You are on page 1of 14

AIAA Propulsion and Energy Forum 10.2514/6.

2017-4892
10-12 July 2017, Atlanta, GA
53rd AIAA/SAE/ASEE Joint Propulsion Conference

Prediction of Thermoacoustic Instabilities in a Premixed


Combustor based on FFT-based Dynamic Characterization

Sudeepta Mondal1 , Chandrachur Bhattacharya1, Pritthi Chattopadhyay1, Achintya Mukhopadhyay2 and Asok Ray1a
1
Department of Mechanical & Nuclear Engineering, Pennsylvania State University, University Park, PA 16802, USA
2
Department of Mechanical Engineering, Jadavpur University, Kolkata 700 032, India
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

The classification of a combustion process into stable and unstable regimes is essential for the design of
combustor systems, which include their monitoring and control mechanisms; an unsupervised method for
monitoring of combustion instabilities is essential from these perspectives. The traditional methods of
combustion stability assessment are often computationally intensive and thus may not be suitable for real-time
monitoring and control. This paper proposes a fast Fourier transform (FFT)-based method for dynamic
characterization of combustion instabilities. The proposed method has been tested on experimental data of
pressure time-series from a lean-premixed swirl-stabilized combustor, and the results have been compared with
the well-known method of visibility graph.

I.Introduction

Instabilities in combustion systems are usually related to the spontaneous excitation of one or more natural acoustic

modes of the combustor [1]. These phenomena are typically manifested by large amplitude self-sustained oscillations

in the combustion chamber, which results from a feedback loop established between the heat release rate from the

flame and the combustion chamber acoustics [2]. The detrimental effects of combustion instabilities include

generation of externally audible tones at intolerable levels, and (possibly) resonance leading to failure in the

mechanical structures if the pressure oscillations match the natural frequency of the system [3, 4].

Various control strategies for suppression of combustion instabilities have been proposed by researchers over the

years, which include various open loop [5-7] and closed loop active control techniques [8]. The review articles by

Candel [9] and Culick [10] list several such techniques. Timely identification of unstable combustion phenomena is

important from the perspectives of monitoring and control, more so in a closed loop active control system. This paper

proposes to develop metrics for dynamic data-driven application systems (DDDAS) [11], which can be used for online

monitoring and control of combustion processes. Time series of pressure dynamics from a laboratory-scale swirl-

1a
Professor, Department of Mechanical Engineering, Pennsylvania State University
1
Research Assistant, Department of Mechanical Engineering, Pennsylvania State University
1,2
Corresponding author email: axr2@psu.edu; achintya.mukho@gmail.com;

1
American Institute of Aeronautics and Astronautics
Copyright © 2017 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
stabilized combustor have been used to validate the instability prediction algorithms that are developed based on

frequency domain analysis of limit-cycle oscillations. The frequency spectrum of combustion pressure and

chemiluminescence time-series data has been commonly used in the combustion literature for instability

characterization, based on visible identification of sharp excitations at harmonics of excited acoustic modes. Several

researchers have recently proposed dynamic data-driven methods of instability prediction. For example, Gotoda et al.

[12] have used nonlinear time series analysis to characterize combustion instabilities close to lean blowout conditions.

Nair et al. [13] have proposed a method of predicting the onset of instabilities by identifying regions of intermittent
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

and randomly occurring high-amplitude periodic oscillations, which are interceded with regions of low-amplitude

oscillations. Sarkar et al. [14] have proposed a dynamic data-driven method of detecting early onsets of combustion

instabilities by using the spatio-temporal co-dependence among data obtained from heterogeneous sensors, in the

framework of symbolic time series analysis via generalized D-Markov machine construction [15,16]. A comparison

of several state-of-the-art data-driven techniques in terms of computational complexity have been also reported by

Sarkar et al. [14]. Apparently, a structured framework of training an instability classifier has not been reported in

open literature, based on frequency domain representation of (limit-cycle) pressure oscillations.

This paper proposes a fast Fourier transform (FFT)-based method for dynamic characterization of combustion

instabilities; the performance of the proposed FFT-based classifiers has also been compared with a well-known

visibility graph method that has been used by other researchers to predict combustion instabilities based on dynamic

data-driven time series analysis of pressure oscillations. The proposed method has been tested on experimental data

of pressure time-series from a partially premixed combustor, and the results have been compared with the visibility

graph algorithm.

Complex networks have been used by many researches to study the dynamic nature of a time series, by extraction

of features underlying in the time-series data [17, 18]. There are several methods to convert a time-series into a

corresponding complex network, one of which is the visibility graph method given by Lacasa et. al. [19]. The first

attempt to study combustion instability and thermoacoustics using a complex network approach was done by

Murugesan et. al [20], where they showed that combustion noise is scale invariant, and that this invariance vanishes

when thermoacoustic instability sets in and the nature of the oscillations begins to follow a more ordered pattern; a

standard measure was introduced to classify the stability of a combustion system. Along this line, this paper proposes

to establish FFT-based metrics for classification of pressure time-series data into stable and unstable regimes, where

2
American Institute of Aeronautics and Astronautics
the underlying algorithms are significantly computationally efficient than the aforementioned data-driven tools. Thus,

the proposed FFT-based metrics are suited for real-time combustion instability identification and control. The

following section describes the experimental apparatus and data acquisition details.

A. Experimental Apparatus

A swirl-stabilized, lean-premixed, laboratory-scale combustor has been used for validation of the FFT-based

algorithms with experimental data. Figure 1 depicts a schematic diagram of the variable-length combustor apparatus

that consists of an inlet section, an injector, a combustion chamber, and an exhaust section. There is an optically-
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

accessible quartz section followed by a variable length steel section.

Fig.1: Schematic diagram of the combustion apparatus

B. Experimental procedure and data acquisition

High pressure air is delivered to the apparatus from a compressor system after passing through filters to remove

any liquid or particles that might be present. The air supply pressure is set to 180 psig using a dome pressure regulator.

The air is pre-heated to a maximum temperature of 250◦C by an 88kW electric heater. The fuel for this study is natural

gas (approximately 95% methane) which is supplied to the system at a pressure of 200 psig. The flow rates of the air

and natural gas are measured by thermal mass flow meters. The desired equivalence ratio and mean inlet velocity is

set by adjusting these flow rates with needle valves. Tests are conducted at a nominal combustor pressure of 1 atm

over a range of operating conditions, as listed in Table 1. There are 780 such operating conditions using combinations

of the parameter values. Under each operating condition, 8 seconds pressure time series are collected at a sampling

rate 8192 Hz.

3
American Institute of Aeronautics and Astronautics
Table 1. Parameters of Experimental Apparatus under Different operating conditions

Parameters Values
Equivalence Ratio 0.525, 0.55, 0.60, 0.65
Inlet Velocity 25-30 m/s with increments of 5m/s
Combustor Length 25-59 inch in 1 inch increments

C. Experimental Observations

Figure 2 shows time series plots of the pressure samples which had the greatest (1.5424 Psi) and the least
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

𝑃𝑟𝑚𝑠 (0.0243 Psi). It can be easily seen that the unstable pressure signature has significantly high amplitudes over the

recorded period of time. It had been observed experimentally that the combustor becomes unstable above a threshold

𝑃𝑟𝑚𝑠 of 0.07 Psi. Thus 𝑃𝑟𝑚𝑠 =0.07 Psi has been used as the ground truth for training the proposed classifiers.

Fig. 2: Pressure time series over 8 seconds for the most stable and most unstable cases

II.Analysis of Experimental Data

This section is dedicated to analysis of experimental data of pressure dynamics and consists of the following three

subsections: (A) Stability metrics based on FFT, (B) Description of the visibility algorithm, and (C) Decision

making based on the choice of an optimal threshold, as described below.

A. Stability Metrics based on FFT

Power spectra of pressure and heat release oscillations have been commonly used for investigating phenomena of

thermoacoustic instabilities [5]. The excitation of unstable modes can be verified by analyzing the power spectrum of

the pressure and chemiluminescence signals [21]. Two FFT-based metrics, namely, Metrics A and B, have been

proposed in this paper for instability characterization. These FFT-based techniques have been used over various

4
American Institute of Aeronautics and Astronautics
lengths of sampling time ranging from 1 s to 8 s, and the results have been compared with the ground-truth provided

by root mean square of pressure fluctuations (𝑃𝑟𝑚𝑠 ) over the entire duration of the 8 s interval, which allows

observation of frequency response for a typical stable and an unstable signal as depicted in Fig. 3. Out of the 780

operating conditions, the pressure time series having the greatest (1.5424 Psi) and the least 𝑃𝑟𝑚𝑠 (0.0243 Psi) were

picked for the FFT analysis. It is seen in Fig. 3 that the power spectrum of the unstable pressure series (marked in red)

shows strong excitations of the harmonics of the unstable mode, while that of the stable case (marked in blue) lacks

such a behavior due to the absence of an excited mode. It is also worth noting that the maximum amplitude of the
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

power spectrum of the stable signal is about thousand-order smaller than that of the unstable signal.

Fig. 3. Typical Power Spectrum of a Stable and an Unstable operating condition



Energy of a time series 𝑥(𝑡) is conventionally denoted as: 𝐸 ≜ ∫−∞|𝑥(𝑡)|2 𝑑𝑡 for a finite energy signal (square

integrable function with respect to Lebesgue measure). By Parseval’s theorem, an analogous expression is obtained
∞ ∞
for the signal energy in the frequency domain, which states that = ∫−∞|𝑥(𝑡)|2 𝑑𝑡 = ∫−∞|𝑥̂(𝑓)|2 𝑑𝑓, where 𝑥̂(𝑓) is the


Fourier transform of 𝑥(𝑡), i.e. 𝑥̂(𝑓)=∫−∞ 𝑒 −2𝜋𝑖𝑓𝑡 𝑥(𝑡) 𝑑𝑡, where 𝑓 denoted the frequency in Hz [22]. Therefore, the

area under magnitude-squared power spectrum amplitude over a compact support of the frequency range would

indicate the energy content in chosen bandwidth. This idea is utilized in the selection of the following two metrics

for instability characterization.

 Metric A calculates the ratio of the energy content of the dominant frequency (i.e., the frequency

corresponding to the maximum amplitude in the power spectrum) to the energy content of the second most

5
American Institute of Aeronautics and Astronautics
dominant frequency that is a non-harmonic of the dominant frequency. Since combustion instability causes

excitation of harmonics of the unstable mode, as seen in Fig. 3, most of the signal energy would be

concentrated in the unstable mode and its harmonics. So, for a stable signal, the ratio is expected to be

considerably higher than that of an unstable signal, because a stable combustion system would most likely

resemble a near-uniform distribution of energy over the entire spectrum without any significant excitation of

a mode, as seen in the blue profile of Fig.3. In this setting, Metric A classifies a signal as stable if the ratio

is greater than a threshold value. The area under the squared amplitude corresponding to the dominant
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

frequency has been calculated over a bandwidth support of 2 Hz, centered at the dominant frequency, which

is assumed to yield its energy content.

 Metric B calculates the ratio of the energy content of the dominant frequency to the total energy content of

the signal. An unstable signal would have a significant energy content in the excited dominant frequency in

contrast to a stable signal. Thus, Metric B classifies a signal as unstable if the ratio is greater than a threshold

value. Similar to Metric A, a 2Hz bandwidth support has been used to calculate the energy content of the

dominant frequency.

1
It is noted that 𝑃𝑟𝑚𝑠 ≜ √ ∑𝑁 ̅ 2 ̅ 1 𝑁
𝑖=1[(𝑃𝑖 − 𝑃 ) ] is dependent on the computed mean, 𝑃 ≜ ∑𝑖=1 𝑃𝑖 , of the sampled
𝑁 𝑁

pressure data, which is not available until the Nth sample [23]. Furthermore, for a non-stationary process, 𝑃𝑟𝑚𝑠 (which

is approximately the standard deviation) is expected to be a function of time as demonstrated in Fig. 4 that shows the

RMS values of a stable pressure time series calculated at time intervals of 0.5 sec. Therefore, a threshold based only

on 𝑃𝑟𝑚𝑠 may not be a suitable measure for online prediction in a non-stationarity process.

0.08
Rms value
0.075

0.07
Pressure (psi)

0.065

0.06

0.055

0.05
1 2 3 4 5 6 7 8
Time (seconds)

Fig. 4. Variation of 𝑷𝒓𝒎𝒔 as a function of time

6
American Institute of Aeronautics and Astronautics
It is seen in Fig. 4 that the threshold of 0.07 psi is exceeded at 1.5, 2 and 4.5 sec marks, while 𝑃𝑟𝑚𝑠 for the entire 8

sec long data is 0.065 Psi. Since the stability label would depend on the duration of sampling if the criterion of 𝑃𝑟𝑚𝑠

only is used as the threshold, it is possible that an actually stable system could be erroneously identified as unstable.

Furthermore, the computed 𝑃𝑟𝑚𝑠 is dependent on the combustor geometry and its operating conditions, which is

specific to a particular combustor. Therefore, FFT-based techniques are proposed as an alternative for pattern

classification, which is shown later in the paper to yield satisfactory results over varying time windows.
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

B. Description of the Visibility Graph

The visibility graph, presented by Lacasa et. al. [19], has been used to convert the pressure time-series signal

generated by the experimental apparatus into a complex network. To do this, only the positive peaks of the time-series

p(t) are considered and they are represented as vertical bars. The tip of each bar corresponds to a node, denoted as pi,

in the complex network that is formed between each point and all the other points it can ‘see’. So, if nodes i and j are

connected by a straight line without intersecting any intermediate node, Ai,j =1, else it is 0. To avoid self-connections,

it is mandated that Ai,i=0 for all i. Thus, the visibility matrix contains information about the nodes and their

interconnections.

Nunez et. al. [24] introduced the concept of a threshold ‘ε’ to better visualize the results when the signal is noisy.

This threshold is defined as a product of a multiplying factor ‘e’ and the mean of all the positive peak values of the

time series. The threshold accounts for the cycle-to-cycle variation of the time series data. It is thus necessary to choose

the value of ‘e’ appropriately to obtain a good classification, as explained later in the paper.

Mathematically, the elements of the visibility matrix (A) is defined as:

𝑡𝑘 − 𝑡𝑖
1, 𝑖𝑓, 𝑝𝑘 + 𝜀 < 𝑝𝑖 + (𝑝𝑗 − 𝑝𝑖 )
𝐴𝑖,𝑗 = { 𝑡𝑗 − 𝑡𝑖 (𝑖 < 𝑘 < 𝑗)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒, 𝑤ℎ𝑒𝑟𝑒 𝜀 = 𝑒 × 𝑚𝑒𝑎𝑛(𝑝)

Now a degree of a node is defined as the number of other nodes it is connected to, and a distribution P(k) denotes

the percentage of nodes of degree k. Then, the degree distribution graph is plotted as log(P(k)) vs log(k). As shown by

Murugesan and Sujith [25], the degree distribution can be used for classification among combustion noise,

intermittency, and combustion instability. It has been seen that, during instability (for a suitable threshold), the degree

distribution shows approximately as a single point, while the number of such points exceeds one for other cases. This

paper classifies a given pressure time-series of a combustor run for a binary classification of either stable or unstable.

7
American Institute of Aeronautics and Astronautics
It follows from the degree distribution plot that the time-series is unstable if it shows or a single point and stable if it

contains multiple points.

Here again, in order to choose the appropriate value of the threshold, the visibility algorithm is used to classify

each of the 780 time-series data for different experimental conditions as either stable or unstable. Upon comparison

of the observations with the ground truth, the minimum average Bayes’ cost criterion is used to find the optimum

threshold [26]. A plot of total cost versus the multiplying factor ‘e’ is computed and the optimal value of ‘e’

corresponds to that with the lowest total cost.


Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

C. Decision Making based on the Choice of an Optimal Threshold

As discussed above, Metrics A and B as well as the visibility algorithm (discussed in Section B) leads to

identification of an optimal threshold. The results presented in this paper have been based on this optimal threshold

(𝜆∗ ) selected by using Bayes’ criterion of minimum average cost [26], as explained below.

Let us consider a detection problem, where the proposed algorithms make a decision 𝐷𝑖 (𝑖 = 1 𝑜𝑟 0) that exactly

one of the two hypotheses 𝐻1 (unstable case) or 𝐻0 (stable case) has occurred, where i=j corresponds to a correct

decision and incorrect otherwise. Bayes’ criterion assigns “costs” to the decision outcomes as explained in Table 2.

Table 2: Costs associated with different decisions

Actual class Predicted class Type of decision Cost


𝑯𝟎 𝐻0 True negative 𝐶00
𝑯𝟎 𝐻1 False Alarm 𝐶10
𝑯𝟏 𝐻0 Missed Detection 𝐶01
𝑯𝟏 𝐻1 True Positive 𝐶11

Since incorrect decisions generally cost more than correct decisions, false alarm and missed detection would have

greater costs associated with them as compared to the correctly detected cases. Considering the individual costs, the

total average cost or risk associated with a decision is expressed as:

𝐶 = 𝑃0 [𝐶00 𝑃(𝐷0 |𝐻0 ) + 𝐶10 𝑃(𝐷1 |𝐻0 )] + 𝑃1 [𝐶01 𝑃(𝐷0 |𝐻1 ) + 𝐶11 𝑃(𝐷1 |𝐻1 )] (1)

where 𝑃0 and 𝑃1 are the prior probabilities of the two hypotheses; 𝑃(𝐷1 |𝐻0 ) is the false alarm probability and

𝑃(𝐷0 |𝐻1 ) is the probability of missed detection, while 𝑃(𝐷0 |𝐻0 ) and 𝑃(𝐷1 |𝐻1 ) are the corresponding probabilities of

correct decisions. They are related as: 𝑃(𝐷0 |𝐻0 ) + 𝑃(𝐷1 |𝐻0 ) = 1 and 𝑃(𝐷0 |𝐻1 ) + 𝑃(𝐷1 |𝐻1 ) = 1. The probabilities

are generally functions of the selected threshold such that the optimal threshold minimizes total average cost of the

1
decision. Uniform costs and equal priors are chosen in this paper: (𝐶00 = 𝐶11 = 0, 𝐶01 = 𝐶11 = 1, 𝑃1 = 𝑃0 = ),
2

8
American Institute of Aeronautics and Astronautics
1
that, in the absence of any additional information, yields the average cost in Eq. (1) as 𝐶 = (𝑃(𝐷1 |𝐻0 ) + 𝑃(𝐷0 |𝐻1 )).
2

The optimal threshold is the one that has a minimum sum of the false alarm probability and the missed detection

arg 𝑚𝑖𝑛 1
probability, 𝑖. 𝑒., (𝜆∗ = 𝜆𝐶 ). Neglecting the factor , the sum 𝑃(𝐷1 |𝐻0 ) + 𝑃(𝐷0 |𝐻1 ) is taken as the total cost,
2

for which the optimal threshold 𝜆∗ is determined. It is to be noted that usually missed detection for an unstable system

may incur greater penalties, and thus the framework could be modified to use a higher value of 𝐶01 relative to 𝐶10 ,

which may affect the choice of 𝜆∗ .


Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

III.Results and Discussions

This section presents comparisons of the results of different classification performances using the proposed FFT-

based algorithms with those generated by the visibility algorithm.

A. Results on FFT-based Stability Metrics

The stability maps, provided by Metrics A and B, are presented here and the results are compared for various sizes of

time windows that were chosen for FFT analysis.

 Metric A: Table 3 lists the total cost, false alarm and missed detection probabilities for time windows of 1s,

2s, 4s, and 8s, respectively, for a sample space of the 780 pressure series data. It is seen in Table 3 that Metric

A yields comparable results across all time windows, and is thus suitable for classifying stability labels by

processing over a short interval of pressure time series. It is noted that the percentage of total

misclassifications has a decreasing trend as the time window is increased.

Table 3: Various Performance parameters for Metric A at different window lengths

Time Optimal Misdetection for False-Alarms for Optimal Total Misclassification


Window Threshold Optimal Threshold Threshold for Optimal Threshold
(seconds) (percent) (percent) (percent)
1.0 0.10 11.74 9.45 11.15
2.0 0.07 9.67 11.94 10.22
4.0 0.06 7.59 12.47 9.11
8.0 0.03 9.84 9.95 9.87

Figure 5 shows that the Bayes’ cost is minimized at an optimal threshold of 0.1, which is used for identifying the

stability criterion over a 1 s length of time series. As the threshold is increased, the false alarm rate increases while

the missed detection rate decreases. Figure 6 shows the stability map predicted by using the optimal threshold criterion

on the same 1 s long time series. The horizontal line indicates the Prms threshold of 0.07 psi that has been used as a

9
American Institute of Aeronautics and Astronautics
standard for comparing the results of different classification methods. The plot thus gives a map of the classification

obtained from our analysis with respect to the ground truth. Thus, any time series predicted stable but with a Prms

greater than 0.07 Psi is considered to be a missed detection; and a time series predicted unstable but with a Prms less

than or equal to 0.07 Psi is considered to be a false alarm. The circles above the horizontal line are the missed detection

cases, and the crosses below the horizontal line are the false alarm cases. The run-time required for the algorithm

under Metric A for classifying one time-series data has been clocked at 15 milliseconds for Matlab R2014a executed
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

on an Intel Core i5 CPU with 8 GB memory.

Fig. 5. Average Cost, Missed Detection and False Alarm at different thresholds for window size=1 sec

Fig. 6: Stability Map using Metric A on a 1 second long time series data

10
American Institute of Aeronautics and Astronautics
 Metric B: Similar results are presented for stability characterization using Metric B in Table 4 for different

window sizes for the frequency based analysis. The different performance parameters show similar trends as

those with Metric A.

Table 4: Various Performance parameters for Metric B at different window lengths

Time Optimal Misdetection for False-Alarms for Optimal Total Misclassification


Window Threshold Optimal Threshold Threshold for Optimal Threshold
(seconds) (percent) (percent) (percent)
1.0 0.54 8.98 12.92 10.00
2.0 0.50 7.59 9.45 8.07
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

4.0 0.40 5.52 10.44 6.79


8.0 0.35 7.77 7.46 7.69

The results in Tables 3 and 4 clearly show that both FFT-based metrics are able to generate stability labels by

analyzing a pressure signature of the combustor quickly and with considerable accuracy. Even the classification

performance is considerably good for very short time windows of acquired data. The performance parameters of the

algorithms are dependent on the relative weightage of false alarm and missed detection probabilities, as preset by the

user. So it is possible to obtain a very small missed detection rate, if needed, at the expense of a higher false alarm

rate, if the user decides to weigh the missed detected cases with a higher cost. The algorithm offers flexibility in this

regard for the end user to decide on an optimal classifier based on the performance requirements.

For comparison of Metrics A and B in Tables 3 and 4, respectively, it appears that Metric B performs slightly

better than Metric A with the total misclassification error being less for each of the four cases. Furthermore, the run-

time clocked by using Metric B on a single time series data is about 8 milliseconds, which is about half the time

required by Metric A for classifying a time series data. This can be a result of the lesser computational cost involved

in the algorithm proposed by Metric B. Thus, both these data-driven metrics can be used for a very quick

characterization of time series samples.

B. Results on Visibility Graph

It was observed that the optimal threshold varied by a very slight amount depending on the window size of the

pressure time-series data. Hence, similar to the previous cases in Section III.A, a total cost analysis is done on 4

windows sizes ranging from 1s to the total data length of 8 s. Table 5 lists the optimal threshold for each case along

with the misdetection and false alarm rates for each. In order to have a common threshold irrespective of window

11
American Institute of Aeronautics and Astronautics
length, an average value of the optimal thresholds is chosen as a common threshold which leads to a slight change in

the misdetection and false-alarm rates for each window length.

Table 5: Various Performance parameters for the Visibility Graph at different window lengths

Time Optimal Misdetections False Alarms Total Common Misdetections False


Window Threshold for Optimal for Optimal Misclassification Threshold for Common Alarms for
(seconds) Threshold Threshold Rate Threshold Common
(percent) (percent) (percent) (percent) Threshold
(percent)
1s 2.27 8.8 6.96 8.33 2.32 8.29 9.45

2s 2.32 8.46 7.46 8.21 2.32 8.46 7.46


Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

4s 2.33 8.29 7.46 8.08 2.32 9.15 7.46

8s 2.36 8.81 7.46 8.46 2.32 9.33 7.46

Figure 7 presents a plot of average minimum cost, missed detection and false alarm at different thresholds for

window size of 1 s, where the average execution time of the visibility algorithm is 70 milliseconds to classify one time

series data.

Fig.7. Average Cost, Missed Detection and False Alarm at different thresholds for window size of 1 s

12
American Institute of Aeronautics and Astronautics
IV.Summary, Conclusions and Future Work

This paper has developed two fast Fourier transform (FFT)-based methods and associated metrics for classification

of combustion pressure time series data into stable and unstable cases. Both methods have been validated using a

ground-truth knowledge based on the 𝑃𝑟𝑚𝑠 data from a lean-premixed swirl-stabilized combustor, and the results have

been compared with the well-known method of visibility matrix that has been used by researchers for a similar

classification problem. The results generated by the proposed FFT-based methods are comparable to those of the
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

visibility algorithm, and the processing time is faster due to its computational simplicity. Thus, both of these FFT-

based methods are well-suited for online instability detection. It has also been shown that although a single 𝑃𝑟𝑚𝑠

provides a very simple threshold criterion for instability characterization, it might not be a suitable measure for online

detection because of its time dependency in case of a non-stationary process.

A topic of future research is extension of the underlying algorithms for online monitoring and active control of

combustion instabilities. Also, Wavelet transform based analysis would be helpful to track the non-stationarity of the

combustion process. For an online detection problem, wavelet analysis would be helpful to detect the magnitude and

time of occurrence of instability.

Acknowledgments

This work has been supported in part by the U.S. Air Force Office of Scientific Research (AFOSR) under Grant

No. FA9550-15-1-0400. Any opinions, findings and conclusions or recommendations expressed in this publication

are those of the authors and do not necessarily reflect the views of the sponsoring agencies. The authors are thankful

to Professor Domenic Santavica at Penn State, who has kindly provided the experimental data.

References

1. Lieuwen, T.C,.Yang., V., Combustion instabilities in gas turbine engines: operational experience, fundamental
mechanisms and modeling, Combustion instabilities in gas turbine engines: operational experience, fundamental
mechanisms and modeling. American Institute of Aeronautics and Astronautics, 2005: p. Chapter 1.
2. Matveev, K., Thermoacoustic instabilities in the Rijke tube: experiments and modeling. California Institute of
Technology, 2003.
3. Lieuwen, T.C., Yang, V., Incorporation Of Combustion Instability Issues Into Design Process: Ge Aeroderivative And
Aero Engines Experience. Combustion Instabilities In Gas Turbine Engines: Operational Experience, Fundamental
Mechanisms, and Modeling, Progress in Astronautics and Aeronautics, 2005: p. 43-63, 163-175.
4. Lieuwen, T.C., Yang, V., Monitoring Combustion Instabilities: E.ON UK's Experience. Combustion instabilities in gas
turbine engines : operational experience, fundamental mechanisms and modeling, 2005: p. 163-175.
5. Mcmanus, K.R., Vandsburger, U., Bowman, C.T., Combustor performance enhancement through direct shear layer
excitation Combustion and Flame, 1990. 82(1): p. 75-92.

13
American Institute of Aeronautics and Astronautics
6. Richards GA, Y.M., Robey EE, Cowell LL, Rawlins DD., Combustion Oscillation Control by Cyclic Fuel Injection. J.
Eng. Gas Turbines Power, 1997. 19(2): p. 340-343.
7. Gopakumar, R., Mondal, S., Paul, R., Mahesh, S., Chaudhuri, S., Mitigating instability by actuating the swirler in a
combustor Combustion and Flame, 2016. 165: p. 361-363.
8. Lang, W., Poinsot, T., Candel, S., Active control of combustion instability. Combustion and Flame, 1987. 70: p. 281-289.
9. Candel, S.M., Combustion instabilities coupled by pressure waves and their active control. 24th Symposium on
Combustion, 1992. 24(1): p. 1277-1296.
10. Culick, F.E.C., Combustion instabilities in liquid-fueled propulsion systems – an overview. AGARD Conference
Proceedings, 1988. 450.
11. Darema F. Dynamic data driven applications systems: new capabilities for application simulations and measurements.
In: 5th international conference on computational science—ICCS 2005, Atlanta, GA, 2005, pp.610-615.

12. Gotoda, H., Amano, M., Ikawa, T. et al, Characterization of complexities in combustion instability in a lean premixed
gas-turbine model combustor. Chaos, 2012. 22(043128).
Downloaded by PENNSYLVANIA STATE UNIVERSITY on February 27, 2019 | http://arc.aiaa.org | DOI: 10.2514/6.2017-4892

13. Nair, V., Thampi, G. and Sujith, R, Intermittency route to thermoacoustic instability in turbulent combustors. Journal of
Fluid Mechanics, 2014. 756: p. 470-487.
14. Sarkar, S., Chakravarthy, S.R., Ramanan, V. and Ray, A., Dynamic data-driven prediction of instability in a swirl-
stabilized combustor. International Journal of Spray and Combustion Dynamics, 2016, doi:
10.1109/TCYB.2015.2508024.
15. Ray, A., Symbolic dynamic analysis of complex systems for anomaly detection, Signal Process, 2004. 84: p. 1115–1130.
16. Mukherjee, K. and Ray, A. State splitting and state merging in probabilistic finite state automata for signal representation
and analysis. Signal Processing, 2014, 104: p. 105–119.

17. Zhang, J., Small, M., Complex Network from Pseudoperiodic Time Series: Topology versus Dynamics. Physical Review
Letters, 2006. 96,238701.
18. Lacasa, L., Luque, B., Ballesteros, F., et al, From time series to complex networks: The visibility graph. Proceedings of
the National Academy of Sciences, 2008. 105(13): p. 4972-4975.
19. Lacasa, L., Nicosia, V. and Latora, V., Network structure of multivariate time series. Scientific Reports, 2015. 5(15508).
20. Murugesan, M., Sujith,R., Combustion noise is scale-free: transition from scale-free to order at the onset of
thermoacoustic instability. Jornal of Fluid Mechanics, 2015. 772: p. 225-245.
21. Fischera, A., Losenno, C., Pagano, A., Experimental analysis of thermo-acoustic combustion instability. Applied Energy,
2001. 70: p. 179-191.
22. Oppenheim, A.V., Schafer, R.W., Discrete-Time Signal Processing. 1999. 2nd edition: p. 60.
23. Yi, T., Gutmark, J., Real-Time Prediction of Incipient Lean Blowout in Gas Turbine Combustors. AIAA Journal, 2007.
45(7).

24. Nunez, A.M., Lacasa, L., Gomez, J.P. et. al, Visibility Algorithms: A Short Review. New Frontiers in Graph Theory, Dr.
Yagang Zhang (Ed.), 2012
25. Murugesan, M., Sujith,R., Detecting the Onset of an Impending Thermoacoustic Instability Using Complex Networks.
Journal of Propulsion and Power, 2016. 32(3).

26. McDonough, R.N., Whalen, A.D., Detection of Signals in Noise. Academic Prses, 1995. 2nd Edition: p. 152-155.

14
American Institute of Aeronautics and Astronautics

You might also like