Main Content

emd

Empirical mode decomposition

Description

[imf,residual] = emd(x) returns intrinsic mode functions imf and residual signal residual corresponding to the empirical mode decomposition of x. Use emd to decompose and simplify complicated signals into a finite number of intrinsic mode functions required to perform Hilbert spectral analysis.

example

[imf,residual,info] = emd(x) returns additional information info on IMFs and residual signal for diagnostic purposes.

example

[___] = emd(___,Name,Value) performs the empirical mode decomposition with additional options specified by one or more Name,Value pair arguments.

example

emd(___) plots the original signal, IMFs, and residual signal as subplots in the same figure.

example

Examples

collapse all

Load and visualize a nonstationary continuous signal composed of sinusoidal waves with a distinct change in frequency. The vibration of a jackhammer and the sound of fireworks are examples of nonstationary continuous signals. The signal is sampled at a rate fs.

load("sinusoidalSignalExampleData.mat","X","fs")
t = (0:length(X)-1)/fs;

plot(t,X)
xlabel("Time (s)")

Figure contains an axes object. The axes object with xlabel Time (s) contains an object of type line.

The mixed signal contains sinusoidal waves with different amplitude and frequency values.

To create the Hilbert spectrum plot, you need the intrinsic mode functions (IMFs) of the signal. Perform empirical mode decomposition to compute the IMFs and residuals of the signal. Since the signal is not smooth, specify 'pchip' as the interpolation method.

[imf,residual,info] = emd(X,Interpolation="pchip");

The table generated in the command window indicates the number of sift iterations, the relative tolerance, and the sift stop criterion for each generated IMF. This information is also contained in info. You can hide the table by adding the 'Display',0 name value pair.

Create the Hilbert spectrum plot using the imf components obtained using empirical mode decomposition.

hht(imf,fs)

Figure contains an axes object. The axes object with title Hilbert Spectrum, xlabel Time (s), ylabel Frequency (Hz) contains 9 objects of type patch.

The frequency versus time plot is a sparse plot with a vertical color bar indicating the instantaneous energy at each point in the IMF. The plot represents the instantaneous frequency spectrum of each component decomposed from the original mixed signal. Three IMFs appear in the plot with a distinct change in frequency at 1 second.

This trigonometric identity presents two different views of the same physical signal:

52cos2πf1t+14(cos2π(f1+f2)t+cos2π(f1-f2)t)=(2+cos2πf2t)cos2πf1t.

Generate two sinusoids, s and z, such that s is the sum of three sine waves and z is a single sine wave with a modulated amplitude. Verify that the two signals are equal by calculating the infinity norm of their difference.

t = 0:1e-3:10;
omega1 = 2*pi*100;
omega2 = 2*pi*20;
s = 0.25*cos((omega1-omega2)*t) + ...
    2.5*cos(omega1*t) + ...
    0.25*cos((omega1+omega2)*t);
z = (2+cos(omega2/2*t).^2).*cos(omega1*t);

norm(s-z,Inf) 
ans = 
3.2729e-13

Plot the sinusoids and select a 1-second interval starting at 2 seconds.

plot(t,[s' z'])
xlim([2 3])
xlabel('Time (s)')
ylabel('Signal')

Figure contains an axes object. The axes object with xlabel Time (s), ylabel Signal contains 2 objects of type line.

Obtain the spectrogram of the signal. The spectrogram shows three distinct sinusoidal components. Fourier analysis sees the signals as a superposition of sine waves.

pspectrum(s,1000,'spectrogram','TimeResolution',4)

Figure contains an axes object. The axes object with title Fres = 3.9101 Hz, Tres = 4 s, xlabel Time (s), ylabel Frequency (Hz) contains an object of type image.

Use emd to compute the intrinsic mode functions (IMFs) of the signal and additional diagnostic information. The function by default outputs a table that indicates the number of sifting iterations, the relative tolerance, and the sifting stop criterion for each IMF. Empirical mode decomposition sees the signal as z.

[imf,~,info] = emd(s);

The number of zero crossings and local extrema differ by at most one. This satisfies the necessary condition for the signal to be an IMF.

info.NumZerocrossing - info.NumExtrema
ans = 
1

Plot the IMF and select a 0.5-second interval starting at 2 seconds. The IMF is an AM signal because emd views the signal as amplitude modulated.

plot(t,imf)
xlim([2 2.5])
xlabel('Time (s)')
ylabel('IMF')

Figure contains an axes object. The axes object with xlabel Time (s), ylabel IMF contains an object of type line.

Simulate a vibration signal from a damaged bearing. Perform empirical mode decomposition to visualize the IMFs of the signal and look for defects.

A bearing with a pitch diameter of 12 cm has eight rolling elements. Each rolling element has a diameter of 2 cm. The outer race remains stationary as the inner race is driven at 25 cycles per second. An accelerometer samples the bearing vibrations at 10 kHz.

fs = 10000;
f0 = 25;
n = 8;
d = 0.02;
p = 0.12;

Schematic of a damaged bearing with eight rolling element. On the left, the bearing unit. On the right, the inner race and the outer race of a rolling element.

The vibration signal from the healthy bearing includes several orders of the driving frequency.

t = 0:1/fs:10-1/fs;
a = 0.2*[1 0.5 0.2 0.1 0.05];
f = f0*[1 2 3 4 5];
yHealthy = a*sin(2*pi*f'.*t);

A resonance is excited in the bearing vibration halfway through the measurement process.

yHealthy = (1+1./(1+linspace(-10,10,length(yHealthy)).^4)).*yHealthy;

The resonance introduces a defect in the outer race of the bearing that results in progressive wear. The defect causes a series of impacts that recur at the ball pass frequency outer race (BPFO) of the bearing:

BPFO=12nf0[1-dpcosθ],

where f0 is the driving rate, n is the number of rolling elements, d is the diameter of the rolling elements, p is the pitch diameter of the bearing, and θ is the bearing contact angle. Assume a contact angle of 15° and compute the BPFO.

ca = 15;
bpfo = n*f0/2*(1-d/p*cosd(ca));

Use the pulstran function to model the impacts as a periodic train of 5-millisecond sinusoids. Each 3 kHz sinusoid is windowed by a flat top window. Use a power law to introduce progressive wear in the bearing vibration signal.

fImpact = 3000;
tImpact = 0:1/fs:5e-3-1/fs;
wImpact = flattopwin(length(tImpact))'/10;
xImpact = sin(2*pi*fImpact*tImpact).*wImpact;

tx = 0:1/bpfo:t(end);
tx = [tx; 1.3.^tx-2];

nWear = 49000;
nSamples = 100000;
yImpact = pulstran(t,tx',xImpact,fs)/5;
yImpact = [zeros(1,nWear) yImpact(1,(nWear+1):nSamples)];

Generate the BPFO vibration signal by adding the impacts to the healthy signal. Plot the signal and select a 0.3-second interval starting at 5.0 seconds.

yBPFO = yImpact + yHealthy;

xLimLeft = 5.0;
xLimRight = 5.3;
yMin = -0.6;
yMax = 0.6;

plot(t,yBPFO)

hold on
[limLeft,limRight] = meshgrid([xLimLeft xLimRight],[yMin yMax]);
plot(limLeft,limRight,"--")
hold off
xlabel("Time (seconds)")
ylabel("Amplitude")

Figure contains an axes object. The axes object with xlabel Time (seconds), ylabel Amplitude contains 3 objects of type line.

Zoom in on the selected interval to visualize the effect of the impacts.

xlim([xLimLeft xLimRight])

Figure contains an axes object. The axes object with xlabel Time (seconds), ylabel Amplitude contains 3 objects of type line.

Add white Gaussian noise to the signals. Specify a noise variance of 1/1502.

rn = 150;
yGood = yHealthy + randn(size(yHealthy))/rn;
yBad = yBPFO + randn(size(yHealthy))/rn;

plot(t,yGood,t,yBad)
xlim([xLimLeft xLimRight])
legend("Healthy","Damaged")
xlabel("Time (seconds)")
ylabel("Amplitude")

Figure contains an axes object. The axes object with xlabel Time (seconds), ylabel Amplitude contains 2 objects of type line. These objects represent Healthy, Damaged.

Use emd to perform an empirical mode decomposition of the healthy bearing signal. Compute the first five intrinsic mode functions (IMFs). Use the Display name-value argument to show a table with the number of sifting iterations, the relative tolerance, and the sifting stop criterion for each IMF.

imfGood = emd(yGood,MaxNumIMF=5,Display=1);
Current IMF  |  #Sift Iter  |  Relative Tol  |  Stop Criterion Hit  
      1      |        3     |     0.017132   |  SiftMaxRelativeTolerance
      2      |        3     |      0.12694   |  SiftMaxRelativeTolerance
      3      |        6     |      0.14582   |  SiftMaxRelativeTolerance
      4      |        1     |     0.011082   |  SiftMaxRelativeTolerance
      5      |        2     |      0.03463   |  SiftMaxRelativeTolerance
Decomposition stopped because maximum number of intrinsic mode functions was extracted.

Use emd without output arguments to visualize the first three modes and the residual.

emd(yGood,MaxNumIMF=5)

Figure contains 5 axes objects. Axes object 1 with ylabel Signal contains an object of type line. This object represents data. Axes object 2 with ylabel IMF 1 contains an object of type line. This object represents data. Axes object 3 with ylabel IMF 2 contains an object of type line. This object represents data. Axes object 4 with ylabel IMF 3 contains an object of type line. This object represents data. Axes object 5 with ylabel Residual contains an object of type line. This object represents data.

Compute and visualize the IMFs of the defective bearing signal. The first empirical mode reveals the high-frequency impacts. This high-frequency mode increases in energy as the wear progresses. The third mode shows the resonance in the vibration signal.

imfBad = emd(yBad,MaxNumIMF=5,Display=1);
Current IMF  |  #Sift Iter  |  Relative Tol  |  Stop Criterion Hit  
      1      |        2     |     0.041274   |  SiftMaxRelativeTolerance
      2      |        3     |      0.16695   |  SiftMaxRelativeTolerance
      3      |        3     |      0.18428   |  SiftMaxRelativeTolerance
      4      |        1     |     0.037177   |  SiftMaxRelativeTolerance
      5      |        2     |     0.095861   |  SiftMaxRelativeTolerance
Decomposition stopped because maximum number of intrinsic mode functions was extracted.
emd(yBad,MaxNumIMF=5)

Figure contains 5 axes objects. Axes object 1 with ylabel Signal contains an object of type line. This object represents data. Axes object 2 with ylabel IMF 1 contains an object of type line. This object represents data. Axes object 3 with ylabel IMF 2 contains an object of type line. This object represents data. Axes object 4 with ylabel IMF 3 contains an object of type line. This object represents data. Axes object 5 with ylabel Residual contains an object of type line. This object represents data.

The next step in the analysis is to compute the Hilbert spectrum of the extracted IMFs. For more details, see the Compute Hilbert Spectrum of Vibration Signal example.

Load and visualize a nonstationary continuous signal composed of sinusoidal waves with a distinct change in frequency. The vibration of a jackhammer and the sound of fireworks are examples of nonstationary continuous signals. The signal is sampled at a rate fs.

load("sinusoidalSignalExampleData.mat","X","fs")
t = (0:length(X)-1)/fs;
plot(t,X)
xlabel("Time(s)")

Figure contains an axes object. The axes object with xlabel Time(s) contains an object of type line.

The mixed signal contains sinusoidal waves with different amplitude and frequency values.

Perform empirical mode decomposition to plot the intrinsic mode functions and residual of the signal. Since the signal is not smooth, specify "pchip" as the interpolation method.

emd(X,Interpolation="pchip",Display=1)
Current IMF  |  #Sift Iter  |  Relative Tol  |  Stop Criterion Hit  
      1      |        2     |     0.026352   |  SiftMaxRelativeTolerance
      2      |        2     |    0.0039573   |  SiftMaxRelativeTolerance
      3      |        1     |     0.024838   |  SiftMaxRelativeTolerance
      4      |        2     |      0.05929   |  SiftMaxRelativeTolerance
      5      |        2     |      0.11317   |  SiftMaxRelativeTolerance
      6      |        2     |      0.12599   |  SiftMaxRelativeTolerance
      7      |        2     |      0.13802   |  SiftMaxRelativeTolerance
      8      |        3     |      0.15937   |  SiftMaxRelativeTolerance
      9      |        2     |      0.15923   |  SiftMaxRelativeTolerance
Decomposition stopped because the number of extrema in the residual signal is less than the 'MaxNumExtrema' value.

Figure contains 5 axes objects. Axes object 1 with ylabel Signal contains an object of type line. This object represents data. Axes object 2 with ylabel IMF 1 contains an object of type line. This object represents data. Axes object 3 with ylabel IMF 2 contains an object of type line. This object represents data. Axes object 4 with ylabel IMF 3 contains an object of type line. This object represents data. Axes object 5 with ylabel Residual contains an object of type line. This object represents data.

emd generates an interactive figure with the original signal, the first 3 IMFs, and the residual. The table generated in the command window indicates the number of sift iterations, the relative tolerance, and the sift stop criterion for each generated IMF. You can hide the table by removing the Display name-value argument or specifying it as 0.

Right-click on the white space in the figure to open the IMF selector window. Use IMF selector to selectively view the generated IMFs, the original signal, and the residual.

When you right-click on the white space in the Empirical Mode Decomposition figure, these options appear: IMF Selector, Characteristics, Grid, and Properties.

Select the IMFs to be displayed from the list. Choose whether to display the original signal and residual on the figure.

IMF Selector window. You can select "Show Signal", "Show residual", and the IMFs that you want to plot from the list displayed.

The selected IMFs are now displayed on the figure.

Empirical Mode Decomposition figure. The following plots appear from top to bottom: Signal, IMF 4, IMF 5, IMF 6, and Residual.

Use the figure to visualize individual components decomposed from the original signal along with the residual. Note that the residual is computed for the total number of IMFs, and does not change based on the IMFs selected in the IMF selector window.

Input Arguments

collapse all

Time-domain signal, specified as a real-valued vector, or a single-variable timetable with a single column. If x is a timetable, x must contain increasing, finite row times.

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Before R2021a, use commas to separate each name and value, and enclose Name in quotes.

Example: 'MaxNumIMF',5

Cauchy-type convergence criterion, specified as the comma-separated pair consisting of 'SiftRelativeTolerance' and a positive scalar. SiftRelativeTolerance is one of the sifting stop criteria, that is, sifting stops when the current relative tolerance is less than SiftRelativeTolerance. For more information, see Sift Relative Tolerance.

Maximum number of sifting iterations, specified as the comma-separated pair consisting of 'SiftMaxIterations' and a positive scalar integer. SiftMaxIterations is one of the sifting stop criteria, that is, sifting stops when the current number of iterations is larger than SiftMaxIterations.

SiftMaxIterations can be specified using only positive whole numbers.

Maximum number of IMFs extracted, specified as the comma-separated pair consisting of 'MaxNumIMF' and a positive scalar integer. MaxNumIMF is one of the decomposition stop criteria, that is, decomposition stops when number of IMFs generated is equal to MaxNumIMF.

MaxNumIMF can be specified using only positive whole numbers.

Maximum number of extrema in the residual signal, specified as the comma-separated pair consisting of 'MaxNumExtrema' and a positive scalar integer. MaxNumExtrema is one of the decomposition stop criteria, that is, decomposition stops when number of extrema is less than MaxNumExtrema.

MaxNumExtrema can be specified using only positive whole numbers.

Signal to residual energy ratio, specified as the comma-separated pair consisting of 'MaxEnergyRatio' and a scalar. MaxEnergyRatio is the ratio of the energy of the signal at the beginning of sifting and the average envelope energy. MaxEnergyRatio is one of the decomposition stop criteria, that is, decomposition stops when current energy ratio is larger than MaxEnergyRatio. For more information, see Energy Ratio.

Interpolation method for envelope construction, specified as the comma-separated pair consisting of 'Interpolation' and either 'spline' or 'pchip'.

Specify Interpolation as:

  • 'spline', if x is a smooth signal

  • 'pchip', if x is a nonsmooth signal

'spline' interpolation method uses cubic splines, while 'pchip' uses piecewise-cubic Hermite interpolating polynomials.

Toggle information display in the command window, specified as the comma-separated pair consisting of 'Display' and either 0 or 1. The table generated in the command window indicates the number of sift iterations, the relative tolerance, and the sift stop criterion for each generated IMF. Specify Display as 1 to show the table or 0 to hide the table.

Output Arguments

collapse all

Intrinsic mode function (IMF), returned as a matrix or timetable. Each IMF is an amplitude and frequency modulated signal with positive and slowly varying envelopes. To perform spectral analysis of a signal, you can apply the Hilbert-Huang transform to its IMFs. See hht and Intrinsic Mode Functions.

imf is returned as:

  • A matrix whose each column is an imf, when x is a vector

  • A timetable, when x is a single data column timetable

Residual of the signal, returned as a column vector or a single data column timetable. residual represents the portion of the original signal x not decomposed by emd.

residual is returned as:

  • A column vector, when x is a vector.

  • A single data column timetable, when x is a single data column timetable.

Additional information for diagnostics, returned as a structure with the following fields:

  • NumIMF — Number of IMFs extracted

    NumIMF is a vector from 1 to N, where N is the number of IMFs. If no IMFs are extracted, NumIMF is empty.

  • NumExtrema — Number of extrema in each IMF

    NumExtrema is a vector equal in length to the number of IMFs. The kth element of NumExtrema is the number of extrema found in the kth IMF. If no IMFs are extracted, NumExtrema is empty.

  • NumZerocrossing — Number of zero crossings in each IMF

    Number of zero crossings in each IMF. NumZerocrossing is a vector equal in length to the number of IMFs. The kth element of NumZerocrossing is the number of zero crossings in the kth IMF. If no IMFs are extracted, NumZerocrossing is empty.

  • NumSifting — Number of sifting iterations used to extract each IMF

    NumSifting is a vector equal in length to the number of IMFs. The kth element of NumSifting is the number of sifting iterations used in the extraction of the kth IMF. If no IMFs are extracted, NumSifting is empty.

  • MeanEnvelopeEnergy — Energy of the mean of the upper and lower envelopes obtained for each IMF

    If UE is the upper envelope and LE is the lower envelope, MeanEnvelopeEnergy is mean(((LE+UL)/2).^2). MeanEnvelopeEnergy is a vector equal in length to the number of IMFs. The kth element of MeanEnvelopeEnergy is the mean envelope energy for the kth IMF. If no IMFs are extracted, MeanEnvelopeEnergy is empty.

  • RelativeTolerance — Final relative tolerance of the residual for each IMF

    The relative tolerance is defined as the ratio of the squared 2-norm of the difference between the residual from the previous sifting step and the residual from the current sifting step to the squared 2-norm of the residual from the ith sifting step. The sifting process stops when RelativeTolerance is less than SiftRelativeTolerance. For additional information, see Sift Relative Tolerance. RelativeTolerance is a vector equal in length to the number of IMFs. The kth element of RelativeTolerance is the final relative tolerance obtained for the kth IMF. If no IMFs are extracted, RelativeTolerance is empty.

More About

collapse all

Empirical Mode Decomposition

The empirical mode decomposition (EMD) algorithm decomposes a signal x(t) into intrinsic mode functions (IMFs) and a residual in an iterative process. The core component of the algorithm involves sifting a function x(t) to obtain a new function Y(t):

  • First find the local minima and maxima of x(t).

  • Then use the local extrema to construct lower and upper envelopes s(t) and s+(t), respectively, of x(t). Form the mean of the envelopes, m(t).

  • Subtract the mean from x(t) to obtain the residual: Y(t) = x(t) − m(t).

An overview of the decomposition is as follows:

  1. To begin, let r0(t) = x(t), where x(t) is the initial signal, and let i = 0.

  2. Before sifting, check ri(t):

    1. Find the total number (TN) of local extrema of ri(t).

    2. Find the energy ratio (ER) of ri(t) (see Energy Ratio).

  3. If (ER > MaxEnergyRatio) or (TN < MaxNumExtrema) or (number of IMFs > MaxNumIMF) then stop the decomposition.

  4. Let ri,Prev(t) = ri(t).

  5. Sift ri,Prev(t) to obtain ri,Cur(t).

  6. Check ri,Cur(t)

    1. Find the relative tolerance (RT) of ri,Cur(t) (see Sift Relative Tolerance).

    2. Get current sift iteration number (IN).

  7. If (RT < SiftRelativeTolerance) or (IN > SiftMaxIterations) then stop sifting. An IMF has been found: IMFi(t) = ri,Cur(t). Otherwise, let ri,Prev(t) = ri,Cur(t) and go to Step 5.

  8. Let ri+1(t) = ri(t) − ri,Cur(t).

  9. Let i = i + 1. Return to Step 2.

For additional information, see [1] and [3].

Intrinsic Mode Functions

The EMD algorithm decomposes, via an iterative sifting process, a signal x(t) into IMFs imfi(t) and a residual rN(t):

X(t)=i=1NIMFi(t)+rN(t)

When first introduced by Huang et al. [1], an IMF was defined to be a function with two characteristics:

  • The number of local extrema — the total number of local minima and local maxima — and the number of zero crossings differ by at most one.

  • The mean value of the upper and lower envelopes constructed from the local extrema is zero.

However, as noted in [4], sifting until a strict IMF is obtained can result in IMFs that have no physical significance. Specifically, sifting until the number of zero crossings and local extrema differ by at most one can result in pure-tone like IMFs, in other words, functions very similar to what would be obtained by projection on the Fourier basis. This situation is precisely what EMD strives to avoid, preferring AM-FM modulated components for their physical significance.

Reference [4] proposes options to obtain physically meaningful results. The emd function relaxes the original IMF definition by using Sift Relative Tolerance, a Cauchy-type stop criterion. The emd function iterates to extract natural AM-FM modes. The IMFs generated may fail to satisfy the local extrema-zero crossings criteria. See Zero Crossings and Extrema in Intrinsic Mode Function of Sinusoid.

Sift Relative Tolerance

Sift Relative Tolerance is a Cauchy-type stop criterion proposed in [4]. Sifting stops when current relative tolerance is less than SiftRelativeTolerance. The current relative tolerance is defined as

Relative Tolerancerprev(t)rcur(t)22rprev(t)22.

Because the Cauchy criterion does not directly count the number of zero crossings and local extrema, it is possible that the IMFs returned by the decomposition do not satisfy the strict definition of an intrinsic mode function. In those cases, you can try reducing the value of the SiftRelativeTolerance from its default value. See [4] for a detailed discussion of stopping criteria. The reference also discusses the advantages and disadvantages of insisting on strictly defined IMFs in empirical mode decomposition.

Energy Ratio

Energy ratio is the ratio of the energy of the signal at the beginning of sifting and the average envelope energy [2]. Decomposition stops when current energy ratio is larger than MaxEnergyRatio. For the ith IMF, the energy ratio is defined as

Energy Ratio10log10(X(t)2ri(t)2).

References

[1] Huang, Norden E., Zheng Shen, Steven R. Long, Manli C. Wu, Hsing H. Shih, Quanan Zheng, Nai-Chyuan Yen, Chi Chao Tung, and Henry H. Liu. “The Empirical Mode Decomposition and the Hilbert Spectrum for Nonlinear and Non-Stationary Time Series Analysis.” Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences 454, no. 1971 (March 8, 1998): 903–95. https://doi.org/10.1098/rspa.1998.0193.

[2] Rato, R.T., M.D. Ortigueira, and A.G. Batista. “On the HHT, Its Problems, and Some Solutions.” Mechanical Systems and Signal Processing 22, no. 6 (August 2008): 1374–94. https://doi.org/10.1016/j.ymssp.2007.11.028.

[3] Rilling, Gabriel, Patrick Flandrin, and Paulo Gonçalves. "On Empirical Mode Decomposition and Its Algorithms." IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing 2003. NSIP-03. Grado, Italy. 8–11.

[4] Wang, Gang, Xian-Yao Chen, Fang-Li Qiao, Zhaohua Wu, and Norden E. Huang. “On Intrinsic Mode Function.” Advances in Adaptive Data Analysis 02, no. 03 (July 2010): 277–93. https://doi.org/10.1142/S1793536910000549.

Extended Capabilities

Version History

Introduced in R2018a

See Also

Apps

Functions