You are on page 1of 12

Analytica Chimica Acta 544 (2005) 280–291

Caterpillar—an adaptive algorithm for detecting process changes


from acoustic emission signals
Geir Rune Flåtena , Ron Belchamberb , Mike Collinsb , Anthony D. Walmsleya,∗
a CPACT, Department of Chemistry, The University of Hull, Cottingham Road, Hull HU6 7RX, UK
b Process Analysis and Automation Ltd., Fernhill Road, Farnborough, Hampshire GU14 9RX, UK

Received 6 October 2004; received in revised form 16 December 2004; accepted 16 December 2004

Abstract

An adaptive algorithm for detecting process changes in a process monitored by acoustic emission is presented as an alternative to traditional
modelling techniques based on fixed or static models. This approach significantly reduces the need to remodel the process as operating
conditions change.
The central idea is that two moving windows are moved through the data side by side. The signal variation in one of them is modelled by
a principal component analysis (PCA) model, and the samples in the other window are compared to the critical borders of the PCA model.
Significant differences are interpreted as a process change, i.e. the acoustic emission from the process has changed. In this work acoustic
emission data from a fluidised bed is analysed. Optimal settings for the algorithm are proposed and the robustness towards noise and other
signal degradations is shown to be good. The algorithm seems to be batch independent, which means that regular re-calibration (blank runs)
which is needed by reference model approaches can be avoided.
© 2004 Elsevier B.V. All rights reserved.

Keywords: Acoustic emission; Fluidised bed; Adaptive; Process monitoring; Chemometrics; SIMCA; Caterpillar

1. Introduction ever, the acoustic emission spectra are less stable and have
lower resolution than optical spectra, and often frequencies
The feasibility of using acoustic emission as a process cannot be assigned to different functional groups. Moreover,
monitoring tool has been studied in several industries [1,2] acoustic spectra from industrial fluidised beds, reflect extrin-
such as wood [3], pharmaceutical [4], reaction monitoring [5] sic physical properties such as density, hardness, particle size
and petrochemical [6]. Acoustic emission is particularly an distribution of the fluidised particles as compared to optical
attractive tool for monitoring the performance of industrial spectra which (primarily) reflect intrinsic chemical proper-
fluidised beds which are used in a number of major processes ties e.g. functional groups, concentrations. Hence acoustic
including catalytic cracking of oil, production of polyethy- emission monitoring can be regarded as a complementary
lene and other important plastics and within the pharmaceu- technique to optical monitoring.
tical and food industries for granulation and drying of parti- The underlying motivation in many acoustic emission-
cles. The acoustic emissions are recorded and power spectra monitoring set ups is that a trained operator can hear when
are calculated by applying the Fourier transform to the raw the process is going out of specification or there is a change
data. The power spectra give the intensity of acoustic signals in state. Everyday examples include gear changes in cars and
at different frequencies analogously to an infrared spectrum washing machine cycles [7]. Acoustic emission monitoring is
where the absorbance at different wavelengths is given. How- thus a natural approach to measure and predict process upsets
or phase changes. Due to the generally large amount of data
∗ Corresponding author. in process monitoring chemometric techniques are necessary
E-mail address: a.d.walmsley@hull.ac.uk (A.D. Walmsley). in order to analyse and interpret the acoustic monitoring data

0003-2670/$ – see front matter © 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.aca.2004.12.043
G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291 281

[8]. SIMCA (Soft Independent Modelling of Class Analo- 2. Theory


gies) [9] is a widely used technique for classifying acoustic
emission signals. In SIMCA a reference state e.g. an end- 2.1. Caterpillar
point or a normal operation state, is modelled using a princi-
pal component. New samples are compared to this model in Caterpillar is a method, which uses the recent variation to
order to detect end-point conditions or any deviations from determine whether the now-variation is ‘normal’ or whether
the norm (upsets). A problem for the SIMCA approach is any changes have occurred in the process. In the case of
that physical changes in the process, such as changed parti- acoustic measurements this emulates the function of the op-
cle size distribution and processing conditions may require erator’s ears: can we hear any changes in our fluidised bed?
re-calculation of the reference model. The Caterpillar algorithm is outlined below while the cal-
The application of MSPC (Multivariate Statistical Process culation details are given in Sections 2.1.1–2.1.3:
Control) techniques [10,11] to acoustic monitoring data has
so far not been reported in the literature. However, analo- 1. Two moving windows of width w are placed side by side.
gously to the SIMCA approach these techniques rely on a Hence the total width is 2w samples. They are named the
steady state or reference state model and are thus sensitive ‘modelling window’ and the ‘prediction window’, and the
towards process changes. ‘prediction window’ is the rightmost, i.e. the most recent
Recently, a moving window approach named ‘Caterpil- measurements.
lar’ was suggested for analysing acoustic monitoring data 2. The systematic variation in the modelling window is rep-
for a fluidised bed reactor [6]. The general idea was to copy resented by a k component PCA model, and the critical
the operator’s method of detection of a process change from variation dcrit is calculated using the T2 statistics.
changes in the acoustic emission from the process by com- 3. The d values (T2 statistics) for the samples in the pre-
paring the now-variation with recent variation. Two windows diction window is calculated and compared to dcrit . The
side by side are moved step-wise through the data. One win- number of samples larger than the critical value is counted
dow is used as a reference window and a PCA model is used and the result is shown in the occurrences plot, which is
to describe the variation for the samples in this window. The explained in detail in Section 2.1.1.
samples in the other window, the prediction window, are com- 4. The windows are moved to time t + 1 (or t + a where a is
pared to the model for the reference or modelling window. If an integer) and the process is repeated from step 2. In the
several of the samples in the prediction window are signifi- case of an online set up the windows are moved for each
cantly different from the reference PCA model this is inter- new recording of data.
preted as a process change. Since both windows are moved The window sizes w and the number of components k are
through the data the reference model will adapt to any pro- determined before starting the Caterpillar monitoring. This
cess changes so in the case of a process with several phases can be done using historical data or knowledge of similar
the Caterpillar algorithm will detect the onset of every new systems as suggested in Section 2.1.2.
phase.
The adaptive nature of the Caterpillar algorithm also re-
2.1.1. PCA decomposition and T2 statistics
moves the need for periodic remodelling or recalibration of
Assume that X is the m × n acoustic data matrix containing
the system as the algorithm can adapt to time dependent
all the monitoring data describing any process, m the number
changes such as drift and offset. Obviously, SIMCA or other
of time points or measurements in the data and n is the num-
reference model approaches must be regularly recalibrated
ber of measured variables or frequencies. The modelling and
or updated. Easy maintenance is especially important in in-
prediction windows mentioned in the Caterpillar algorithm
dustrial applications as recalibration is time consuming and
outline (previous section) are then two w × n matrices, which
may be operationally impossible. The continuous update of
can be referred to by Xmod and Xpred , respectively. In the fur-
the reference model also makes the Caterpillar algorithm less
ther discussion, Xmod is assumed to be mean centred and
sensitive to normal process variation, which results in rela-
Xpred is assumed to be scaled accordingly.
tively few false alarms.
The k component PCA model can be written as:
In this work, the Caterpillar algorithm is applied for
analysing monitoring data from a laboratory scale flu- Xmod = Tmod Ptmod + E (1)
idised bed reactor where an agglomeration event is in-
duced. Agglomeration in industrial fluid bed reactors is an where Tmod is the w × k scores matrix, Ptmod is the k × n
undesired condition. Serious agglomeration incidents are loadings matrix (t indicates matrix transpose) and E is the
very costly to rectify and can result in considerable plant w × n residual matrix. As always the k scores are mutually
downtime. orthogonal, the k loadings are mutually orthonormal and the
The settings of the algorithm such as window sizes and residuals matrix contains the variation not explained by the
number of components in the reference model are discussed. PCA model.
Caterpillar’s robustness towards noise and other signal upsets The scores for the w samples in the prediction window are
such as drift and shift is also checked. calculated by projection onto the PCA model describing the
282 G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291

variation in the modelling window: used in the PCA modelling. However, as opposed to SIMCA
where the reference model is system specific the Caterpillar
Tpred = Xpred Pmod (2)
approach is also working when only experiences with similar
where Xpred is the w × n data matrix containing the w sam- systems are applied. This is due to the more adaptive nature
ples in the prediction window. Tpred is the w × k scores matrix of Caterpillar.
and Pmod is the n × k loadings matrix obtained from (1). If reference data are available the suggested approach is
The T2 statistics for sample i, di , are calculated as (remem- to analyse these data by the Caterpillar algorithm using dif-
ber that the scores are centered already) [11]: ferent window sizes and different number of components.
The number of atypical samples (d > dcrit ) in each predic-
ti tti tion window is counted. As the number of atypical samples
di =   (3)
1 t should be low (∼0) for the reference data, a combination
w−1 Tmod Tmod
of window sizes and number of components that gives few
where ti is the 1 × k score values for sample i and Tmod is atypical samples are chosen. An additional criterion is that
the w × k scores matrix obtained from (1). Eq. (3) is used for the modelling window should be wide in order to give the
samples from both the prediction and the modelling window. best possible representation of the variation in all stages of
The critical value for the T2 statistics, dcrit , is calculated the process. This means that some subjectivity is connected
as [11]: to the selection of w and k. The modelling and prediction
window should be of the same size so that ‘like is compared
k(w2 − 1)
dcrit = Fα,k,w−k (4) with like’.
w(w − k) Obviously, the significance level α must also be chosen,
where k and w correspond to the number of components but this is done according to prevailing statistical considera-
and the number of samples used to build the reference PCA tions. In the case of monitoring process changes as few false
model, respectively. F is the F-statistics at significance level positives as possible is wanted so the α-value is set high, i.e.
α with the given degrees of freedom. α = 0.99.
Typically a range of window sizes are tested but not neces-
2.1.2. Choosing window sizes and number of sarily as a series since the magnitude is more important than
components the exact window size. Here, w is tested from 10 to 100 in
Analogously to SIMCA approaches it is suggested to use steps of 5, while models with 1–10 components are tested.
areas with reference variation to determine reasonable win- This search strategy gives response surfaces similar to the one
dow sizes w and the right number of components k to be shown in Fig. 1. The response is the total number of atypi-

Fig. 1. The response surface shows the occurrences for different combinations of window sizes and number of components for the reference region in the
analysed data. The optimal settings for the Caterpillar algorithm are combinations with few occurrences and wide windows.
G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291 283

cal samples (d > dcrit ) for the given settings of window size 2.2. SIMCA
and number of components. The details of the interpretation
of Fig. 1 is discussed in Section 4, but generally a combina- In SIMCA a set of reference samples, Xref , which de-
tion of window size and number of components that give few scribes the reference state e.g. stable process or smoothly
atypical samples are chosen. running fluidised bed, is selected. The dimension of the ref-
Importantly, the systematic testing of window sizes and erence set is r × n, where r is the number of reference sam-
number of components gives information on how many atypi- ples and n the number of variables, here frequencies. A PCA
cal samples one can expect from any combination. The num- model describing the group of reference samples is built anal-
ber of atypical samples found in the reference data for the ogously to Eq. (1) using:
chosen window size and number of components is used as
a base level in the later application of the Caterpillar al- Xref = Tref Ptref + E (5)
gorithm. More atypical samples than this base level is de- where the matrices Tref and Pref represent the scores and
manded in order to accept atypical samples as a sign of loadings, respectively. They consist of k score and loading
process changes. vectors, which describe the variation in the reference set.
Usually, the number of components is decided by means of
2.1.3. Operator information—occurrences plot cross-validation [12]. E is the r × n residual matrix, which
The occurrences plot is suggested as a feasible operator contains the variation in the reference set that is not explained
tool for detecting process changes using the Caterpillar algo- by the model. Non-reference samples are compared to the
rithm since the occurrences plot compresses the information model using:
in the prediction windows into single numbers. ‘Occurrences’
is the term chosen to represent atypical samples (d > dcrit ). At tfit = xi Pref (6)
each time point the number of atypical samples in the cor-
ei = xi − tfit Ptref (7)
responding prediction window is counted and the result is
indicated in the occurrences plot. A high number of atypical where the vector tfit represents the scores for sample xi when
samples compared to the base level (see previous section) in- fitted to the model, and ei is the deviation from the model for
dicates that a process change, e.g. upset, is happening. How- sample i, i.e. the residual. The residuals are used to calculate
ever, it should be emphasized that the Caterpillar algorithm the residual standard deviations (R.S.D.) according to:
detects any unusual changes in the measurement signal so 
the cause of the upset must be determined by other means. In eti ei
addition to being a useful tool for monitoring the entire pre- R.S.D.i = (8)
n−k
diction window, the occurrences plot is robust in the sense that
a single outlier in the prediction window is not interpreted as Eq. (8) is applied for the samples that do not belong to the ref-
a process change. erence set. For the R reference
 samples
 R.S.D.
 i is multiplied
r−1
A true process change is always reflected by a series of by a correction factor c c = r−k−1 in order to obtain
occurrences. This is especially important since a confidence R.S.D.i,ref [9].
level of 0.95 means that 5 out of every 100 samples can be The mean of the R.S.D. values for the reference set,
atypical before there is a significant change in the process. R.S.D.ref , is used to determine the border of the model,
Hence, if a window size w of 20 is used, 1 atypical sam- R.S.D.crit , using:
ple can be expected in every window. Another feature of 
the occurrences plot is that the ‘base level’ can be set ac- R.S.D.crit = R.S.D.2ref · Fcrit (9)
cording to the preliminary analyses for determining window
widths and number of components. Thus the normal num- where Fcrit is the F value at the selected level of significance α,
ber of occurrences can be ignored or eliminated from the i.e. Fα,(n−k),(n−k)(r−k−1) . Samples with R.S.D. values larger
plot and all attention is easily directed towards significant than R.S.D.crit are judged to be different from the reference
process changes. samples.
The information reflected in the occurrences plot at time t
is the number of occurrences observed in the prediction win-
dow including the measurements from time t − w to t. This 3. Experimental/data
means that the information in the occurrences plot is lagged
with w time points, i.e. changes in the process can be de- The data analysed in this work are passive acoustic moni-
tected up to w time points after they actually occurred. In the toring data describing a laboratory scale fluidised bed, which
case of acoustics monitoring, this is usually not a problem as is sprayed with a liquid in order to induce agglomeration.
new measurements easily are sampled each second or faster. The acoustics monitoring is done by commercially avail-
Hence the lag is only 20 s at the most if the window size is able equipment (GranuMet, Process Analysis and Automa-
20 and this is small enough to still give an early warning for tion Ltd., Farnborough, UK). A resonant acoustic emission
upsets such as agglomeration. sensor (resonant frequency 190 kHz) transducer is mounted
284 G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291

external to the fluidised bed. The signal is converted to an (33 points wide second order polynomial) data with the time
RMS representation electronically, and this RMS signal is profile indicated by increasing size of the points. The process
digitally sampled at a rate appropriate to the process (typ- starts in the region near the point (−0.5, 0), it moves to the
ically in the order of 100 Hz). From this time series data a left for the transient agglomeration and moves over on the
spectrum is calculated using the FFT algorithm. To improve right-hand side with scores larger than 0 along component
the reproducibility of the spectrum several spectra are aver- 1 for the major agglomeration event. However, in process
aged. In this case, the averaged spectrum has 64 points or monitoring the major objective is to detect process changes
discrete frequencies, which is then used as ‘finger-print’ of as they happen. In this perspective PCA of historical data is
the process condition. not feasible or appropriate.
The spectra are imported into MATLAB 6.5 (MathWorks, In the Caterpillar approach two windows are moved side
Natick, MA), and all calculations are done in this environment by side through the data as it is collected. Possible process
using a combination of in-house routines, routines from the changes are detected by comparing the samples in the pre-
PLS-toolbox (Eigenvector Research, Manson, WA), and rou- diction window (most recent samples) with a PCA model
tines from the Statistics toolbox (MathWorks, Natick, MA). describing the systematic variation for the samples in the
modelling window. The first step in the Caterpillar approach
is to choose the window size, for the moving windows, and the
4. Results and discussion number of components to use for modelling. This can be done
by systematic testing of different window sizes and number
The time profiles for the 64 frequencies are shown in Fig. 2 of components of data from a reference region as outlined
and the three main phases of the process is indicated. The in Section 2.1.2. Time point 1–400 is defined as the refer-
first phase is a steady-state phase where the fluidised bed is ence region for the data analysed in this work. The response
operated under normal conditions. In the second phase, the surface for different window sizes and number of compo-
fluidised powders are sprayed with a liquid and a transient nents in the reference model is shown in Fig. 1. The window
agglomeration is induced. The process returns to almost nor- sizes w are varied in the range of 10–100 in steps of 5, and
mal operating conditions before it is sprayed again and a ma- the models with 1–10 components are tested. The response
jor agglomeration (non-recoverable) event is induced (third function is the total number of occurrences throughout the ref-
phase). erence region, i.e. samples in the prediction windows that are
If acoustic emission data for the whole process is available significantly different (d > dcrit ) from the models describing
it is usually easy to spot the process phases for a fluidised the variation in the corresponding modelling windows. Large
bed using a score plot. In Fig. 3, the score plot of the two window sizes and few number of components is generally
first scores of the mean centred Savitzky-Golay smoothed performing poorly (many occurrences) while small window

Fig. 2. The changes in the acoustic signal over time are shown. The three phases of the process are indicated as blocks below the data.
G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291 285

Fig. 3. The score plot of the two first components from the PCA model clearly shows the time profile of the process. Time is indicated by increasing size of
the time points.

sizes and many components are generally performing well as ‘Reference’ in Fig. 2). A six component SIMCA model
(few occurrences). This is as expected, but the challenge is of the mean-centred data was built and the relative standard
to find a combination with a relatively wide window and a deviations (R.S.D.) for the remaining samples were calcu-
reasonable number of components (parsimonious principle) lated. In Fig. 4b the R.S.D. values for the different times are
with a good performance. The area of feasible combinations shown and the critical R.S.D. value is indicated (horizontal
is indicated in Fig. 1. In the further analyses, six components line). The reference samples are indicated by a dashed line
and a window size of 65 points has been used. and their R.S.D. values are (mostly) below R.S.D.crit while
It is suggested to use a reference region to find the optimal most of the non-reference samples have higher R.S.D.fit val-
algorithm settings, although this is not always critical. For the ues than the critical value. Clearly, it is difficult to recognise
data discussed in this paper a subset including the samples the different stages in the process since the reference set is
from time 1200 to 1643 were tested for different window pre-defined and fixed.
sizes and number of components. The main features of the
resulting response surface were similar to those in Fig. 1. 4.1. Shift in signal
In spite of this, it is recommended to use a reference region
to find the optimal algorithm settings. Especially as using a A shift in the signal is simulated by adding a random
reference region known to contain no process changes gives value in the range [ 0 1 ] to each of the frequencies, and
a known reference number of occurrences, e.g. zero. the added values are allowed to be different from frequency
Running the Caterpillar algorithm with six components in to frequency. The calculation of the shifted data, Xshift , can
the reference models and a window size w of 65 resulted in be formulated as:
the occurrences plot shown in Fig. 4a. The horizontal dashed
line indicates the base level (i.e. [1]) for the occurrence plot, Xshift = X + 1 · a (10)
i.e. the number of occurrences in the reference region as ob-
tained from Fig. 1. Occurrences above the base level indicate
a change in the acoustic signal, e.g. a process change. Clearly, Table 1
the occurrences plot reports five events in the data as indi- Occurrence Time Possible event
cated with the letters A–E in Fig. 4a. The five detected events A 09:41:04 Spraying starts
coincide with different process events as listed in Table 1. All B 09:44:20 Transient agglomeration
major process events are detected. C 09:50:27 End of transient agglomeration.
The data was also analysed using the SIMCA. The samples D 09:54:43 Spraying
E 09:57:21 Agglomeration
from time 1 to 400 were included in a reference set (labelled
286 G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291

Fig. 4. (a) The occurrences plot from the Caterpillar algorithm is shown. The five major events in the analysed process is easily recognised and labelled by
letters A–E in the figure. The horizontal line is the base level of the occurrences plot and (b) the R.S.D. values for the samples fitted to a six-component SIMCA
model built from the 400 first-time points are shown. The reference samples are indicated by a dashed line, and the horizontal line is the critical R.S.D. value,
i.e. the border of the SIMCA model.

where X is the original data, 1 is a m × 1 column vector of data are indicated by negative occurrences in the same figure.
ones and a is a 1 × n vector of random values. Clearly the results are the same.
The occurrences plot obtained when analysing the shifted In Fig. 5b the R.S.D.fit values for the shifted data are
data is shown in Fig. 5a. The results using the non-shifted shown. The same model as developed for the original
G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291 287

Fig. 5. (a) The occurrences plot for the shifted data is shown. The occurrences for the non-disturbed data is shown as negative occurrences. The horizontal line
is the base level of the occurrences plot and (b) the R.S.D. values for the shifted samples fitted to a six-component SIMCA model built from the 400 first-time
points from the non-disturbed data are shown. The reference data interval is indicated by a dashed line, but do note that the shown R.S.D. values are those for
the shifted data. The horizontal line is the critical R.S.D. value, i.e. the border of the SIMCA model.

data is used in order to simulate a shift in the signal be- if the data are scaled by the means from the reference
tween calibration and measurement, and the R.S.D.crit bor- set, all samples are judged to be outside the model, and
der is indicated by the horizontal line. The results seem this would be a likely approach in an online applica-
to be the same as for the non-shifted data. However, tion.
288 G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291

4.2. Drift in data different for the different frequencies and u is the vector of
time points. The maximum value for lj is chosen randomly on
Drift in the data is simulated by adding a m × 1 linear the interval [ 0 1 ] and bj is set accordingly. This is relatively
contribution lj to each of the n frequencies. The linear con- large values since the intensity of the original signal varies on
tribution is of the type lj = bj u where bj is a constant which is the interval [ 2 8 ] depending on the frequency. The results

Fig. 6. (a) The occurrences plot for the data with drift is shown. The occurrences for the non-disturbed data is shown as negative occurrences. The horizontal
line is the base level of the occurrences plot and (b) the R.S.D. values for the samples from data with drift fitted to a six-component SIMCA model built from
the 400 first-time points from the non-disturbed data are shown. The reference data interval is indicated by a dashed line, but do note that the shown R.S.D.
values are those for the data with drift. The horizontal line is the critical R.S.D. value, i.e. the border of the SIMCA model.
G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291 289

from running the Caterpillar algorithm on the data where drift inal data is used, and the R.S.D.crit border is indicated by the
is added are shown in Fig. 6a. Comparing with the original horizontal line. In this case the reference set (dashed line) as
occurrences, which are indicated as negative values in the well as the non-reference samples are outside the model and
figure the main events in the data are still detected in spite of their distance to the model is increasing with time (expected
the added drift. due to the drift). It is also worth noticing that any features of
The R.S.D.fit values for the data where drift is added are the non-reference samples are lost compared to the original
shown in Fig. 6b. The same model as developed for the orig- R.S.D.fit values (Fig. 4b).

Fig. 7. The occurrences plot for the data with different types of noise is shown. The occurrences for the noise-free data is shown as negative occurrences. The
horizontal line is the base level of the occurrences plot. (a) Random noise; (b) heteroscedastic noise; (c) 1/f noise.
290 G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291

Fig. 7. (Continued ).

4.3. Noise in data mal distribution with mean 0 and standard deviation 0.05
giving noise with variation spans up to [ −1.5 1.3 ]. The
The robustness of the Caterpillar algorithm towards dif- heteroscedastic noise was added to the data matrix and the
ferent types of noise including random or additive noise, het- Caterpillar algorithm was run using the earlier mentioned
eroscedastic and 1/f noise is tested. Additive noise results settings.
from electrical noise in the measurement system, interfer- The results from the Caterpillar algorithm are shown in
ence from other processes, etc., and is generally independent Fig. 7b. Again, all the main events in the data are detected in
of the process in question. 1/f noise and other heteroscedastic the occurrences plot. In the SIMCA modelling all the sam-
noises are related to the natural variation in the process. The ples including the reference samples are classified as being
1/f component is associated with slow drifts in the process different from the noise-free reference model. Using a refer-
conditions. ence model with noise gives SIMCA performance similar to
the noise-free data (Fig. 4b).
4.3.1. Random noise
The random noise is simulated by adding random numbers 4.3.3. 1/f noise
from a normal distribution with mean zero, standard devia- 1/f noise is inversely proportional to the frequency. The
tion 0.1 and variance 0.01. The results from the Caterpillar 1/f noise is simulated by first creating a linear profile l of
algorithm using the earlier mentioned settings are shown in length n. The first element is 1, the last 0.2, and the oth-
Fig. 7a. All the main events in the data are still detected as ers are chosen so that l is linear. The elements in the lin-
can be seen in Fig. 7a. SIMCA does not perform well when ear profile is repeatedly multiplied by a random number
noise is added and the noise-free model is used as the refer- from a uniform distribution on the interval [ 4 4.2 ] in or-
ence. Most of the samples including the reference samples are der to build the m × n 1/f noise matrix. The 1/f noise ma-
outside the model. However, if the reference region from the trix is added to the original data and the Caterpillar algo-
noisy data is used SIMCA perform similarly to the noise-free rithm is run using the same settings as earlier. The results
data (Fig. 4b). are shown in Fig. 7c. As for the other types of noise the
main events in the fluidised bed process are detected, and in
4.3.2. Heteroscedastic noise this case all the events seen in Fig. 4a are contained. Again,
Heteroscedastic (or multiplicative) noise varies in mag- all the samples including the reference samples are classi-
nitude with the intensity of the signal. In this work het- fied as being different from the noise-free reference model in
eroscedastic noise was simulated by multiplying each ele- the SIMCA modelling unless a reference model with noise
ment in the signal matrix with random numbers from a nor- is used.
G.R. Flåten et al. / Analytica Chimica Acta 544 (2005) 280–291 291

5. Conclusions [2] J.W.R. Boyd, J. Varley, The uses of passive measurement of acoustic
emissions from chemical engineering processes, Chem. Eng. Sci. 56
The Caterpillar algorithm is presented and demonstrated (2001) 1749–1767.
[3] A. Björk, L.-G. Danielsson, Spectra of Wavelet scale coeffi-
on acoustic data from a laboratory scale fluidised bed. The
cients from process acoustic measurements as input for PLS mod-
settings for the algorithm are discussed and it is proposed elling of pulp quality, J. Chemometr. 16 (8–10) (2002) 521–
how the algorithm performance can be optimised by using 528.
a reference region to choose the optimal algorithm settings. [4] M. Whitaker, et al., Application of acoustic emission to the monitor-
It is also shown that the algorithm is robust towards differ- ing and end point determination of a high shear granulation process,
Int. J. Pharm. 205 (2000) 79–91.
ent types of noise, and shift and drift in the signal for the
[5] A. Nordon, et al., Monitoring of a heterogeneous reaction by acoustic
studied data. This is a possible indication that the Caterpillar emission, Analyst 129 (5) (2004) 463–467.
algorithm settings are transferable between similar systems [6] G.R. Flåten, A.D. Walmsley, Tracking process changes in a fluidised
and batch independent. Clearly, system transferability and bed using acoustic sensors, Chemometr. Intell. Lab. Syst., submitted
batch independency are very advantageous features and thus for publication.
[7] C.C. Tan, N.F. Thornhill, R.M. Belchamber, Principal component
interesting to confirm in future research.
analysis of spectra with application to acoustic emissions from me-
The algorithm was robust towards the types of signal dis- chanical equipment, Trans. Inst. Meas. Contr. 24 (4) (2002) 333–
turbances tested in this work implying the hypothesis that 353.
the need for recalibration is minimal. If this hypothesis gen- [8] K.H. Esbensen, et al., Acoustic chemometrics—from noise to infor-
erally holds the building and maintenance of training models, mation, Chemometr. Intell. Lab. Syst. 44 (1998) 61–76.
[9] S. Wold, Pattern recognition by means of disjoint principal compo-
which are work intensive parts of using SIMCA models are
nents models, Pattern Recogn. 8 (1976) 127–139.
minimised. Additionally, the Caterpillar algorithm seems to [10] B.M. Wise, N.B. Gallagher, The process chemometrics approach to
perform as well or better than the SIMCA approach in mon- process monitoring and fault detection, J. Process Contr. 6 (6) (1996)
itoring changes in the process. 329–348.
[11] T. Kourti, J.F. MacGregor, Process analysis, monitoring and diagno-
sis, using multivariate projection methods, Chemometr. Intell. Lab.
References Syst. 28 (1995) 3–21.
[12] S. Wold, Cross-validatory estimation of the number of components
in factor and principal components models, Technometrics 20 (1978)
[1] R.M. Belchamber, et al., Quantitative study of acoustic-emission
397–405.
from a model chemical process, Anal. Chem. 58 (8) (1986)
1873–1877.

You might also like