the amplitude value in the order spectrum A (k) is positioned, the highest order of the amplitude value is compared with the inherent order of the rotating machine under the working condition of changing the rotation speed, and whether the rotating machine has faults under the working condition of changing the rotation speed is determined.
Example 2:
referring to fig. 1 to 9, a method for adaptively extracting instantaneous phase of a rotary machine based on depth signal separation includes the steps of:
1) A series of simulation template signals with different instantaneous frequencies are generated by simulating fault signals of the rotary machine under the working condition of variable rotation speed. The training samples come from constructed variable-speed signals with slopes changing randomly, and the test samples come from simulation signals under different variable-speed conditions;
1.1 The training template signal is mainly composed of a series of variable speed signals with the instantaneous frequency increasing and then decreasing, any change and the instantaneous frequency increasing and then fixing and then decreasing. The specific expression of the training template signal can be expressed as:
wherein f i (u) is the instantaneous frequency, i is the instantaneous frequency order, n (t) is the white gaussian noise, n (t) = -3dB. The sampling frequency of the training template signal was 1.0417kHz. The duration of each training template signal is 10s. The following is a specific expression of the instantaneous frequency:
f 2 (u)＝64sin(u1.5)+64(3)
1.2 By arbitrarily changing f 1 (u)、f 2 (u) and f 3 (u) slope obtaining constructed training samples, wherein f 1 (u)、f 2 (u) and f 3 (u) changing the slope to obtain 11, 10 and 11 training template signals, respectively. Thus, the total duration of the training template signal is 320 seconds. The test specimen was constructed as a signal of a speed rise under variable speed conditions, and the total duration of the test specimen was 10s. In order to obtain more training samples and test samples, the invention adopts truncated simulation template signals to test samples and training samplesThe present invention extends. The signals of the constructed training samples and the test samples are then truncated. The length of each section of the training sample after cutting is 20 sample points, and the height of each section is 256 sample points. Meanwhile, 10 sample points are moved each time to conduct training sample truncation. The length of each section of the test sample after cutting is 20 sample points, and the height of each section is 256 sample points. Meanwhile, the test sample is cut off by moving 20 sample points each time. Thus, each segment of the test sample after truncation does not overlap. Then, the training samples obtained by the invention are 33,317, and the test samples are 514.
2) Adding Gaussian white noise into the constructed simulation template signal, and transforming the simulation template signal to a time-frequency domain by utilizing short-time Fourier transformation for further analysis;
2.1 Performing short-time fourier transform on the constructed simulation signal, wherein the obtained time-frequency characterization is as follows:
where t is a time variable, f is a frequency variable, and h (τ) is a window function.
3) And obtaining a training target of the depth binary masking signal separation model through time-frequency characterization combination time-frequency binary masking operation of fundamental harmonic waves and higher order harmonic waves in the constructed simulation template signals. The training target of the depth model is a time-frequency binary mask with optimal frequency conversion basic harmonic components;
3.1 The depth binary masking signal separation model constructed by the invention is mainly aimed at the frequency conversion fundamental harmonic f 1 Adaptive separation of (t). The binary mask is derived from time-frequency characterization of the converted fundamental harmonic component and other harmonic components. Thus, the time-frequency characterization of the converted fundamental harmonic component with other harmonic components is:
wherein S (t, f 1 ) For time-frequency characterization of the fundamental harmonic component of the conversion frequency, S (t, f i ) Time-frequency characterization of other harmonic components.
3.2 According to the time-frequency representation of the above-mentioned frequency conversion basic harmonic component, can obtain the binary mask objective function of the binary mask signal separation model of the depth as:
4) Training a depth binary masking signal separation model by utilizing a variable speed signal time-frequency representation and a frequency conversion fundamental harmonic time-frequency binary mask training target in the constructed training template signal, and obtaining a nonlinear mapping relation between a frequency conversion fundamental harmonic component time-frequency representation and an original signal time-frequency representation when the time-frequency representation and the frequency conversion fundamental harmonic time-frequency binary mask of the original signal reach the minimum mean square error in the depth binary masking signal separation model;
4.1 The depth binary masking signal separation neural network consists of an input layer, a full connection layer, an activation function layer, a batch normalization layer, a dropout layer and an output layer.
4.2 Calculation of the full connection layer is as follows:
in the method, in the process of the invention,
for the output of the full connection layer, < >>
Input for full connection layer, < >>
And->
The weights and offsets of the full connection layer, respectively.
4.3 Through full-connected layer convolution operations, features in the original signal will appear in a nonlinear form. The calculation formula of the activation function is as follows:
4.4 During the training of the depth model, there is a significant computational load and the data input distribution at each layer is shifted by covariates. In order to inhibit this phenomenon, the training process is accelerated, and a data batch normalization layer is performed after each full-connection operation. The formula of the batch normalization layer is as follows:
In the method, in the process of the invention,
outputting data for full connection layer>
Variance of->
Outputting data for full connection layer>
Mean value of->
And->
Representing the scale and displacement parameters, respectively.
4.5 To prevent the proposed depth signal separation model from over-fitting during training, a dropout layer is introduced to prevent this. The specific operation is that in the training process, the exit layer randomly samples the parameters of the full connection layer according to a certain probability P, and the sampled network is used as an updated target network. The calculation method of the dropout layer is as follows:
wherein W is a subset obtained by sampling the output data of the upper layer according to a certain probability P.
p
5) Establishing a mean square error loss function between the time-frequency representation of the original signal and a frequency conversion fundamental harmonic time-frequency binary mask, and storing network parameters of a trained depth binary mask signal separation model;
5.1 The objective function of the depth binary masking signal separation model is to minimize the mean square error between the input signal time-frequency representation and the target time-frequency binary mask. And when the minimum value is reached, outputting the time-frequency representation of the frequency conversion fundamental harmonic wave which needs to be separated. Thus, the mean square error equation for calculating the loss function is:
in the method, in the process of the invention,
signals output by the model are separated for depth binary masking signals.
5.2 Therefore, the time-frequency characterization of the final output frequency conversion fundamental harmonic of the depth binary masking signal separation model provided by the invention is as follows:
S optimal (t,f 1 )＝argmin(L)(15)
6) In the application process of the depth binary masking signal separation model, as the high-frequency resonance harmonic components of the rotary machine are quite rich and complex and the frequency conversion harmonic is mainly distributed in the low frequency band, the invention only carries out the depth separation operation on the low frequency band of the signal. Firstly, carrying out low-pass filtering on an acquired vibration signal, limiting a frequency conversion harmonic component in a narrow band for analysis, then converting the rotating machinery low-pass filtering vibration signal into a time-frequency domain for analysis by utilizing short-time Fourier transform, and then inputting a trained depth binary masking signal separation model to obtain a time-frequency representation of the frequency conversion fundamental harmonic component in a self-adaptive manner;
7) And transforming the time-frequency representation of the frequency-converted basic harmonic component separated by the depth model into a time domain signal by using an inverse short-time Fourier transform, and obtaining the instantaneous phase information of the frequency-converted basic harmonic component separated by using a Hilbert transform. And then, carrying out angular domain resampling on the acquired original vibration signal according to the instantaneous phase information, and obtaining an order spectrum through Fourier transformation, thereby being used for diagnosing the fault of the rotary machine under the working condition of variable rotating speed.
7.1 And (3) performing short-time Fourier inverse transformation by utilizing the time-frequency representation of the frequency conversion basic harmonic component separated by the depth binary masking signal separation model to obtain a time domain signal of the frequency conversion basic harmonic component. The calculation formula is as follows:
7.2 A time domain signal x for obtaining a frequency-converted fundamental harmonic component 1 After (t), it is an important step in order tracking to obtain instantaneous phase information of the converted fundamental harmonic component. X is of 1 The instantaneous phase information of (t) can be calculated as follows:
wherein H (x) 1 (t) is x 1 Hilbert transform of (t), unwrap [ + ]]Is a phase unwrapping operator that can unwrap x 1 The instantaneous phase information of (t) is from interval [ -pi, pi]Mapped to the cumulative phase space. The essence of the order tracking method is the conversion between the time and angular domains of the original signal. The method is a process of performing incremental angle analog sampling on an original signal according to phase information. The main steps of the conversion process are as follows:
7.2.1 Angular interval of equal angle resampling
Maximum analysis order O max And the length N of the resampled data is shown in equation (17). The equiangular sampling rate should be greater than or equal to twice the maximum analysis order according to the sampling theorem.
Wherein T is the total sampling time, f 1 And (t) is a converted harmonic component.
7.2.2 Calculating a resampling time scale T in the angular domain 1 。
Wherein a, b, c, d are equation coefficients, T 0 Is the time domain sample start time.
7.2.3 Obtaining an intra-angular resampling time scale T 1 Corresponding vibration amplitudes. Interpolation is carried out on the signals by utilizing the Langerhans formula to obtain corresponding time marks T 1 The amplitude in the angular domain is shown below:
x(T 1 )＝x(t i )+(x(t i+1 )-x(t i ))(t i+1 -t i )(T 1 -t i ),t i ≤T 1 ≤t i+1 (21)
7.2.4) Peer-to-peer angle resampled signal x (T 1 ) Performing Fourier transform to obtain an order spectrum, wherein the expression is as follows:
8) Converting the time domain non-stationary signal into the angular domain stationary signal by using the angular domain resampling method, and combining the order tracking method with the angular domain stationary signal x (T) 1 ) Obtaining an order spectrum A (k), positioning the highest order in the order spectrum A (k), comparing the obtained order with the inherent order of the rotating machine under the variable-rotation-speed working condition, and determining whether the rotating machine has faults under the variable-rotation-speed working condition.
Example 3:
as shown in fig. 2-5, a method for adaptively extracting instantaneous phase of a rotary machine based on depth signal separation, the specific implementation process and result of the method are as follows:
step 1: a series of simulation template signals with different instantaneous frequencies are generated by simulating fault signals of the rotary machine under the working condition of variable rotation speed. The training samples come from constructed variable-speed signals with slopes changing randomly, and the test samples come from simulation signals under different variable-speed conditions;
Step 1.1: the training template signal is mainly composed of a series of variable speed signals with the instantaneous frequency increasing and then decreasing, any change and the instantaneous frequency increasing and then fixing and then decreasing. The specific expression of the training template signal can be expressed as:
wherein f i (u) is the instantaneous frequency, i is the instantaneous frequency order, n (t) is the white gaussian noise, n (t) = -3dB. The sampling frequency of the training template signal was 1.0417kHz. The duration of each training template signal is 10s, and the constructed training template signal frequency conversion fundamental harmonic time-frequency characterization is shown in fig. 2. The following is a specific expression of the instantaneous frequency:
f 2 (u)＝64sin(u1.5)+64(25)
reconstructing another variable-rotation-speed simulation signal x inconsistent with training samples 2 And (t) performing performance generalization verification of the depth binary masking signal separation model, wherein the fault order is set to be 2.6. The formula is as follows:
wherein f 4 (u) is an instantaneous frequency, the expression of which is as follows:
f 4 (u)＝-1.28u 2 +18u+65(28)
step 1.2: by arbitrarily changing f 1 (u)、f 2 (u) and f 3 (u) slope obtaining constructed training samples, wherein f 1 (u)、f 2 (u) and f 3 (u) changing the slope to obtain 11, 10 and 11 training template signals, respectively. Thus, the total duration of the training template signal is 320 seconds. The test specimen was constructed as a signal of a speed rise under variable speed conditions, and the total duration of the test specimen was 10s. In order to obtain more training samples and test samples, the invention adopts truncated simulation template signals to expand the training samples and the test samples. The signals of the constructed training samples and the test samples are then truncated. The length of each section of the training sample after cutting is 20 sample points, and the height of each section is 256 sample points. Meanwhile, 10 sample points are moved each time to conduct training sample truncation. The length of each section of the test sample after cutting is 20 sample points, and the height of each section is 256 sample points. Meanwhile, the test sample is cut off by moving 20 sample points each time. Thus, each section of the test sample after interception is absent There is an overlap. Then, the training samples obtained by the invention are 33,317, and the test samples are 514. The constructed test samples and the simulated signal time-frequency characterization for generalization performance verification are shown in fig. 3 (a) and fig. 4 (a).
Step 2: adding Gaussian white noise into the constructed simulation template signal, and transforming the simulation template signal to a time-frequency domain by utilizing short-time Fourier transformation for further analysis;
step 2.1: and carrying out short-time Fourier transform on the constructed simulation signal, wherein the obtained time-frequency characterization is as follows:
where t is a time variable, f is a frequency variable, and h (τ) is a window function.
Step 3: and obtaining a training target of the depth binary masking signal separation model through time-frequency characterization combination time-frequency binary masking operation of fundamental harmonic waves and higher order harmonic waves in the constructed simulation template signals. The training target of the depth model is a time-frequency binary mask with optimal frequency conversion basic harmonic components;
step 3.1: the depth binary masking signal separation model constructed by the invention is mainly aimed at the frequency conversion fundamental harmonic f 1 Adaptive separation of (t). The binary mask is derived from time-frequency characterization of the converted fundamental harmonic component and other harmonic components. Thus, the time-frequency characterization of the converted fundamental harmonic component with other harmonic components is:
Wherein S (t, f 1 ) For time-frequency characterization of the fundamental harmonic component of the conversion frequency, S (t, f i ) Time-frequency characterization of other harmonic components. Depth binary masking in this embodimentAfter the signal separation model is trained by the training sample, the test sample and the simulation signal for generalization performance verification are utilized to obtain an optimal time-frequency binary mask diagram, which is shown in fig. 3 (b) and fig. 4 (b).
Step 3.2: according to the time-frequency representation of the frequency conversion fundamental harmonic component, a binary mask objective function of a depth binary mask signal separation model can be obtained as follows:
step 4: training a depth binary masking signal separation model by utilizing a variable speed signal time-frequency representation and a frequency conversion fundamental harmonic time-frequency binary mask training target in the constructed training template signal, and obtaining a nonlinear mapping relation between a frequency conversion fundamental harmonic component time-frequency representation and an original signal time-frequency representation when the time-frequency representation and the frequency conversion fundamental harmonic time-frequency binary mask of the original signal reach the minimum mean square error in the depth binary masking signal separation model;
step 4.1: the depth binary masking signal separation neural network consists of an input layer, a full connection layer, an activation function layer, a batch normalization layer, a dropout layer and an output layer.
Step 4.2: the calculation of the full connection layer is as follows:
in the method, in the process of the invention,
for the output of the full connection layer, < >>
Input for full connection layer, < >>
And->
The weights and offsets of the full connection layer, respectively.
Step 4.3: by full connected layer convolution operation, the features in the original signal will appear in a nonlinear form. The calculation formula of the activation function is as follows:
step 4.4: during the training of the depth model, there is a great computational load, and the data input distribution of each layer is shifted in covariates. In order to inhibit this phenomenon, the training process is accelerated, and a data batch normalization layer is performed after each full-connection operation. The formula of the batch normalization layer is as follows:
in the method, in the process of the invention,
outputting data for full connection layer>
Variance of->
Outputting data for full connection layer>
Mean value of->
And->
Representing the scale and displacement parameters, respectively.
Step 4.5: in order to prevent the proposed depth signal separation model from over fitting during training, a dropout layer is introduced to prevent this. The specific operation is that in the training process, the exit layer randomly samples the parameters of the full connection layer according to a certain probability P, and the sampled network is used as an updated target network. The calculation method of the dropout layer is as follows:
Wherein W is a subset obtained by sampling the output data of the upper layer according to a certain probability P.
p
Step 5: establishing a mean square error loss function between the time-frequency representation of the original signal and a frequency conversion fundamental harmonic time-frequency binary mask, and storing network parameters of a trained depth binary mask signal separation model;
step 5.1: the objective function of the depth binary masking signal separation model is to minimize the mean square error between the input signal time-frequency representation and the target time-frequency binary mask. And when the minimum value is reached, outputting the time-frequency representation of the frequency conversion fundamental harmonic wave which needs to be separated. Thus, the mean square error equation for calculating the loss function is:
in the method, in the process of the invention,
signals output by the model are separated for depth binary masking signals.
Step 5.2: therefore, the time-frequency characterization of the final output frequency conversion fundamental harmonic of the depth binary masking signal separation model provided by the invention is as follows:
S optimal (t,f 1 )＝argmin(L)(39)
the test samples and the simulation signals constructed for generalization performance verification are adaptively separated by a depth binary masking signal separation model to obtain frequency conversion fundamental harmonic components as shown in fig. 3 (c) and fig. 4 (c).
Step 6: in the application process of the depth binary masking signal separation model, as the high-frequency resonance harmonic components of the rotary machine are quite rich and complex and the frequency conversion harmonic is mainly distributed in the low frequency band, the invention only carries out the depth separation operation on the low frequency band of the signal. Firstly, carrying out low-pass filtering on an acquired vibration signal, limiting a frequency conversion harmonic component in a narrow band for analysis, then converting the rotating machinery low-pass filtering vibration signal into a time-frequency domain for analysis by utilizing short-time Fourier transform, and then inputting a trained depth binary masking signal separation model to obtain a time-frequency representation of the frequency conversion fundamental harmonic component in a self-adaptive manner;
Step 7: and transforming the time-frequency representation of the frequency-converted basic harmonic component separated by the depth model into a time domain signal by using an inverse short-time Fourier transform, and obtaining the instantaneous phase information of the frequency-converted basic harmonic component separated by using a Hilbert transform. And then, carrying out angular domain resampling on the acquired original vibration signal according to the instantaneous phase information, and obtaining an order spectrum through Fourier transformation, thereby being used for diagnosing the fault of the rotary machine under the working condition of variable rotating speed.
Step 7.1: and performing short-time inverse Fourier transform by utilizing the time-frequency representation of the frequency conversion basic harmonic component separated by the depth binary masking signal separation model to obtain a time domain signal of the frequency conversion basic harmonic component. The calculation formula is as follows:
step 7.2: obtaining a time domain signal x of a frequency conversion fundamental harmonic component 1 After (t), it is an important step in order tracking to obtain instantaneous phase information of the converted fundamental harmonic component. X is of 1 The instantaneous phase information of (t) can be calculated as follows:
wherein H (x) 1 (t) is x 1 Hilbert transform of (t), unwrap [ + ]]Is a phase unwrapping operator that can unwrap x 1 The instantaneous phase information of (t) is from interval [ -pi, pi]Mapped to the cumulative phase space. The essence of the order tracking method is the conversion between the time and angular domains of the original signal. The method is a process of performing incremental angle analog sampling on an original signal according to phase information. The main steps of the conversion process are as follows:
Step 7.2.1: angular interval of equal angle resampling
Maximum analysis order O max And the length N of the resampled data is shown in equation (17). The equiangular sampling rate should be greater than or equal to twice the maximum analysis order according to the sampling theorem.
Wherein T is the total sampling time, f 1 And (t) is a converted harmonic component.
Step 7.2.2: calculating a resampling time scale T in the angular domain 1 。
Wherein a, b, c, d are equation coefficients, T 0 Is the time domain sample start time.
Step 7.2.3: obtaining the intra-angular resampling time scale T 1 Corresponding vibration amplitudes. Interpolation is carried out on the signals by utilizing the Langerhans formula to obtain corresponding time marks T 1 The amplitude in the angular region is shown below:
x(T 1 )＝x(t i )+(x(t i+1 )-x(t i ))(t i+1 -t i )(T 1 -t i ),t i ≤T 1 ≤t i+1 (45)
Step 7.2.4: peer-to-peer angle resampled signal x (T 1 ) Performing Fourier transform to obtain an order spectrum, wherein the expression is as follows:
step 8: converting the time domain non-stationary signal into the angular domain stationary signal by using the angular domain resampling method, and combining the order tracking method with the angular domain stationary signal x (T) 1 ) Obtaining an order spectrum A (k), positioning the highest order in the order spectrum A (k), comparing the obtained order with the inherent order of the rotating machine under the variable-rotation-speed working condition, and determining whether the rotating machine has faults under the variable-rotation-speed working condition.
And constructing a frequency conversion fundamental harmonic time-frequency representation obtained by separating a simulation signal for verifying generalization performance through a depth binary masking signal separation model, and then transforming the frequency conversion fundamental harmonic time-frequency representation by utilizing short-time Fourier transform to obtain a time domain signal. Meanwhile, the transient phase information is obtained by Hilbert transformation on the time domain signals, as shown in fig. 5 (a), the transient phase information is well matched with a real phase curve, which shows that the proposed model has higher accuracy. And resampling the original non-stationary signal in the time domain by utilizing the instantaneous phase information to obtain a stationary signal in the angular domain. Then, the order spectrum of the angle domain stationary signal is obtained by fourier transformation, and the calculated order spectrum is shown in fig. 5 (b), so that the 2.6 failure order and the high-order failure order can be accurately identified.
Example 4:
as shown in fig. 6-7, a method for adaptively extracting instantaneous phase of a rotary machine based on depth signal separation is implemented as follows:
the invention applies a depth binary masking signal separation model trained by a simulation template signal in the early stage to the adaptive separation application of a frequency conversion base harmonic wave of an accessory gearbox of an aeroengine, and selects an L5 high-speed shaft as a rotating speed reference shaft for keyless phase order tracking analysis, wherein the fault order of an outer ring of a rolling bearing of the aeroengine is 7.759. The sampling frequency is 4.41kHz, and the sampling duration is 204s. Firstly, the acquired original vibration signal Acc1 of the accessory gearbox of the aeroengine is subjected to 5-layer wavelet low-pass filtering, and the sampling frequency is changed to 689Hz. The time-frequency characterization of the original vibration signal Acc1 after low-pass filtering is shown in fig. 6 (a). The low-pass filtered vibration signal is adaptively separated through a depth binary masking signal separation model, and the time-frequency representation of the frequency conversion fundamental harmonic component is shown in fig. 6 (b).
As shown in fig. 6 (b), the depth binary masking signal separation model can effectively separate the frequency conversion fundamental harmonic component of the acquired aeroengine Acc1 vibration signal. Then, the time-frequency representation of the separated frequency-conversion fundamental harmonic component is converted into a time domain signal through short-time Fourier inversion, and then the Hilbert transform is utilized to extract instantaneous phase information, as shown in fig. 7 (a). Angular domain resampling of the Acc2 vibration signal combined with fourier transform yields an order spectrum, as shown in fig. 7 (b). And then an advanced keyless phase order tracking algorithm is selected to be compared with the depth binary masking signal separation model. The specific comparison method comprises MOPA, FKSR and SRCKF, and then analyzing the obtained fault order relative error to evaluate the superiority of the depth binary masking signal separation model. The comparison results are shown in Table 1.
Table 1 comparative results method
As shown in fig. 7 (b), the third order of failure obtained by the depth binary masking signal separation model is 23.25. The third order of failure has no other more obvious order components around it, so it can be explained that the order extracted by the method of the invention is the order of failure of the rolling bearing of the aeroengine related to L5. Compared with the fault orders of the outer ring of the rolling bearing of the aeroengine obtained by different methods in the table 1, the relative error obtained by the depth binary masking signal separation model is 0.116%, is the minimum relative error in the existing algorithm, and can reflect the performance of the method provided by the invention under the real variable-speed working condition. The superiority of the fault diagnosis of the rolling bearing of the accessory gearbox of the aeroengine under the condition of variable rotation speed is verified.
Example 5:
as shown in fig. 8-9, a method for adaptively extracting instantaneous phase of a rotary machine based on depth signal separation is shown as follows:
the depth binary masking signal separation model trained by the simulation template signal in the early stage is applied to the rolling bearing test bed frequency conversion basic harmonic wave self-adaptive separation application, and the fault order of the rolling bearing inner ring is 4.92. The sampling frequency is 12.8kHz, and the sampling time is 36.2s. Firstly, the collected original vibration signal of the rolling bearing is subjected to 7-layer wavelet low-pass filtering. The local time-frequency characterization of the original vibration signal is shown in fig. 8 (a). The original vibration signal is adaptively separated through a depth binary masking signal separation model, and the time-frequency characterization of the frequency conversion fundamental harmonic component is shown in fig. 8 (b).
As shown in fig. 8 (b), the depth binary masking signal separation model can effectively separate the frequency-converted fundamental harmonic components of the rolling bearing vibration signal. Then, the time-frequency representation of the separated frequency-converted fundamental harmonic component is converted into a time domain signal through short-time Fourier inversion, and then the Hilbert transform is utilized to extract instantaneous phase information, as shown in fig. 9 (a). Angular domain resampling of the rolling bearing vibration signal in combination with fourier transform yields an order spectrum, as shown in fig. 9 (b). And then an advanced keyless phase order tracking algorithm is selected to be compared with the depth binary masking signal separation model, and the comparison result is shown in table 2.
Table 2 method for comparing results
As shown in fig. 9 (b), the depth binary masking signal separation model yields a failure order of 4.92006. According to the comparison of the fault orders of the inner ring of the rolling bearing obtained by different methods in the table 2, the relative error obtained by the depth binary masking signal separation model is 0.0012%, is the minimum error in the existing algorithm, and can reflect the fault diagnosis accuracy of the algorithm under the variable rotation speed working condition.
Through the application of the fault diagnosis cases under the two variable speed working conditions, the effectiveness of the depth binary masking signal separation model provided by the invention is verified, and the model can adaptively separate the frequency conversion fundamental harmonic component of the vibration signal, so that the instantaneous phase of the frequency conversion fundamental harmonic component is accurately estimated. The result shows that the depth binary masking signal separation model provided by the invention can provide a powerful tool for self-adaptive diagnosis for fault diagnosis application under the working condition of rotating speed change of the rotary machine. The effectiveness and the universality of the method provided by the invention are verified through the application of vibration signals of a gearbox of an accessory of a civil aircraft engine and vibration signals of a rolling bearing test bed. In conclusion, the depth binary masking signal separation model provides a novel thought and a powerful tool for keyless phase order tracking in a rotating machine speed change state, and has good application prospect.
Example 6:
a rotary mechanical instantaneous phase self-adaptive extraction method based on depth signal separation comprises the following steps:
1) And simulating the fault state of the rotary machine under the variable speed working condition, and acquiring fault signals, so as to generate simulation template signals with different instantaneous frequencies.
Dividing the simulation template signal into a training template signal and a test template signal;
2) Adding Gaussian white noise into the simulation template signal, and performing short-time Fourier transform on the simulation template signal to obtain a time-frequency representation of the variable-speed signal;
3) Establishing a depth binary masking signal separation model, wherein an objective function of the model is an optimal time-frequency binary mask of a frequency conversion basic harmonic component;
4) Training the depth binary masking signal separation model by utilizing the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask in the training template signal until the mean square error of the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask is minimum;
testing the depth binary masking signal separation model by using a test template signal, outputting a nonlinear mapping relation between the time-frequency representation of the frequency conversion basic harmonic component and the time-frequency representation of the speed conversion signal if the test is passed, otherwise, returning to the step 1);
5) Establishing a mean square error loss function between the variable speed signal time-frequency representation and the variable frequency fundamental harmonic time-frequency binary mask according to the variable frequency fundamental harmonic component time-frequency representation and the variable speed signal time-frequency representation nonlinear mapping relation, and storing the network parameters of the trained depth binary mask signal separation model;
6) Monitoring a current rotating machinery vibration signal, and performing low-pass filtering to obtain a rotating machinery low-pass filtering vibration signal;
performing short-time Fourier transform on the rotary mechanical low-pass filtering vibration signal to obtain a vibration signal represented by a time-frequency domain, and inputting the vibration signal represented by the time-frequency domain into a depth binary masking signal separation model to obtain a time-frequency representation of a frequency conversion basic harmonic component;
7) Performing inverse short-time Fourier transform on the time-frequency representation of the frequency-conversion fundamental harmonic component to obtain a frequency-conversion fundamental harmonic component time domain signal, and performing Hilbert transform on the frequency-conversion fundamental harmonic component time domain signal to obtain instantaneous phase information of the frequency-conversion fundamental harmonic component;
8) And (3) carrying out angular domain resampling and Fourier transformation on the rotating machinery vibration signal in the step (6) according to the instantaneous phase information to obtain an order spectrum, and carrying out rotating machinery fault diagnosis under the working condition of variable rotating speed according to the order spectrum.
Example 7:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is as shown in the embodiment 6, wherein the test template signal comprises a signal with rising speed under the variable speed condition;
the training template signal comprises a variable speed signal with the instantaneous frequency firstly increased and then decreased, the instantaneous frequency is arbitrarily changed, the instantaneous frequency is firstly increased and then fixed and then decreased, namely:
wherein f i (u) is the instantaneous frequency; i is the instantaneous frequency order; i=1, 2,3; n (t) is Gaussian white noise; x (t) is a variable speed signal; u is a time variable;
wherein the instantaneous frequency f i (u) is as follows:
f 2 (u)＝64sin(u1.5)+64(3)
where u is a time variable.
Example 8:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is shown in the embodiment 6, wherein in the step 1), a truncated simulation template signal is also adopted to expand a test template signal and a training template signal.
Example 9:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is as shown in the embodiment 6, wherein in the step 2), the simulation template signal S (t, f) of the time domain representation is as follows:
wherein T is a time variable, f is a frequency variable, h (τ -T) is a window function, and T is a time period; x (τ) is the shift signal at τ time.
Example 10:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is as shown in the embodiment 6, wherein the objective function BM (t, f) of the depth binary masking signal separation model is as follows:
wherein the time-frequency representation S (t, f 1 ) Time-frequency representation S (t, f) of harmonic components other than the fundamental harmonic of the conversion i ) The following are respectively shown:
wherein t is a time variable, f 1 、f i H (τ -t) is a window function, which is a frequency variable; t is the time period. x (τ) is the shift signal. n is the number of frequency bins.
Example 11:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is shown in the embodiment 6, wherein the depth binary masking signal separation model comprises an input layer, a full connection layer, an activation function layer, a batch normalization layer, a dropout layer and an output layer;
wherein the full connection layer outputs
The following is shown:
in the method, in the process of the invention,
input for the full connection layer; />
And->
Respectively the weight and the bias of the full connection layer;
the activation function σ (x) of the fully connected layer is as follows:
wherein x is an independent variable;
output of batch normalization layer
The following is shown:
in the method, in the process of the invention,
outputting data for full connection layer >
Is a variance of (2); />
Outputting data for full connection layer>
Is the average value of (2); />
And->
Respectively representing the scale and displacement parameters; />
Is an intermediate parameter;
output of dropout layer
The following are provided:
in the formula, w| p The subset obtained by sampling the output data of the upper layer according to the probability P.
Example 12:
a rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation is disclosed in embodiment 6, wherein the time-frequency representation S of a frequency conversion fundamental harmonic component optimal (t，f 1 ) The following is shown:
S optimal (t,f 1 )＝argmin(L) (14)
wherein the mean square error L of the loss function is as follows:
wherein f (y i j ) Signals output by the model are separated for the depth binary masking signals; n is the number of samples.
Example 13:
a rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation is disclosed in embodiment 6, wherein the frequency conversion fundamental harmonic component time domain signal x 1 (t) is as follows:
instantaneous phase information of frequency conversion fundamental harmonic component
The following is shown:
wherein H (x) 1 (t) is x 1 The Hilbert transform of (t); unwrap [. Times. ]]Is a phase unwrapping operator.
Example 14:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is as shown in embodiment 6, wherein the step of performing angular domain resampling and fourier transformation on the rotary machine vibration signal of step 6) to obtain an order spectrum comprises the following steps:
6.1 Setting an angular interval for equal angle resampling
Maximum analysis order O max And the total length N of the resampled data, i.e.:
wherein T is the total sampling time, f 1 And (t) is a converted harmonic component.
6.2 Calculating a resampling time scale T in the angular domain 1 The method comprises the following steps:
wherein a, b, c, d are equation coefficients, T 0 Starting time for time domain sampling;
6.3 Acquiring intra-angular resampling time scale T 1 Corresponding vibration amplitude, and interpolating the vibration amplitude by utilizing the Langerhans formula to obtain a time mark T 1 Amplitude in the angular region x (T 1 ) The method comprises the following steps:
x(T 1 )＝x(t i )+(x(t i+1 )-x(t i ))(t i+1 -t i )(T 1 -t i ),t i ≤T 1 ≤t i+1 (21)
wherein t is i 、t i+1 Is a time variable;
6.4 For amplitude x (T) 1 ) Fourier transform is performed to obtain an order spectrum a (k), namely:
where k is the length of the resampled data.
Example 15:
the main content of the adaptive extraction method of the instantaneous phase of the rotary machine based on the depth signal separation is as shown in embodiment 6, wherein the method for diagnosing the fault of the rotary machine under the working condition of variable rotating speed according to the order spectrum comprises the following steps:
the highest order of the amplitude value in the order spectrum A (k) is positioned, the highest order of the amplitude value is compared with the inherent order of the rotating machine under the working condition of changing the rotation speed, and whether the rotating machine has faults under the working condition of changing the rotation speed is determined.
Claims (10)
1. The rotary mechanical instantaneous phase self-adaptive extraction method based on depth signal separation is characterized by comprising the following steps of:
1) And simulating the fault state of the rotary machine under the variable speed working condition, and acquiring fault signals, so as to generate simulation template signals with different instantaneous frequencies.
Dividing the simulation template signal into a training template signal and a test template signal;
2) Adding Gaussian white noise into the simulation template signal, and performing short-time Fourier transform on the simulation template signal to obtain a time-frequency representation of the variable-speed signal;
3) Establishing a depth binary masking signal separation model, wherein an objective function of the model is an optimal time-frequency binary mask of a frequency conversion basic harmonic component;
4) Training the depth binary masking signal separation model by utilizing the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask in the training template signal until the mean square error of the variable rotation speed signal time-frequency representation and the rotation frequency fundamental harmonic time-frequency binary mask is minimum;
testing the depth binary masking signal separation model by using a test template signal, outputting a nonlinear mapping relation between the time-frequency representation of the frequency conversion basic harmonic component and the time-frequency representation of the speed conversion signal if the test is passed, otherwise, returning to the step 1);
5) Establishing a mean square error loss function between the variable speed signal time-frequency representation and the variable frequency fundamental harmonic time-frequency binary mask according to the variable frequency fundamental harmonic component time-frequency representation and the variable speed signal time-frequency representation nonlinear mapping relation, and storing the network parameters of the trained depth binary mask signal separation model;
6) Monitoring a current rotating machinery vibration signal, and performing low-pass filtering to obtain a rotating machinery low-pass filtering vibration signal;
performing short-time Fourier transform on the rotary mechanical low-pass filtering vibration signal to obtain a vibration signal represented by a time-frequency domain, and inputting the vibration signal represented by the time-frequency domain into a depth binary masking signal separation model to obtain a time-frequency representation of a frequency conversion basic harmonic component;
7) Performing inverse short-time Fourier transform on the time-frequency representation of the frequency-conversion fundamental harmonic component to obtain a frequency-conversion fundamental harmonic component time domain signal, and performing Hilbert transform on the frequency-conversion fundamental harmonic component time domain signal to obtain instantaneous phase information of the frequency-conversion fundamental harmonic component;
8) And (3) carrying out angular domain resampling and Fourier transformation on the rotating machinery vibration signal in the step (6) according to the instantaneous phase information to obtain an order spectrum, and carrying out rotating machinery fault diagnosis under the working condition of variable rotating speed according to the order spectrum.
2. The method for adaptive extraction of instantaneous phase of a rotary machine based on separation of depth signals according to claim 1, wherein said test template signal comprises a signal of rising speed under variable speed conditions;
the training template signal comprises a variable speed signal with the instantaneous frequency firstly increased and then decreased, the instantaneous frequency is arbitrarily changed, the instantaneous frequency is firstly increased and then fixed and then decreased, namely:
wherein f i (u) is the instantaneous frequency; i is the instantaneous frequency order; i=1, 2,3; n (t) is Gaussian white noise; x (t) is a variable speed signal; u is a time variable;
wherein the instantaneous frequency f i (u) is as follows:
f 2 (u)＝64sin(u/1.5)+64 (3)
where u is a time variable.
3. The adaptive extraction method of instantaneous phase of rotary machine based on depth signal separation according to claim 1, wherein in step 1), the test template signal and the training template signal are further extended by using truncated simulation template signal.
4. The method of claim 1, wherein in step 2), the simulated template signal S (t, f) of the time domain representation is as follows:
wherein T is a time variable, f is a frequency variable, h (τ -T) is a window function, and T is a time period; x (τ) is the shift signal at τ time.
5. The method for adaptive extraction of instantaneous phase of a rotating machine based on depth signal separation according to claim 1, characterized in that the objective function BM (t, f) of the depth binary masking signal separation model is as follows:
wherein the time-frequency representation S (t, f 1 ) Time-frequency representation S (t, f) of harmonic components other than the fundamental harmonic of the conversion i ) The following are respectively shown:
wherein t is a time variable, f 1 、f i Is a frequency variable, h(τ -t) is a window function; t is the time period. x (τ) is the shift signal. n is the number of frequency bins.
6. The rotary machine instantaneous phase adaptive extraction method based on depth signal separation according to claim 1, wherein the depth binary masking signal separation model comprises an input layer, a full connection layer, an activation function layer, a batch normalization layer, a dropout layer and an output layer;
wherein the full connection layer outputs
The following is shown:
in the method, in the process of the invention,
input for the full connection layer; />
And->
Respectively the weight and the bias of the full connection layer;
the activation function σ (x) of the fully connected layer is as follows:
wherein x is an independent variable;
output of batch normalization layer
The following is shown: />
In the method, in the process of the invention,
outputting data for full connection layer >
Is a variance of (2); />
Outputting data for full connection layer>
Is the average value of (2); />
And->
Respectively representing the scale and displacement parameters; />
Is an intermediate parameter;
output of dropout layer
The following are provided:
in the formula, w| p The subset obtained by sampling the output data of the upper layer according to the probability P.
7. The adaptive extraction method of instantaneous phase of rotary machine based on depth signal separation according to claim 1, wherein the time-frequency representation S of the fundamental harmonic component of the frequency conversion optimal (t，f 1 ) The following is shown:
S optimal (t,f 1 )＝arg min(L) (14)
wherein the mean square error L of the loss function is as follows:
in the method, in the process of the invention,
signals output by the model are separated for the depth binary masking signals; n is the number of samples.
8. The adaptive extraction method of instantaneous phase of rotary machine based on depth signal separation according to claim 1, wherein the time domain signal x of fundamental harmonic component of frequency conversion 1 (t) is as follows:
instantaneous phase information of frequency conversion fundamental harmonic component
The following is shown:
wherein H (x) 1 (t) is x 1 Hilbert variant of (t)Changing; unwrap [. Times. ]]Is a phase unwrapping operator.
9. The adaptive extraction method of instantaneous phase of rotary machine based on depth signal separation according to claim 1, wherein the step of performing angular domain resampling and fourier transformation on the rotary machine vibration signal of step 6) to obtain an order spectrum comprises:
1) Setting an angular interval for equal angle resampling
Maximum analysis order O max And the total length N of the resampled data, i.e.:
wherein T is the total sampling time, f 1 And (t) is a converted harmonic component.
2) Calculating a resampling time scale T in the angular domain 1 The method comprises the following steps:
wherein a, b, c, d are equation coefficients, T 0 Starting time for time domain sampling;
3) Acquiring an intra-angular resampling time scale T 1 Corresponding vibration amplitude, and interpolating the vibration amplitude by utilizing the Langerhans formula to obtain a time mark T 1 Amplitude in the angular region x (T 1 ) The method comprises the following steps:
x(T 1 )＝x(t i )+(x(t i+1 )-x(t i ))/(t i+1 -t i )(T 1 -t i ),t i ≤T 1 ≤t i+1 (21)
wherein t is i 、t i+1 Is a time variable;
4) For amplitude x (T) 1 ) Fourier transform is performed to obtain an order spectrum a (k), namely:
where k is the length of the resampled data.
10. The adaptive extraction method of instantaneous phase of rotary machine based on depth signal separation according to claim 1, wherein the method for diagnosing faults of rotary machine under variable rotation speed working condition according to order spectrum comprises:
the highest order of the amplitude value in the order spectrum A (k) is positioned, the highest order of the amplitude value is compared with the inherent order of the rotating machine under the working condition of changing the rotation speed, and whether the rotating machine has faults under the working condition of changing the rotation speed is determined.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211740653.4A CN116256158A (en) | 2022-12-31 | 2022-12-31 | Rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211740653.4A CN116256158A (en) | 2022-12-31 | 2022-12-31 | Rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116256158A true CN116256158A (en) | 2023-06-13 |
Family
ID=86685537
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211740653.4A Pending CN116256158A (en) | 2022-12-31 | 2022-12-31 | Rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116256158A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451057A (en) * | 2023-06-16 | 2023-07-18 | 中汽研新能源汽车检验中心(天津)有限公司 | Instantaneous rotating speed frequency identification method for motor rotor system |
-
2022
- 2022-12-31 CN CN202211740653.4A patent/CN116256158A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451057A (en) * | 2023-06-16 | 2023-07-18 | 中汽研新能源汽车检验中心(天津)有限公司 | Instantaneous rotating speed frequency identification method for motor rotor system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Rolling element bearing fault diagnosis via fault characteristic order (FCO) analysis | |
| Urbanek et al. | Comparison of amplitude-based and phase-based method for speed tracking in application to wind turbines | |
| CN110617964A (en) | Synchronous compression transformation order ratio analysis method for fault diagnosis of rolling bearing | |
| CN111665051A (en) | Bearing fault diagnosis method under strong noise variable-speed condition based on energy weight method | |
| CN110907162B (en) | Rotating machinery fault feature extraction method without tachometer under variable rotating speed | |
| CN112665851B (en) | Key-free phase change rotating speed gearbox fault diagnosis method | |
| Li et al. | Automated and adaptive ridge extraction for rotating machinery fault detection | |
| CN116256158A (en) | Rotary machine instantaneous phase self-adaptive extraction method based on depth signal separation | |
| Li et al. | Adaptive cost function ridge estimation for rolling bearing fault diagnosis under variable speed conditions | |
| Lin et al. | A review and strategy for the diagnosis of speed-varying machinery | |
| CN111256993A (en) | Method and system for diagnosing fault type of main bearing of wind turbine generator | |
| CN111738068A (en) | Transmission shaft fault diagnosis method and system under working condition of rotating speed fluctuation | |
| Dziedziech et al. | A novel method for speed recovery from vibration signal under highly non-stationary conditions | |
| CN108398260B (en) | Method for quickly evaluating instantaneous angular speed of gearbox based on mixed probability method | |
| CN112362343A (en) | Distributed fault feature extraction method for gearbox under variable rotating speed based on frequency modulation dictionary | |
| Li et al. | Application of a Method of Identifiying Instantaneous Shaft Speed from Spectrum in Aeroengine Vibration Analysis | |
| Zhao et al. | Adaptive scaling demodulation transform: Algorithm and applications | |
| Randall et al. | Gear diagnostics under widely varying speed conditions | |
| CN112465068A (en) | Rotating equipment fault feature extraction method based on multi-sensor data fusion | |
| Zhu et al. | Fault diagnosis method for rolling element bearings under variable speed based on TKEO and fast-SC | |
| CN114383718B (en) | High-frequency blade passing frequency extraction method based on vibration signals of external casing of gas turbine | |
| CN116383629A (en) | Method for diagnosing faults of variable-rotation-speed rolling bearing | |
| CN112781723B (en) | Harmonic component detection method based on frequency spectrum variance | |
| He et al. | Adaptive fast Chirplet transform and its application into fault diagnosis of rolling bearing under time-varying speed condition | |
| CN112784702A (en) | Signal processing method based on self-adaptive rotation synchronous extraction frequency modulation transformation |
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 |