CN109919082A - Modal identification method based on LSTM and EMD - Google Patents
Modal identification method based on LSTM and EMD Download PDFInfo
- Publication number
- CN109919082A CN109919082A CN201910165765.3A CN201910165765A CN109919082A CN 109919082 A CN109919082 A CN 109919082A CN 201910165765 A CN201910165765 A CN 201910165765A CN 109919082 A CN109919082 A CN 109919082A
- Authority
- CN
- China
- Prior art keywords
- lstm
- signal
- emd
- identification method
- indicate
- 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
Landscapes
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a kind of modal identification method based on LSTM and EMD, include the following steps: to obtain signal data;Prediction, which is carried out, using left and right ends data of the LSTM algorithm to signal prolongs Shen, the signal data after being expanded;The signal data after above-mentioned expansion is decomposed using EMD algorithm, obtains each rank mode signals of signal;Each rank mode signals are truncated at the left and right endpoint of original signal, form final each rank mode signals.The present invention can efficiently solve end effect and modal overlap effect of the traditional Empirical mode decomposition (EMD) in signal decomposition, to keep the result of non-destructive tests, mechanical fault diagnosis more accurate, robustness is stronger.
Description
Technical field
The present invention relates to artificial intelligence and signal processing technology field, especially a kind of mode based on LSTM and EMD is known
Other method.
Background technique
Engineering practice and research in the fields such as monitoring structural health conditions and non-destructive tests, mechanical fault diagnosis and Meteorological Science
In, it is required to that a large amount of signals are handled and described;Commonly used approach is Hilbert-Huang transformation (HHT) at present.
HHT is a kind of Time-Frequency Analysis Method, and signal is first carried out empirical mode decomposition (EMD) by this method, and generating one group has different spies
The intrinsic mode functions (IMF) for levying time scale, then make Hilbert transformation to each IMF again respectively, obtain Hilbert spectrum,
The spectrum can relatively accurately describe the energy of signal at any time with the changing rule of frequency.
In HHT method, EMD extracts IMF by multiple screening process one by one, will basis in screening each time
The Local Extremum of signal is fitted upper and lower envelope using cubic spline curve respectively.But due to the limited length of signal, end
Extreme value except point can not determine, therefore when with spline curve fitting, envelope usually will appear amplitude very near endpoint
Big phenomenon, referred to as end effect.For the high-frequency I MF component of signal, time scale is small, and distance is small between extreme point, endpoint
Effect is limited to the lesser part in signal both ends;And for low frequency IM F component, time scale is big, the distance between extreme point
Greatly, end effect can travel to the inside of signal.With the continuous screening of EMD, end effect is just constantly inwardly spread, especially
It is that can seriously affect the quality of IMF when signal is shorter, so that decomposition result is without actual physical significance;Therefore, it is necessary to grind
Suitable method is studied carefully to inhibit end Divergent Phenomenon when spline-fit, while avoiding the end of distortion signal special as far as possible again
Sign.
Summary of the invention
Goal of the invention: the end effect in order to overcome the problems, such as HHT and EMD algorithm in the prior art, the present invention provides one
Modal identification method of the kind based on LSTM and EMD can efficiently solve the end effect and modal overlap effect of signal decomposition
It answers, to keep the result of non-destructive tests, mechanical fault diagnosis more accurate, robustness is stronger.
Technical solution: in order to solve the above technical problems, of the present invention be based on LSTM long memory network (Long in short-term
Short Term Memory Network, LSTM) with the modal identification method of EMD, characterized by the following steps:
S1: measurement obtains object vibration Acceleration time course data-signal;
S2: denoising is carried out to resulting vibration acceleration time course data signal;
S3: left and right continuation is carried out to the signal after denoising in S2 using LSTM algorithm, obtains continuation signal;
S4: the continuation signal in S3 is decomposed using EMD algorithm, obtains each rank mode signals;
S5: it will be truncated, held at the left and right ends of the signal obtained in S1 step of each rank mode signals obtained in S4
The smooth each rank mode signals of point.
Further, in step S1, object vibration Acceleration time course data-signal is surveyed using acceleration transducer
Amount.
Further, in step S2, using mean filter method, denoising is carried out to signal.
Further, in step S3, LSTM algorithm expands the signal that S2 is obtained, and each LSTM computing unit uses input gate
Limit forgets thresholding, output thresholding.
Further, in step S3, the output valve of each LSTM computing unit calculates as follows:
ft=σ (Wf·[ht-1, xt]+bf)
it=σ (Wi·[ht-1, xt]+bi)
ot=σ (Wo[ht-1, xt]+bo)
ht=ot*tanh(Ct)
Wherein, ftIt indicates to forget thresholding, itIndicate input threshold,Indicate the state of previous moment neuron, CtIndicate existing
In the state of moment neuron, Ct-1It indicates to calculate ht-1Median in the process, otIndicate output thresholding, htIndicate current time
The output of unit, ht-1Indicate the output of previous moment unit, xtIndicate the input at current time, WfFor the weight square propagated forward
Battle array, WiFor the weight matrix of backpropagation, bfFor the bias matrix propagated forward, biFor the bias matrix of backpropagation, Wf、Wi、
bf、biUsing normal distribution as initialization value, WoFor the weight matrix currently exported, boFor the bias matrix currently exported, σ table
Show activation primitive, WcFor the weight matrix of Current neural member, bcFor the bias matrix of Current neural member.
Further, it in step S3, is updated using the feedback that error back propagation method carries out neuron, wherein criterion letter
Number selects error sum of squares loss (MSE), and general hidden layer excitation function uses Sigmoid function, and output layer excitation function is selected
SoftMax function.
Further, in step S3, LSTM prediction extension points select the 5% of original signal strength.
Further, it in step S4, in EMD decomposition step, is inserted when obtaining signal or more envelope using cubic spline
Value method.
The utility model has the advantages that compared with the prior art, the present invention has the following beneficial effects: LSTM is following after a kind of improvement
Ring neural network can solve the problem of traditional RNN can not handle the dependence of long range;LSTM passes through processing signal time sequence
The long temporal information of column is predicted to reject redundancy and short temporal information to the information that will have an impact of future, so as to benefit
Remove to predict the time sequential value of later a period of time with historical time sequence before;Signal left and right ends are prolonged using LSTM
It opens up, recycles EMD to decompose, can effectively inhibit the end effect of EMD method itself in this way, be able to solve Inclined Cable Vibration signal
Signal decomposition end effect problem during non-destructive tests, thus more accurately identification of damage.
Detailed description of the invention
Fig. 1 is the flow chart for the modal identification method based on LSTM and EMD that the present invention uses;
Fig. 2 is LSTM neural network cellular construction schematic diagram used in the present invention.
Specific embodiment
Technical solution of the present invention is described further with reference to the accompanying drawing.
As shown in Figure 1, being based on LSTM long memory network (Long Short Term Memory Network, LSTM) in short-term
With the modal identification method of EMD, include the following steps:
S1: object vibration Acceleration time course data-signal is measured using acceleration transducer;
S2: using the mean filter method in signal processing, obtained vibration acceleration time course data signal is carried out
Denoising;
S3: left and right continuation is carried out to the signal after denoising in S2 using LSTM algorithm, obtains continuation signal;
S4: continuation signal obtained in S3 is decomposed using EMD algorithm, obtains each rank mode signals;
S5: each rank mode signals obtained in S4 are truncated at the left and right ends of the S1 signal walked, obtain endpoint
More smooth each rank mode signals.
Wherein, in step S3, the signal that S2 is obtained is expanded using LSTM algorithm, each LSTM computing unit uses input gate
Limit forgets thresholding, output thresholding;The output valve of each LSTM computing unit calculates as follows:
ft=σ (Wf·[ht-1, xt]+bf)
it=σ (Wi·[ht-1, xt]+bi)
ot=σ (Wo[ht-1, xt]+bo)
ht=ot*tanh(Ct)
Wherein ftIt indicates to forget thresholding, itIndicate input threshold,Indicate the state of previous moment neuron, CtIndicate existing
In the state of moment neuron, Ct-1It indicates to calculate ht-1Median in the process, otIndicate output thresholding, htIndicate current time
The output of unit, ht-1Indicate the output of previous moment unit, xtIndicate the input at current time, WfFor the weight square propagated forward
Battle array, WiFor the weight matrix of backpropagation, bfFor the bias matrix propagated forward, biFor the bias matrix of backpropagation, Wf、Wi、
bf、biGenerally using normal distribution as initialization value, WoFor the weight matrix currently exported, boFor the bias matrix currently exported, σ
Indicate activation primitive, WcFor the weight matrix of Current neural member, bcFor the bias matrix of Current neural member.
It in step S3, is updated using the feedback that error back propagation method carries out neuron, wherein criterion function, which is selected, misses
Poor quadratic sum loses (MSE), and general hidden layer excitation function uses Sigmoid function, and output layer excitation function selects SoftMax letter
Number.In step S3, LSTM prediction extension points select the 5% of original signal strength.
In step S4, in EMD decomposition step, using cubic spline interpolation method to letter when obtaining signal or more envelope
Several two adjacent extreme points carry out the spline function of interpolation, as follows:
Wherein: interpolation characteristic, S (xi)=f (xi);
Batten is connected with each other, Si-1(xi)=Si(xi), i=1 ..., n-1;
It can continuously lead twice, S 'i-1(xi)=S 'i(xi) and S "i-1(xi)=S "i(xi), i=1 ..., n-1.
In step S5, signal obtained in S4 is truncated at the left and right ends of the S1 signal walked, obtain endpoint compared with
For smooth each rank mode signals.
Claims (8)
1. a kind of modal identification method based on LSTM and EMD, characterized by the following steps:
S1: measurement obtains object vibration Acceleration time course data-signal;
S2: denoising is carried out to resulting vibration acceleration time course data signal;
S3: left and right continuation is carried out to the signal after denoising in S2 using LSTM algorithm, obtains continuation signal;
S4: the continuation signal in S3 is decomposed using EMD algorithm, obtains each rank mode signals;
S5: each rank mode signals obtained in S4 are truncated at the left and right ends of the S1 signal walked, it is smooth to obtain endpoint
Each rank mode signals.
2. the modal identification method according to claim 1 based on LSTM and EMD, it is characterised in that: in step S1, utilize
Acceleration transducer measures object vibration Acceleration time course data-signal.
3. the modal identification method according to claim 1 based on LSTM and EMD, it is characterised in that: in step S2, utilize
Mean filter method carries out denoising to signal.
4. the modal identification method according to claim 1 based on LSTM and EMD, it is characterised in that: in step S3, utilize
LSTM algorithm expands the signal that S2 is obtained, and each LSTM computing unit is using input threshold, forgetting thresholding, output thresholding.
5. the modal identification method according to claim 4 based on LSTM and EMD, it is characterised in that: in step S3, each
The output valve of LSTM computing unit calculates as follows:
ft=σ (Wf·[ht-1, xt]+bf)
it=σ (Wi·[ht-1, xt]+bi)
ot=σ (Wo[ht-1, xt]+bo)
ht=ot*tanh(Ct)
Wherein, ftIt indicates to forget thresholding, itIndicate input threshold,Indicate the state of previous moment neuron, CtIndicate present tense
Carve the state of neuron, Ct-1It indicates to calculate ht-1Median in the process, otIndicate output thresholding, htIndicate current time unit
Output, ht-1Indicate the output of previous moment unit, xtIndicate the input at current time, WfFor the weight matrix propagated forward,
WiFor the weight matrix of backpropagation, bfFor the bias matrix propagated forward, biFor the bias matrix of backpropagation, Wf、Wi、bf、
biUsing normal distribution as initialization value, WoFor the weight matrix currently exported, boFor the bias matrix currently exported, σ indicates to swash
Function living, WcFor the weight matrix of Current neural member, bcFor the bias matrix of Current neural member.
6. the modal identification method according to claim 1 based on LSTM and EMD, it is characterised in that: in step S3, utilize
The feedback that error back propagation method carries out neuron updates, and wherein criterion function selects error sum of squares loss (MSE), implies
Layer excitation function uses Sigmoid function, and output layer excitation function selects SoftMax function.
7. the modal identification method according to claim 1 based on LSTM and EMD, it is characterised in that: in step S3, LSTM
Prediction extension points select the 5% of original signal strength.
8. the modal identification method according to claim 1 based on LSTM and EMD, it is characterised in that: in step S4,
In EMD decomposition step, cubic spline interpolation method is used when obtaining signal or more envelope.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910165765.3A CN109919082A (en) | 2019-03-05 | 2019-03-05 | Modal identification method based on LSTM and EMD |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910165765.3A CN109919082A (en) | 2019-03-05 | 2019-03-05 | Modal identification method based on LSTM and EMD |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109919082A true CN109919082A (en) | 2019-06-21 |
Family
ID=66963373
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910165765.3A Pending CN109919082A (en) | 2019-03-05 | 2019-03-05 | Modal identification method based on LSTM and EMD |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109919082A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110618451A (en) * | 2019-09-17 | 2019-12-27 | 太原理工大学 | Method for detecting seismic exploration weak signal based on NARX neural network |
CN110909931A (en) * | 2019-11-20 | 2020-03-24 | 成都理工大学 | Logging curve prediction method based on modal decomposition reconstruction and depth LSTM-RNN model |
CN111131424A (en) * | 2019-12-18 | 2020-05-08 | 武汉大学 | Service quality prediction method based on combination of EMD and multivariate LSTM |
CN111353482A (en) * | 2020-05-25 | 2020-06-30 | 天津开发区精诺瀚海数据科技有限公司 | LSTM-based fatigue factor recessive anomaly detection and fault diagnosis method |
CN113362853A (en) * | 2020-03-03 | 2021-09-07 | 东北大学秦皇岛分校 | EMD endpoint effect suppression method based on LSTM network |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103440226A (en) * | 2013-06-26 | 2013-12-11 | 燕山大学 | EMD (Empirical Mode Decomposition) endpoint effect suppression method based on HMM (Hidden Markov Model) correction and neural network extension |
-
2019
- 2019-03-05 CN CN201910165765.3A patent/CN109919082A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103440226A (en) * | 2013-06-26 | 2013-12-11 | 燕山大学 | EMD (Empirical Mode Decomposition) endpoint effect suppression method based on HMM (Hidden Markov Model) correction and neural network extension |
Non-Patent Citations (1)
Title |
---|
王鑫等: "基于 LSTM 循环神经网络的故障时间序列预测", 《北京航空航天大学学报》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110618451A (en) * | 2019-09-17 | 2019-12-27 | 太原理工大学 | Method for detecting seismic exploration weak signal based on NARX neural network |
CN110909931A (en) * | 2019-11-20 | 2020-03-24 | 成都理工大学 | Logging curve prediction method based on modal decomposition reconstruction and depth LSTM-RNN model |
CN111131424A (en) * | 2019-12-18 | 2020-05-08 | 武汉大学 | Service quality prediction method based on combination of EMD and multivariate LSTM |
CN111131424B (en) * | 2019-12-18 | 2020-12-18 | 武汉大学 | Service quality prediction method based on combination of EMD and multivariate LSTM |
CN113362853A (en) * | 2020-03-03 | 2021-09-07 | 东北大学秦皇岛分校 | EMD endpoint effect suppression method based on LSTM network |
CN111353482A (en) * | 2020-05-25 | 2020-06-30 | 天津开发区精诺瀚海数据科技有限公司 | LSTM-based fatigue factor recessive anomaly detection and fault diagnosis method |
CN111353482B (en) * | 2020-05-25 | 2020-12-08 | 天津开发区精诺瀚海数据科技有限公司 | LSTM-based fatigue factor recessive anomaly detection and fault diagnosis method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109919082A (en) | Modal identification method based on LSTM and EMD | |
CN108764540B (en) | Water supply network pressure prediction method based on parallel LSTM series DNN | |
CN111539132B (en) | Dynamic load time domain identification method based on convolutional neural network | |
CN106709640A (en) | Method for predicting ship energy consumption based on deep learning and support vector regression | |
CN111324990A (en) | Porosity prediction method based on multilayer long-short term memory neural network model | |
CN110096810B (en) | Industrial process soft measurement method based on layer-by-layer data expansion deep learning | |
CN109472397B (en) | Polymerization process parameter adjusting method based on viscosity change | |
CN111461187B (en) | Intelligent building settlement detection system | |
Ye et al. | Cascaded GMDH-wavelet-neuro-fuzzy network | |
US20040133616A1 (en) | Apparatus and method for efficient adaptation of finite element meshes for numerical solutions of partial differential equations | |
CN114117912A (en) | Sea clutter modeling and inhibiting method under data model dual drive | |
CN112001115A (en) | Soft measurement modeling method of semi-supervised dynamic soft measurement network | |
CN110222825B (en) | Cement product specific surface area prediction method and system | |
CN113468466B (en) | One-dimensional wave equation solving method based on neural network | |
CN113642591B (en) | Multi-beam submarine sediment layer type estimation method and system based on transfer learning | |
CN116542139A (en) | Method and device for predicting roughness of liquid jet polishing surface | |
CN116663428A (en) | Neural network-based face wave imaging joint inversion method | |
CN114942480B (en) | Ocean station wind speed forecasting method based on information perception attention dynamic cooperative network | |
CN108154273B (en) | Rainfall data estimation method and device based on neural network | |
CN113836818B (en) | Ocean current motion prediction algorithm based on BP neural network prediction model | |
CN113627074B (en) | Ground wave propagation delay prediction method based on transfer learning | |
CN111460738B (en) | RNN-ARX modeling method and RNN-ARX model of magnetic suspension system | |
CN112862173B (en) | Lake and reservoir cyanobacterial bloom prediction method based on self-organizing deep confidence echo state network | |
CN114814776A (en) | PD radar target detection method based on graph attention network and transfer learning | |
CN106777918A (en) | The inversion method of the function gradient structure material property based on guided wave and fuzzy algorithmic approach |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190621 |