CN104819846B - Rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and sparse laminated automatic encoder - Google Patents
Rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and sparse laminated automatic encoder Download PDFInfo
- Publication number
- CN104819846B CN104819846B CN201510169697.XA CN201510169697A CN104819846B CN 104819846 B CN104819846 B CN 104819846B CN 201510169697 A CN201510169697 A CN 201510169697A CN 104819846 B CN104819846 B CN 104819846B
- Authority
- CN
- China
- Prior art keywords
- rolling bearing
- data
- fault
- sound signal
- sae
- 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.)
- Expired - Fee Related
Links
Abstract
The invention discloses a rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and a sparse laminated automatic encoder. According to the method of the invention, firstly a smart mobile phone is used for acquiring a sound signal of the rolling bearing fault; then short-time Fourier analysis is performed on the sound signal for obtaining a spectrogram matrix; then the modulus value of the matrix is acquired and gray scale normalization processing is performed; then the normalized data are selected and input into a deep studying network for automatically extracting characteristics; and finally the characteristic which is extracted by a neural net is input into a Softmax classifier for identifying the fault mode. The invention provides the rolling bearing sound signal fault diagnosis method based on smart mobile phone sound signal short-time Fourier transform (STFT) and the sparse laminated automatic encoder (SAE). Through testing result analysis, the rolling bearing sound signal fault diagnosis method can accurately determine the fault mode of the rolling bearing.
Description
Technical field
The present invention relates to the technical field of rolling bearing fault diagnosis, and in particular to it is a kind of based on Short Time Fourier Transform and
The rolling bearing acoustical signal method for diagnosing faults of sparse stacking autocoder.
Background technology
Rolling bearing is one of wide variety of standard component in various plant equipment, and rolling bearing fault is also to cause
One of main reason of mechanical disorder.According to statistics, there are about 30% rotating machinery fault it is relevant with rolling bearing fault.
In the General Maintenance of the railway system, workman taps locomotive wheel with iron hammer, can just judge car according to sound is tapped
Whether crackle is had in wheel.In the same manner, in many engineering fields, veteran attendant can be according to sound during machine operation
Sound feature judges machine whether normal operation.Its potential physical principle is to change its characteristic frequency as parts are damaged,
And then change the tone of sound.Coupled vibration in rolling bearing operation process can equally cause air to compress and generation sound
Sound, wherein includes the fault message of rolling bearing, therefore can obtain the axis of rolling by carrying out feature extraction to acoustical signal
The fault message for holding.Voice signal is a kind of typical non-stationary signal, and it is by the physical motion mistake of sounding body which is non-stationary
What journey was produced, this process is slower compared with the speed of acoustic vibration, it can be assumed that within the short time as 10～30ms be
Smoothly.Fourier analyses are the strong means for analyzing the steady signal steady-state characteristic of linear system peace, and Fourier in short-term
Analysis, is also time-dependent Fourier transformation, is exactly, under the hypothesis of short-term stationarity, to process non-stationary letter with steady-state analysis method
Number a kind of method.Sound spectrograph is that the time intensity of voice signal short-term spectrum is represented.Voice signal is first segmented into
Some sections overlapped (frames), to each section of windowing process, then carry out fast Fourier transform, obtain the short of this signal
Time-frequency Power estimation.
Obtain rolling bearing sound language modal data after, directly carry out as its dimension is higher pattern recognition it is computationally intensive and
Accuracy of identification is not high, so that carrying out the extraction of feature to which.Using deep neural network, can fully describe between feature
Dependency relation.Therefore, we voice time-frequency characteristics send into stacking autocoder deep learning network carry out feature from
It is dynamic to extract.As deep learning network is using the multiple structure of simulation human brain, the extraction of information characteristics can be carried out step by step, most
More satisfactory feature of the end form into suitable pattern classification.This profound structure and human brain processing information have very big similarity, are point
What layer was processed.The modeling technique of deep learning network, can seamlessly with traditional time frequency processing technology and Softmax graders
Combine, in the case where any extra consumption is not caused, significantly lift the degree of accuracy of diagnosis.
It is described above that to carry out diagnosis based on rolling bearing operating sound data be practicable.With electronic information technology
Develop by leaps and bounds, smart mobile phone has become the life instrument that we be unable to do without, however, smart mobile phone is except the life as us
Outside instrument living, it can be with by the use of specialty.Compared with traditional vibrating sensor gathered data, gathered using smart mobile phone
Failure voice data has many advantages：One is motility, and we can carry out data to the operating condition of equipment anywhere or anytime
Collection, without pre-installing sensor in advance in plant equipment, does not also need the installation site of analyte sensors；Two is economical
Property, thousand of units are high then up to ten thousand at least for traditional high-precision sensor, expensive, and only need to the smart mobile phone with sound-recording function
Us can be just helped to collect the data message of specialty, the specialization for realizing daily life instrument is used, convenient and swift, simply
Effectively；Three is the suitability, and different equipment, different operating modes may be needed using different types of vibrating sensor, sensor
Type is more, selects them to need relevant professional knowledge and experience, and smart mobile phone is adapted to gather the acoustic information under various operating modes,
The scope of application is relatively extensive.
Smart mobile phone has become the important component part of daily life, although our recording work(to smart mobile phone
Can be customary, but it is also rarely found in the fault diagnosis of equipment as the data sampling sensor of fault message.
The content of the invention
It is an object of the invention to：A kind of rolling based on Short Time Fourier Transform and sparse stacking autocoder is provided
Bearing acoustical signal method for diagnosing faults, the method gather rolling bearing fault acoustical signal first by smart mobile phone, then
Short Time Fourier Analysis are carried out to acoustical signal, sound spectrograph matrix is obtained, matrix norm value is then obtained and is carried out gray scale normalizing
Change is processed, then is input to deep learning network and carries out automatically extracting for feature the data after normalization Jing after choosing, and finally will
The feature input Softmax graders that neutral net is extracted carry out the identification of fault mode.The method can diagnose rolling exactly
The fault mode of dynamic bearing.
The technical solution used in the present invention is：A kind of rolling based on Short Time Fourier Transform and sparse stacking autocoder
Dynamic bearing acoustical signal method for diagnosing faults, the method step are as follows：
The first step：Obtain data
The acoustical signal during bearing operation is recorded on request using smart mobile phone, and does certain editing process；
Second step：Voice signal Short Time Fourier Analysis (STFT)
Program reads in pretreated acoustical signal, and Matlab obtains its sound spectrograph and language with spectrogram functions
Spectrum matrix；
3rd step：Data normalization and selection
Modulus computing is carried out to the element of sound spectrograph matrix first, then by data by the grey that row normalization is 0～255
According to value, gray matrix is obtained, the data input of each column mid portion certain length is then chosen to stacking autocoder；
4th step：Unsupervised self study process based on SAE
Sound language modal data using the rolling bearing after previous processed under different faults pattern, by unsupervised self-study
Practise training SAE models；Unsupervised self study process is the process of a deep learning, and SAE model frameworks are with two hidden layers
Neutral net, final character representation, and the output of first hidden layer of neutral net are obtained by the Level by level learning to data
As the input of second hidden layer；
5th step：Fault diagnosis based on Softmax regression algorithms
Select Softmax to return as rolling bearing fault algorithm for pattern recognition, the characteristic vector obtained using SAE as
The input of Softmax algorithms, solves classification problem of the rolling bearing under multiple fault modes, by minimizing cost function meter
The probability that each classification results occurs is calculated, if the probit that certain fault signature is calculated is maximum, that is, determines failure for currently estimating
The fault mode of meter.
Present invention advantage compared with prior art is：
By analyzing the rolling bearing fault of the STFT based on acoustical signal proposed by the present invention and stacking autocoder
The result of the test that diagnostic method is applied in rolling bearing laboratory table, it can be deduced that：
1) acoustical signal in the rolling bearing operation process of smart mobile phone collection of the present invention includes its fault message.
2) characteristic vector that the unsupervised self study process of present invention stacking autocoder is obtained can reconstruct initial data,
Further indicating that for feature is effectively realized, is conducive to the substitutive characteristics of mining data.
3) present invention is for only with the relatively low problem of unsupervised self-learning method fault diagnosis precision, reverse by being based on
The fine setting of propagation algorithm, can effectively improve the precision of fault diagnosis so that classification accuracy reaches more than 99%.
4) fault diagnosis model of the present invention based on STFT and SAE is short and sweet effectively, mainly includes three parts：A) use
Smart mobile phone obtains rolling bearing operating sound signal；B) the positive unsupervised self study of initial data is carried out based on SAE, and is adopted
Reversely finely tuned with back-propagation algorithm, the foundation and optimization of implementation model；C) realize rolling by Softmax homing methods
The Fault Pattern Recognition of bearing.
Description of the drawings
Fig. 1 is the Fault Diagnosis of Roller Bearings flow chart based on STFT and SAE；
Fig. 2 is signal framing schematic diagram；
Fig. 3 is that sound spectrograph produces schematic flow sheet；
Fig. 4 is that acoustical signal language composes gray-scale maps；
Single order feature representation schematic diagrams of the Fig. 5 for bearing data；
Second order feature representation schematic diagrams of the Fig. 6 for bearing data；
Fig. 7 is the bearing fault classification schematic diagram based on Softmax regression models；
Fig. 8 is SAE model structure schematic diagrams；
Fig. 9 is Level by level learning network diagram；
Figure 10 is that two hidden layers are laminated autocoder schematic diagram；
Figure 11 is cylinder roller bearing testing stand figure.
Specific embodiment
Below in conjunction with the accompanying drawings and specific embodiment further illustrates the present invention.
A kind of rolling bearing acoustical signal method for diagnosing faults flow process based on STFT and SAE of the present invention is as shown in Figure 1.Tool
Body flow process can be summarized as following five step：
The first step：Obtain data
The acoustical signal during bearing operation is recorded on request using smart mobile phone, and does certain editing process.
Second step：Acoustical signal Short Time Fourier Analysis (STFT)
Program reads in pretreated acoustical signal, and Matlab obtains its sound spectrograph and language with spectrogram functions
Spectrum matrix.
As a whole, the parameter for characterizing acoustical signal is real-time change, but very short time (20～
It is relatively stable in the range of 30ms), thus a quasi-steady state process can be regarded as.The purpose of adding window is exactly that acoustical signal is divided into
One short time period.By acoustical signal adding window, framing, frame length is designated as N (in units of ms), and it is left that frame number per second is about 30 frames
The right side, the general method using overlapping segmentation.As shown in Figure 2.
It is rectangular window and Hamming window (Hamming) function below, its expression formula is shown in formula (1) and (2), and N is frame length.
Rectangular window：
Hanning window：
Acoustical signal adding window is the gradient for reducing two ends.The secondary lobe of rectangular window is too high, undesirable.
The Short Time Fourier Transform (STFT) of signal s (t) is defined as follows：
Wherein, w (t) is certain window function.
The discrete time STFT expression formula of n is as follows at any time：
Discrete STFT can be obtained by frequency sampling：
S (n, k)=S (t, f) |F=k/N, t=nT (5)
Wherein, N is the sum of data point in window function, and the frequency sampling factor.Wushu (5) brings formula (4) into, obtain from
Scattered STFT：
Wherein 0≤k≤(N-1), then | | x (n, k) | | is exactly the estimation of x (n) short time amplitudes spectrum, and frequency spectrum energy at time m
Metric density function (or power spectrum function) P (n, k) is：
P (n, k)=| x (n, k) |2=(x (n, k)) x (conj (x (n, k))) (7)
Then P (n, k) is two-dimentional non-negative real-valued function, and it is easy to show that it is the short-time autocorrelation function of signal x (n)
Fourier transformation.With time n as abscissa, k makees vertical coordinate, and it is exactly sound spectrograph that the value of P (n, k) is expressed as pcolor.
Spectrogram algorithms are a kind of parsers, and the two dimensional image form output that it produces a voice signal is (same
Shi Yike obtains numerical matrix).Sound spectrograph use time n does abscissa, frequency f as vertical coordinate, by energy density spectral function
Value be expressed as two dimension pcolor.The time-frequency figure of this reaction voice signal dynamic spectrum characteristic has weight in speech analysises
The practical value wanted, also becomes " visual language ".
The generation schematic flow sheet of sound spectrograph：
Some frequency-domain analysiss parameters (such as formant, pitch period etc.) can be obtained from sound spectrograph with speech utterance process
The situation of change of (time)；Situation of change of the energy with speech utterance process (time), each pixel of image can also be obtained
Pseudo- color-values (or gray value) size represent the signal energy density of corresponding moment and corresponding frequencies.
In this research, the author selects frame length (window) for 44100 (equal to sample frequencys), slip length
(noverlap) it is 44000, discrete Fourier transform length (nfft) is 44100 (equal with window length, sample frequency), is sampled
Frequency fs=44100 is generating sound spectrograph.Window function selects Hanning windows.
3rd step：Data normalization and selection
For convenience of follow-up data process work, by fast Fourier transform (FFT) result of each frame by row normalization it is
Gray value, method for normalizing adopt min-max standardization (Min-Max Normalization), also referred to as deviation standardization.
Min-max standardization is the linear transformation to initial data, and data are mapped as the integer value between [0～255], is obtained such as figure
Gray-scale maps shown in 4.Transfer function is as follows：
Wherein, min is row minima, and max is row maximum.
Modulus computing is carried out to the element of sound spectrograph matrix first, then by data by the grey that row normalization is 0～255
According to value, gray matrix is obtained, the data input of each column mid portion certain length is then chosen to stacking autocoder.
4th step：Unsupervised self study process based on SAE
Using sound language modal data of the rolling bearing after processing in advance under different faults pattern, by unsupervised self-study
Practise training SAE models.Unsupervised self study process is the process of a deep learning, and SAE model frameworks are with two hidden layers
Neutral net, final character representation, and the output of first hidden layer of neutral net are obtained by the Level by level learning to data
As the input of second hidden layer.
Specifically, in the unsupervised self study of Rolling Bearing Fault Character based on SAE,
1st, SAE model structures
SAE algorithm models are deep learning neutral nets being made up of the sparse autocoder of multilamellar, its preceding layer
Input of the output of self-encoding encoder as later layer self-encoding encoder.
In terms of the setting of network parameter, the initialization of SAE Model Neural parameters adopts successively greedy coaching method, i.e.,
First with the ground floor that carrys out training network is originally inputted, weights and offset parameter is obtained；Then network first tier will be originally inputted
The vector (assuming that the vector is A) being made up of hidden unit activation value is transformed into, now single order mark sheets of the A for initial data
Show, i.e. the expression of the bottom；Then A is continued the parameter that training obtains the second layer as the input of the second layer, the layer is joined
Number is the second order feature obtained by single order feature reconstruction, is more abstract expression；Finally, to each layer below using same
The strategy of sample, will front layer output as next layer input mode train successively, and train each layer parameter when
Fix other each layer parameters and keep constant.The output of last hidden layer is input to softmax classification by we
Device carries out pattern recognition so that the Grad of the classification error of grader (softmax) is directly opposite and is broadcast to coding layer (essence
Adjust).
For the SAE models of the unsupervised self study process of rolling bearing initial data, first with training sample x(k)Training
First self-encoding encoder, can learn the single order character representation h for obtaining being input into(1)(k)(as shown in Figure 5).
Then initial data is input in the above-mentioned self-encoding encoder for training, for each input x(k), can obtain
To its corresponding single order character representation h(1)(k), then again with these single order features as another self-encoding encoder input instructing
Practice study and obtain its second order feature h(2)(k)(as shown in Figure 6).
Finally using second order feature as softmax graders input, training obtains one and can arrive second order Feature Mapping
The model (as shown in Figure 7) of digital label.
Finally, the single-layer model that we can be trained by more than is connected in turn, and builds one and hides comprising two
The stack autoencoder network of layer and a final softmax grader layer, i.e., required SAE models, its overall structure such as Fig. 8
It is shown.
2nd, the unsupervised Level by level learning (" pre-training ", " coarse adjustment ") of SAE
The purpose of pre-training is that all weights are linked and biased to be limited in certain parameter space, prevents random initial
The quality factor that each hidden layer occurs and then reduces of change, are easy to carry out systemic parameter optimization to whole neutral net, should
The core of algorithm is all initialized the input layer of SAE and hidden layer with unsupervised method, then again with successively greed instruction
Practice algorithm and each hidden layer is trained for into auto-associating device, realize the reconstruct of input data, its basic step may be summarized as follows：
1) ground floor of neutral net is trained in unsupervised mode, is output it as the minimum reconstruct being originally inputted
Error；
2) input of the output of each implicit unit as next layer of neutral net, with unlabeled data sample to next layer
It is trained, by error control within the specific limits；
3) repeat step 2), till completing the training of specified quantity hidden layer；
4) using the output of last hidden layer as the input for having monitor layer, and initialization has the parameter of monitor layer,
Pre-training process is as shown in Figure 9.
By lasting pre-training, SAE can replace RBM to build depth structure, the depth structure stability built by it
It is higher more reliable.Importantly, network parameter can be adjusted according to the needs of different task, controllability is higher.
3rd, the trim process (" accurate adjustment ") based on the back-propagation algorithm (BP) for having supervision
Fine setting be called accurate adjustment, be build SAE steps necessary, complete generally using BP algorithm this task (Newton method,
Conjugate gradient method, MOBP and SDBP etc., the deformation of BP algorithm can also be used for accurate adjustment).The core concept of accurate adjustment is by automatic encoding
The input layer of device, output layer and all hidden layers are considered as an entirety, are further adjusted through pre- instruction with supervised learning algorithm
Experienced neutral net, after successive ignition, all weights and biasing can be optimised.As last hidden layer can only be defeated
Go out the reconstruct of initial data, because not with Classification and Identification function.There is the function of Classification and Identification in order to allow SAE, need complete
Into after the output layer of the neutral net of accurate adjustment add SoftMax graders, by whole neural metwork training into can complete layering
Feature extraction or the multiple perceptron of data sorting task.
After the activation value of all neurons during forward conduction calculates network, we are directed to neutral net hidden layer and defeated
Go out each node i of layer, calculate residual error δi (l), affect to characterize the residual error on final output value.Final output residual error
Difference between data and initial data is reconstructed as.And for each hidden layer, can be using based on node residual error
Weighted average value calculating method.Trim step based on back-propagation algorithm is as follows：
1. output layer n1The residual computations method of each output unit i is：
Wherein yiFor real output value, aiFor neural network prediction output valve, f is transmission function.In calculating process, lead to
F'(z is obtained ofteni (l))=ai (l)(1-ai (l))。
2. for nl-1,nl-2..., 2 layers, the residual computations method of i-th node is：
δ(l)=((W(l))Tδ(l+1))·f'(z(l)) (10)
Weights of the wherein W for l layers, δ(l+1)Residual error of the layer for last layer.
3. calculation cost function weights and biasing local derviation value：
Therefore, we can update network weight and offset parameter according to calculated residual error come iteration, so as to realize
The optimization of whole SAE model parameters, as shown in Figure 10.
The SAE for completing pre-training and accurate adjustment has the basic feature of biological nervous system, reflects people to a certain extent
Some functions of brain, are the successful simulations to biosystem, realize the combination of unsupervised learning and supervised learning, and not
Different features can be acquired with hidden layer, low order hidden layer provides feature mode for high-order hidden layer, be easy to high-order hidden layer to enter
Row mode is combined.Meanwhile, the topological structure of SAE is complete, with powerful nonlinear fitting ability, can find the distributed of data
Character representation, extracts the potential layered characteristic of data, obtains " the hierarchical packet " or " part is overall to decompose " being originally inputted
Structure, thus input data can be preferably represented, it is excellent with large-scale parallel, distributed treatment, self-organizing and self study etc.
Point.
5th step：Fault diagnosis based on Softmax regression algorithms
Softmax regression algorithms are selected as rolling bearing algorithm for pattern recognition.Made with the characteristic vector that SAE algorithms are obtained
For the input of Softmax algorithms, classification problem of the rolling bearing under multiple fault modes is solved.By minimizing cost function
The probability that each classification results occurs is calculated, if the probit that certain fault signature is calculated is maximum, that is, determines that failure is current
The fault mode of estimation.
Specifically, during the rolling bearing fault diagnosis returned based on Softmax,
Rolling bearing has multiple fault modes such as inner ring failure, outer ring failure, rolling element failure under normal circumstances, for
The classification of multiple faults pattern, the method that this programme is returned using Softmax, the probability size for calculating every kind of fault mode respectively are made
For the judgment basis of failure modes.Assume a total of k class categories, the equation of system is in softmax is returned：
Parameter θ therein is no longer column vector, but a matrix, it is the ginseng of a classification correspondence grader per a line
Number, amounts to k rows.Therefore, θ can be written as form：
When using gradient descent method Optimized model parameter, in Softmax recurrence, the partial derivative of loss function is：
Itself it is one vectorial, its l-th elementIt is J (θ) to θjL-th component local derviation
Number.1 {. } is an indicative function, i.e., when the value in braces is true time, the result of the function is just 1, and otherwise its result is just
For 0.
In order to avoid the not unique problem of optimal solution, in loss function regularization term is added to make cost function become strict
Convex function, so as to define new partial derivative be：
By minimizingAn effective softmax disaggregated model can just be built.
Application case of the present invention is as follows：
1st, rolling bearing voice data prepares
As shown in figure 11, bearing is cylinder roller bearing to rolling bearing test table.In process of the test, speed setting is
1200r/min, corresponding axle frequency are 20Hz.Voice data collection is being gathered using the recording software in Samsung note3 mobile phones
During by mobile phone near the bearing of bearing test-bed, sample frequency is 44.1kHz.Institute's gathered data covers normal shape
State, outer ring failure, inner ring failure and rolling element failure totally 4 kinds of fault modes.
2nd, rolling bearing fault class test analysis under the conditions of voice data
Under the conditions of voice data, cylinder roller bearing fault diagnosis model is made up of three parts, and STFT spectrum analysis (compose by language
Figure), it is laminated sparse autocoder and Softmax disaggregated models.
STFT spectrum analysis parameter settings：
Window function acquiescence selects to be hanning windows that window width (window) i.e. frame length is 44100 (equal to sample frequencys),
Slip length (noverlap) is 44000, and discrete Fourier transform length (nfft) is 44100 (equal with window width), and sampling is frequently
Rate fs=44100.
It is laminated sparse autocoder parameter setting：
Ground floor：Input number of nodes 6500, hidden layer nodes 1000, degree of rarefication are set to 0.1, and cost function is openness
Penalty factor is set to 3, and cost function weight decay factor is set to 0.003.
The second layer：Input number of nodes 1000 (the hidden layer nodes of the sparse autocoder of ground floor), hides node layer
Number 100, degree of rarefication are set to 0.1, and the openness penalty factor of cost function is set to 3, and cost function weight decay factor is arranged
For 0.003.
Third layer：Input number of nodes 100 (the hidden layer nodes of the sparse autocoder of the second layer), hidden layer nodes
10, degree of rarefication is set to 0.1, and the openness penalty factor of cost function is set to 3, and cost function weight decay factor is set to
0.003。
The parameter setting of Softmax disaggregated models：10 (the hidden layer section of the sparse autocoder of third layer of input number of nodes
Points), output node number 4 (carries out the other classification of 4 species altogether：Normally, rolling element wear-out failure, inner ring wear-out failure and outer
Circle wear-out failure).
3rd, analysis of experiments result
Using the sparse autocoder of the stacking of above-mentioned configuration and Softmax disaggregated models, justified using acoustical signal
Post roller bearing fault diagnosis.All data are halved, are selected successively using the method for 2 groups of crosschecks by model performance test
One group in each group of data test data for diagnostic cast is selected, remaining one group is model training data, and data composition situation is such as
Shown in table 1.
1 cylinder roller bearing of table, 2 groups of crosscheck data compositions
The failure for being laminated sparse autocoder and Softmax disaggregated models two groups of data above-mentioned to cylinder roller bearing is examined
Disconnected result is as shown in table 2.After weight fine tuning, the classification accuracy rate of each group is above 97%, the classification accuracy of two groups of data
Meansigma methodss have reached 97.84%.
Table 2 cylinder roller bearing, 2 groups of crosscheck fault diagnosis results
Be laminated sparse autocoder and Softmax disaggregated models it is normal to the cylinder roller bearing of above-mentioned two groups of data,
Rolling element wear-out failure, four kinds of states of inner ring wear-out failure and outer ring wear-out failure Fault Pattern Recognition result as shown in table 3,
Fault Pattern Recognition result is the result after weight fine tuning.It is laminated sparse autocoder and Softmax disaggregated models pair
Four kinds of states of cylinder roller bearing are respectively provided with good recognition effect, and identification error rate meansigma methodss are respectively less than 0.03.
The each Fault Pattern Recognition result of table 3 cylinder roller bearing, 2 groups of crosschecks
The techniques well known being related in the present invention is not elaborated.
Claims (1)
1. a kind of rolling bearing acoustical signal fault diagnosis side based on Short Time Fourier Transform and sparse stacking autocoder
Method, it is characterised in that：The method step is as follows：
The first step：Obtain data
The acoustical signal during bearing operation is recorded on request using smart mobile phone, and does certain editing process；
Second step：Acoustical signal Short Time Fourier Analysis (STFT)
Program reads in pretreated acoustical signal, and Matlab obtains its sound spectrograph and sound spectrograph with spectrogram functions
Matrix；
3rd step：Data normalization and selection
Modulus computing is carried out to the element of sound spectrograph matrix first, then by data by the gradation data that row normalization is 0～255
Value, obtains gray matrix, then chooses the data input of gray matrix each column mid portion certain length to stacking automatic encoding
Device；
4th step：Unsupervised self study process based on SAE
Sound language modal data using previously processed rolling bearing under different faults pattern, is instructed by unsupervised self study
Practice SAE models；Unsupervised self study process is the process of a deep learning, and SAE model frameworks are the god with two hidden layers
Jing networks, obtain final character representation, and the output conduct of first hidden layer of neutral net by the Level by level learning to data
The input of second hidden layer；
5th step：Fault diagnosis based on Softmax regression algorithms
Select Softmax to return as rolling bearing fault algorithm for pattern recognition, the characteristic vector obtained using SAE as
The input of Softmax algorithms, solves classification problem of the rolling bearing under multiple fault modes, by minimizing cost function meter
The probability that each classification results occurs is calculated, if the probit that certain fault signature is calculated is maximum, that is, determines failure for currently estimating
The fault mode of meter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510169697.XA CN104819846B (en) | 2015-04-10 | 2015-04-10 | Rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and sparse laminated automatic encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510169697.XA CN104819846B (en) | 2015-04-10 | 2015-04-10 | Rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and sparse laminated automatic encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104819846A CN104819846A (en) | 2015-08-05 |
| CN104819846B true CN104819846B (en) | 2017-03-22 |
Family
ID=53730205
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510169697.XA Expired - Fee Related CN104819846B (en) | 2015-04-10 | 2015-04-10 | Rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and sparse laminated automatic encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104819846B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023099849A1 (en) * | 2021-12-02 | 2023-06-08 | Safran Aircraft Engines | Method for automatically diagnosing a part |
Families Citing this family (47)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105139864B (en) * | 2015-08-17 | 2019-05-07 | 北京眼神智能科技有限公司 | Audio recognition method and device |
| CN105225049B (en) * | 2015-10-09 | 2020-02-14 | 清华大学深圳研究生院 | High-expandability fault diagnosis expert system |
| CN105975988B (en) * | 2016-04-29 | 2019-10-25 | 东莞理工学院 | Gear distress test method based on convolution spectrum autocoding support vector machines |
| CN106124212B (en) * | 2016-06-16 | 2018-09-04 | 燕山大学 | Fault Diagnosis of Roller Bearings based on sparse coding device and support vector machines |
| CN106053074B (en) * | 2016-08-02 | 2019-12-20 | 北京航空航天大学 | Rolling bearing sound signal fault feature extraction method based on STFT and rotational inertia entropy |
| CN106251375B (en) * | 2016-08-03 | 2020-04-07 | 广东技术师范学院 | Deep learning stack type automatic coding method for general steganalysis |
| CN106323636A (en) * | 2016-08-16 | 2017-01-11 | 重庆交通大学 | Adaptive extraction and diagnosis method for degree features of mechanical fault through stack-type sparse automatic coding depth neural network |
| CN106596123B (en) * | 2016-11-22 | 2020-03-27 | 东软集团股份有限公司 | Method, device and system for diagnosing equipment fault |
| CN106847309A (en) * | 2017-01-09 | 2017-06-13 | 华南理工大学 | A kind of speech-emotion recognition method |
| EP3585711A4 (en) * | 2017-02-22 | 2021-01-27 | CMTE Development Limited | Optical acoustic sensing system and method |
| CN106980831A (en) * | 2017-03-17 | 2017-07-25 | 中国人民解放军国防科学技术大学 | Based on self-encoding encoder from affiliation recognition methods |
| CN106709482A (en) * | 2017-03-17 | 2017-05-24 | 中国人民解放军国防科学技术大学 | Method for identifying genetic relationship of figures based on self-encoder |
| CN107702922B (en) * | 2017-09-11 | 2020-10-09 | 南京信息工程大学 | Rolling bearing fault diagnosis method based on LCD and stacking automatic encoder |
| CN107992843B (en) * | 2017-12-13 | 2021-08-17 | 武汉科技大学 | Sparse feature extraction method for fault signals of rotary machine |
| CN108062572B (en) * | 2017-12-28 | 2021-04-06 | 华中科技大学 | Hydroelectric generating set fault diagnosis method and system based on DdAE deep learning model |
| CN108847252B (en) * | 2018-04-26 | 2022-12-02 | 杭州电子科技大学 | Acoustic feature extraction method based on acoustic signal spectrogram texture distribution |
| CN110619342B (en) * | 2018-06-20 | 2023-02-03 | 鲁东大学 | Rotary machine fault diagnosis method based on deep migration learning |
| CN109192226A (en) * | 2018-06-26 | 2019-01-11 | 深圳大学 | A kind of signal processing method and device |
| CN108956145A (en) * | 2018-07-17 | 2018-12-07 | 中国科学院沈阳自动化研究所 | Based on the lossless sparse Fault Diagnosis of Roller Bearings from coding of constraint noise reduction |
| CN109086888B (en) * | 2018-09-05 | 2020-10-27 | 北京机械设备研究所 | Automobile engine fault determination method and device based on voice recognition |
| CN109087634A (en) * | 2018-10-30 | 2018-12-25 | 四川长虹电器股份有限公司 | A kind of sound quality setting method based on audio classification |
| CN109446997A (en) * | 2018-10-31 | 2019-03-08 | 国网山东省电力公司淄博供电公司 | Document code automatic identifying method |
| CN109632309A (en) * | 2019-01-17 | 2019-04-16 | 燕山大学 | Based on the rolling bearing fault intelligent diagnosing method for improving S-transformation and deep learning |
| CN109975702B (en) * | 2019-03-22 | 2021-08-10 | 华南理工大学 | Direct-current gear reduction motor quality inspection method based on circulation network classification model |
| CN110059601B (en) * | 2019-04-10 | 2021-04-13 | 西安交通大学 | Intelligent fault diagnosis method for multi-feature extraction and fusion |
| CN110132554B (en) * | 2019-04-17 | 2020-10-09 | 东南大学 | Rotary machine fault diagnosis method based on deep Laplace self-coding |
| CN110322896A (en) * | 2019-06-26 | 2019-10-11 | 上海交通大学 | A kind of transformer fault sound identification method based on convolutional neural networks |
| CN112326210A (en) * | 2019-07-17 | 2021-02-05 | 华北电力大学（保定） | Large motor fault diagnosis method combining sound vibration signals with 1D-CNN |
| CN110890102A (en) * | 2019-09-07 | 2020-03-17 | 创新奇智(重庆)科技有限公司 | Engine defect detection algorithm based on RNN voiceprint recognition |
| CN111122163A (en) * | 2019-09-19 | 2020-05-08 | 人本集团有限公司 | Bearing fault detection system |
| CN110595780B (en) * | 2019-09-20 | 2021-12-14 | 西安科技大学 | Bearing fault identification method based on vibration gray level image and convolution neural network |
| CN110852154B (en) * | 2019-09-29 | 2022-10-14 | 广东石油化工学院 | Rolling bearing fault diagnosis method and device based on deep learning and sound waveform images and readable storage medium |
| CN112949671B (en) * | 2019-12-11 | 2023-06-30 | 中国科学院声学研究所 | Signal classification method and system based on unsupervised feature optimization |
| CN111026058B (en) * | 2019-12-16 | 2021-01-12 | 浙江大学 | Semi-supervised deep learning fault diagnosis method based on Watherstein distance and self-encoder |
| CN111276187B (en) * | 2020-01-12 | 2021-09-10 | 湖南大学 | Gene expression profile feature learning method based on self-encoder |
| CN111256806A (en) * | 2020-01-20 | 2020-06-09 | 福州大学 | Non-contact vibration frequency composition measuring method |
| CN111432305A (en) * | 2020-03-27 | 2020-07-17 | 歌尔科技有限公司 | Earphone alarm method and device and wireless earphone |
| CN112613481A (en) * | 2021-01-04 | 2021-04-06 | 上海明略人工智能(集团)有限公司 | Bearing abrasion early warning method and system based on frequency spectrum |
| CN112926014B (en) * | 2021-01-19 | 2023-08-29 | 北京化工大学 | Rolling bearing acoustic signal multiband fusion fault diagnosis method based on RLS and RSSD |
| CN112863667B (en) * | 2021-01-22 | 2023-10-24 | 杭州电子科技大学 | Lung sound diagnostic device based on deep learning |
| CN113095137B (en) * | 2021-03-10 | 2023-12-29 | 南京大学 | Signal characteristic recognition device and method based on machine learning |
| CN112994740B (en) * | 2021-04-23 | 2021-07-23 | 成都天锐星通科技有限公司 | Frequency hopping signal parameter estimation method and device, electronic equipment and readable storage medium |
| CN113253682B (en) * | 2021-05-17 | 2022-04-29 | 中国石油大学(华东) | Nonlinear chemical process fault detection method |
| CN113643722B (en) * | 2021-08-27 | 2024-04-19 | 杭州电子科技大学 | Urban noise identification method based on multilayer matrix random neural network |
| CN115013298B (en) * | 2022-06-22 | 2023-03-31 | 浙江石水泵业科技有限公司 | Real-time performance online monitoring system and monitoring method of sewage pump |
| CN114861891A (en) * | 2022-07-05 | 2022-08-05 | 中国民航大学 | Improved sparse autoencoder-based situation feature extraction method for air traffic control system |
| CN115375210B (en) * | 2022-10-27 | 2023-01-24 | 安徽省中网云联科技有限公司 | Occupational skill level identification management method and system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1598569A1 (en) * | 2003-02-28 | 2005-11-23 | Thk Co., Ltd. | Condition-detecting device, method, and program, and information-recording medium |
| WO2007099730A1 (en) * | 2006-02-28 | 2007-09-07 | Thk Co., Ltd. | State detection device, state detection method, state detection program, and information recording medium |
| CN104156736A (en) * | 2014-09-05 | 2014-11-19 | 西安电子科技大学 | Polarized SAR image classification method on basis of SAE and IDL |
| CN104655425A (en) * | 2015-03-06 | 2015-05-27 | 重庆大学 | Bearing fault classification diagnosis method based on sparse representation and LDM (large margin distribution machine) |
-
2015
- 2015-04-10 CN CN201510169697.XA patent/CN104819846B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1598569A1 (en) * | 2003-02-28 | 2005-11-23 | Thk Co., Ltd. | Condition-detecting device, method, and program, and information-recording medium |
| WO2007099730A1 (en) * | 2006-02-28 | 2007-09-07 | Thk Co., Ltd. | State detection device, state detection method, state detection program, and information recording medium |
| CN104156736A (en) * | 2014-09-05 | 2014-11-19 | 西安电子科技大学 | Polarized SAR image classification method on basis of SAE and IDL |
| CN104655425A (en) * | 2015-03-06 | 2015-05-27 | 重庆大学 | Bearing fault classification diagnosis method based on sparse representation and LDM (large margin distribution machine) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023099849A1 (en) * | 2021-12-02 | 2023-06-08 | Safran Aircraft Engines | Method for automatically diagnosing a part |
| FR3130028A1 (en) * | 2021-12-02 | 2023-06-09 | Safran Aircraft Engines | Automatic part diagnostic method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104819846A (en) | 2015-08-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104819846B (en) | Rolling bearing sound signal fault diagnosis method based on short-time Fourier transform and sparse laminated automatic encoder | |
| CN112149316B (en) | Aero-engine residual life prediction method based on improved CNN model | |
| Shao et al. | Modified stacked autoencoder using adaptive Morlet wavelet for intelligent fault diagnosis of rotating machinery | |
| CN110702411B (en) | Residual error network rolling bearing fault diagnosis method based on time-frequency analysis | |
| CN110261109B (en) | Rolling bearing fault diagnosis method based on bidirectional memory cyclic neural network | |
| CN108256556A (en) | Wind-driven generator group wheel box method for diagnosing faults based on depth belief network | |
| CN109632309A (en) | Based on the rolling bearing fault intelligent diagnosing method for improving S-transformation and deep learning | |
| Abdul et al. | A hybrid temporal feature for gear fault diagnosis using the long short term memory | |
| CN107702922B (en) | Rolling bearing fault diagnosis method based on LCD and stacking automatic encoder | |
| CN110795843B (en) | Method and device for identifying faults of rolling bearing | |
| CN112362756B (en) | Concrete structure damage monitoring method and system based on deep learning | |
| CN113128338B (en) | Intelligent diagnosis method for printer roller faults under small sample | |
| CN106326929A (en) | Rail train running part rolling bearing fault diagnosis method | |
| CN104330255A (en) | Gear fault diagnosing method based on multi-sensor information fusion | |
| CN105424366A (en) | Bearing fault diagnosis method based on EEMD adaptive denoising | |
| CN112329914B (en) | Fault diagnosis method and device for buried transformer substation and electronic equipment | |
| CN105678343A (en) | Adaptive-weighted-group-sparse-representation-based diagnosis method for noise abnormity of hydroelectric generating set | |
| CN113314144A (en) | Voice recognition and power equipment fault early warning method, system, terminal and medium | |
| Xu et al. | Deep coupled visual perceptual networks for motor fault diagnosis under nonstationary conditions | |
| CN105448291A (en) | Parkinsonism detection method and detection system based on voice | |
| CN111275108A (en) | Method for performing sample expansion on partial discharge data based on generation countermeasure network | |
| CN110455530A (en) | Compose the gear case of blower combined failure diagnostic method of kurtosis combination convolutional neural networks | |
| CN106053074A (en) | Rolling bearing sound signal fault feature extraction method based on STFT and rotation inertia entropy | |
| CN112364706A (en) | Small sample bearing fault diagnosis method based on class imbalance | |
| CN112160877A (en) | Fan bearing fault diagnosis method based on SFA and CNN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170322 |