Content uploaded by Yuvaraja Visagathilagar
Author content
All content in this area was uploaded by Yuvaraja Visagathilagar on Aug 22, 2017
Content may be subject to copyright.
PROCEEDINGS OF SPIE
SPIEDigitalLibrary.org/conference-proceedings-of-spie
Nuisance alarm suppression
techniques for fibre-optic intrusion
detection systems
Seedahmed S. Mahmoud
Yuvaraja Visagathilagar
Jim Katsifolis
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
Nuisance alarm suppression techniques for fibre-optic intrusion
detection systems
Seedahmed S. Mahmoud, Yuvaraja Visagathilagar and Jim Katsifolis
Future Fibre Technologies Pty Ltd.
10 Hartnett Close, Mulgrave, VIC 3170, Australia
ABSTRACT
The suppression of nuisance alarms without degrading sensitivity in fibre-optic intrusion detection systems is important
for maintaining acceptable performance. Signal processing algorithms that maintain the POD and minimize nuisance
alarms are crucial for achieving this. A level crossings algorithm is presented for suppressing torrential rain-induced
nuisance alarms in a fibre-optic fence-based perimeter intrusion detection system. Results show that rain-induced
nuisance alarms can be suppressed for rainfall rates in excess of 100 mm/hr, and intrusion events can be detected
simultaneously during rain periods. The use of a level crossing based detection and novel classification algorithm is also
presented demonstrating the suppression of nuisance events and discrimination of nuisance and intrusion events in a
buried pipeline fibre-optic intrusion detection system. The sensor employed for both types of systems is a distributed
bidirectional fibre-optic Mach Zehnder interferometer.
Keywords: Event classification, fibre-optic sensor, high traffic volume, intrusion detection, nuisance alarms
1. INTRODUCTION
In recent years, high performance distributed fibre-optic sensors have been applied to both outdoor and buried intrusion
detection systems. The success of any intrusion detection system depends on three important performance parameters:
the probability of detection (POD), the nuisance alarm rate (NAR), and the false alarm rate (FAR). The POD provides an
indication of a system’s ability to detect an intrusion. A nuisance alarm is any alarm which is not generated by an event
of interest. Nuisance alarms are typically generated by environmental sources such as rain, wind, snow, wildlife and
vegetation, as well as man-made sources such as traffic crossings, industrial noises and other ambient noise sources.
Since increasing the sensitivity of an intrusion detection system to achieve a high POD will also increase its sensitivity to
nuisance events, signal processing algorithms that maintain the POD and eliminate or minimize nuisance alarms are
crucial in perimeter intrusion detection systems1. The most important challenge for a distributed fibre-optic intrusion
detection system is to minimize the NAR without compromising the system sensitivity or POD for a wide range of
operating environments. The minimization of the NAR of any intrusion detection system, and indeed of any sensing
system, is therefore critical for its successful performance and confidence of operation.
A number of sensor related signal processing algorithms have been presented in the literature for suppressing nuisance
alarms. Jiang et al2 proposed a classification method for a MZ interferometric sensor using wavelet packet transform for
denoising and feature extraction and implemented a neural network as a classifier. This method however would not be
suitable to eliminate nuisance alarms due to torrential rain as the sensor’s signal amplitude would be saturated in the time
domain. Vries3 proposed an acoustic based perimeter intrusion classification system that deploys a neural network with
frequency domain features to detect different types of intrusion events such as climbing, cutting and jumping over the
fence. The system however suffers from performance degradation when the quality of the sound (SNR) generated by the
intruders and the surrounding environment decreases. Moreover, the frequency domain features are not robust enough to
distinguish between nuisance and intrusion events.
smahmoud@fftsecurity.com; phone +61 3 9590-3100; fax +61 3 9560-8000; www.fftsecurity.com
Third Asia Pacific Optical Sensors Conference, edited by John Canning, Gangding Peng,
Proc. of SPIE Vol. 8351, 83513J · © 2012 SPIE · CCC code: 0277-786X/12/$18 · doi: 10.1117/12.915950
Proc. of SPIE Vol. 8351 83513J-1
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
Yousefi et al4 presented a fence breach detection system which can detect activity on the fence and discriminate different
types of activity. The hardware of the system comprises a 3-axis accelerometer and a RISC microprocessor. The system
employs an algorithm that detects activity and non-activity on the fence. It also recognizes the type of breach whether it
is due to rattling caused by strong wind or a person climbing on the fence. This system uses signal variation features
along with the energy of two bandpass filters to separate the rattle and climb frequency components. While this shows
some success, it is not possible to discriminate between classes that have a similar impact on a fence. Moreover, this
algorithm is used to classify a small number of classes (limited to two classes). Min et. al5 proposed a real-time
monitoring system using an audio sensor to detect abnormal activities in the vicinity of buried gas pipes. They extracted
a frequency domain feature using a nonlinear scale filter bank method and cepstral mean subtraction along with a
combination of two classifiers using a Gaussian mixture model and multi-layer perceptron. Their system achieved a 92%
detection rate to abnormal activities such as hammer drilling and digging. The detection rate of intrusion was however
degraded in the presence of background noise such as traffic in the vicinity of the sensor.
In this paper, signal processing algorithms are presented for suppressing nuisance alarms in both outdoor fence-mounted
and buried fibre-optic intrusion detection systems. The use of a real-time level crossing algorithm to suppress rain-
induced nuisance alarms and discriminate between continuous nuisance and non-continuous intrusion events in perimeter
intrusion detection systems is presented. The use of a level crossing based detection method and novel classification
algorithm is also presented for the suppression and discrimination of nuisance events from intrusion events in a buried
intrusion detection system. Results are shown from real systems.
2. FIBRE OPTIC INTRUSION DETECTION SYSTEM
The intrusion detection system used in this work is based on the Future Fibre Technologies Microstrain Locator
technology as applied to fence perimeter and pipeline security applications6. The Microstrain Locator is based on the use
of a bidirectional Mach Zehnder Interferometer (MZI) as a distributed sensor to detect and locate an intrusion anywhere
along the sensing length, LS, as shown in Fig 1. The two sensing fibers and the lead out fibre are typically housed in a
standard single mode fibre-optic cable which is mounted on the perimeter fence. In this paper, the intrusion detection
system will be referred to as the Locator.
Fig. 1. A basic FFT Microstrain Locator system using a bi-directional MZ with input polarisation control.
C1 – C5 are all 50:50 fibre couplers.
The deployed sensing system consists of an industrial computer which houses a highly coherent 1550nm laser source
which injects continuous wave counter-propagating light into the MZ. Two detectors, also housed in the sensing
controller, receive the clockwise (CW) and counter-clockwise (CCW) signals from the MZ to analyze the signals. The
sensing controller also includes polarization controllers PCCW and PCCCW to maximize the MZ’s fringe visibility and
optimize the location accuracy by actively compensating for changes in fibre birefringence. Detection of an event is
based on analyzing the interferometric signals, while an event’s location along the sensing length LS is resolved by
measuring the time difference between received counter propagating signals. Additionally, using the event signals
detected by both detectors it is possible to apply the appropriate signal processing techniques to classify the signals and
Proc. of SPIE Vol. 8351 83513J-2
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
perform both signal recognition and signal discrimination. The installation of a fence-mounted fibre-optic locator system
is shown in Fig. 2.
Fig. 2: A fence-mounted fibre-optic FFT Microstrain Locator system. (SS=start sensor and ES= end sensor).
In buried systems, the sensing cable is typically buried next to the pipeline to detect third party interference (TPI)
activities as shown in Fig. 3. Inevitably it will also be sensitive to other non-intrusion events such as those from nearby
traffic and railway crossings.
Fig. 3 Cross section of a buried fibre-optic intrusion detection system for detecting third party interference.
The quality of installation of the fibre-optic sensing cable in both fence systems and buried systems is critical for
achieving optimal system performance and is very often under-estimated. For example, in fence systems, the fence
construction needs to be built according to the relevant standards and the sensing cable attached correctly. Poor fence
construction and sensor cable attachment contribute to excessive nuisance alarms and long term performance
degradation. While good installation practices will not eliminate nuisance alarms, it will ensure that excessive nuisance
signals are not generated due to hypersensitivity of the fence to environmental noise, and enable optimal performance of
any nuisance mitigation algorithms employed.
3. SUPPRESSION OF RAIN-INDUCED NUISANCE ALARMS IN FENCE SYSTEMS
All perimeter intrusion detection systems (PIDS) will inevitably experience nuisance events. In the case of fence based
PIDS this includes, torrential rain, strong winds, adjacent vehicular traffic, and other man made sources. The ideal way to
deal with these signals is to be able to classify them and make a decision about their nature without compromising
system sensitivity. In this section, a nuisance mitigation technique using a level crossings algorithm will be discussed.
Proc. of SPIE Vol. 8351 83513J-3
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
3.1 Level crossings based nuisance mitigation
Suppression of continuous nuisance alarms such as those induced by torrential rain is one of the most challenging tasks
for outdoor PIDS. A real-time level crossings (LC) algorithm to mitigate rain-induced nuisance alarms in fence based
fibre-optic intrusion detection systems was proposed and implemented7. This algorithm is computationally non-intensive
and it can be used to eliminate rain-induced nuisance alarms for torrential rainfall rates up to and in excess of 100 mm/hr.
The LC-based algorithm is also used to discriminate between continuous nuisances such as rain and non-continuous
intrusion events, which allows for simultaneous detection of intrusion events. The algorithm also employs a dynamic
event threshold to be able to automatically adjust to varying rainfall rates.
The LC algorithm can be defined by the number of level crossings, in the positive direction, of an input vector through a
given threshold and can be expressed as7,8:
()( ){}
∑
−
=<−≥Ψ= 1
0)1(&)(
N
nthreshnxthreshnxLC , (1)
where x is a signal of length N, the parameter thresh is the level threshold, and the indicator function Ψ is 1 if its
argument is true, or 0 otherwise. By applying Equation 1 to event signals received by the fibre-optic based Locator
system described in Figure 1, a level crossings representation of the detected signal can be formed against blocks of time.
Based on this representation, a number of features can be extracted to suppress nuisance alarms and discriminate
between nuisance and intrusion events.
3.2 Fence System Results
The LC algorithm was integrated into the Locator sensing controller which has been installed in numerous sites
worldwide that experience torrential rainfall. These sites experience rainfall rates up to and in excess of 100 mm/hour (>
4 inches/hr). Results from these sites have demonstrated the elimination of rain-induced nuisance alarms with the
simultaneous detection of intrusion events. The LC-based algorithm is used to discriminate between continuous
nuisances such as rain and non-continuous intrusion events. Due to its continuous nature, torrential rainfall will generate
a fairly consistent level-crossing rate per time block period. This feature can be used to suppress rainfall induced alarms
from the system. By monitoring for any changes in the level crossing rate, non-continuous intrusion events such as fence
climbing or cutting can be detected during the rainfall period. Using a dynamic intrusion event threshold has also proven
to be effective in automatically adjusting to variable rainfall rates. Figure 4 shows an example of the detected torrential
rain signal on a 3.2 km long chain link fence perimeter. The LC representation (see inset Nuisance Level graph) can be
used to detect intrusions which are buried inside the rain signal. The LC algorithm can also be adapted to deal with other
continuous or semi-continuous nuisances such as nearby traffic noise in a similar way.
Fig. 4 Real-time elimination of rain-induced nuisance alarms with simultaneous intrusion detection on a 3.2 km chain link fence.
Proc. of SPIE Vol. 8351 83513J-4
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
4. NUISANCE SUPPRESSION IN BURIED SYSTEMS
Buried-fibre-optic sensors, such as those implemented for protecting buried oil and gas pipelines as well as buried
communications links, are designed to detect physical disturbances generated by third party interference (TPI) which
includes accidental or deliberate tampering, digging or excavation activities. These are susceptible to a range of ground
based nuisance events such as road and railway traffic and other nearby construction activities. These events can reduce
an intrusion detection system’s effectiveness with an unacceptably high rate of nuisance alarms. In this section, some
novel signal processing techniques are outlined to mitigate the effect of these nuisance events on buried intrusion
detection systems by suppressing particular nuisance induced alarms without affecting alarms generated by intrusion
events of interest. Figs. 5a and 5b show detected signal examples of typical intrusion events that should be detected,
while Figs. 6a and 6b show typical signals of nuisance alarms due to traffic that should be rejected on a 2.7 km buried
gas pipeline.
00.5 11.5 22.5
x 10
4
-5
-4
-3
-2
-1
0
1
2
3
4
5
No. of Sample s
Amplitude
Original signal
00.5 11.5 22.5
x 10
4
-5
-4
-3
-2
-1
0
1
2
3
4
5
No. of Samples
Amplitude
Original signal
(a) (b)
Fig. 5 (a) Intrusion signal caused by digging with a pick-axe above a 2.7 km buried gas pipeline protected by a locator system. (b)
Intrusion signal caused by digging with a back-hoe above a 2.7 km buried gas pipeline protected by a locator system.
00.5 11.5 22.5
x 10
4
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
No. of Sample s
Amplit ude
Original signal
00.5 11.5 22.5
x 10
4
-6
-4
-2
0
2
4
6
No. of Samples
Amplitude
Original signal
(a) (b)
Fig. 6 (a) Nuisance signal from traffic on a nearby road for a 2.7 km gas pipeline intrusion detection system. (b) Strong periodic
nuisance signal from a railway crossing for a 2.7 km gas pipeline intrusion detection system. The railway runs perpendicularly over
the pipeline.
4.1 Pre-processing and feature extraction of the proposed solution
The novel nuisance alarm suppression algorithm described herein consists of an event detection portion, a feature
extracted portion for extracting a number of appropriate signal features from the time domain representation of the
signals, and a simple decision tree classifier. Event detection is based on the previously mentioned LCs algorithm. Fig. 7
shows the pre-processing and feature extraction stages of the nuisance suppression algorithm.
A number of extracted features are generated as shown in Fig. 7. Only three of these features will be considered for
nuisance alarm suppression. These features are described below:
1- Continuity of the signal: This is a measure of how continuous the signal is over its duration. It is determined
by using the maximum amplitude versus segment information from the pre-processing stage (see Fig. 7) and can
be given by.
Proc. of SPIE Vol. 8351 83513J-5
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
(
)
segmentsofNo.
2
threshamplitude with segmentsofNo.
Continuity >
= (2)
The parameter thresh2 is normally set above the system noise of the time domain signal as in the case of LCs.
The maximum possible continuity is unity.
2- Maximum amplitude strength (MAS) (count %): To evaluate the MAS feature, first the amplitude strength
of each segment needs to be measured. The amplitude strength relates to how much of a signal is above a given
amplitude threshold, Thresh1, and is defined by equation 3. The parameter Thresh1 is normally application
dependent. After evaluating the amplitude strength using equation 3, the MAS feature is calculated as the
maximum value calculated by equation 3 over the whole duration of the signal (see Fig. 7). It is effectively a
measure of what percentage of a signal is above a given threshold value and is given as a percentage value. This
feature is important for distinguishing digging events from traffic nuisances that have similar continuity values.
The intrusion signals will typically have higher maximum amplitude strengths.
(
)
()
100
segmentspecifiedtheinsamplesofsamplestotal
1
ThreshsegmentspecifiedtheinsamplesofNo.
segmenteachofstrength amplitudeThe ∗
>
= (3)
3- Maximum deviation (MD): The first step towards the evaluation of MD is by the evaluation of the maximum
amplitude in each segment as is the case of the continuity feature (see Fig. 7). The MD is then calculated by
subtracting the mean of the segment’s amplitudes from the maximum segment’s amplitude.
Maximum deviation = Maximum segment amplitude – Mean segment amplitude (4)
This feature is important to discriminate between digging intrusion events and adjacent nuisance events of
comparably long continuities, even if they have roughly similar amplitudes. In this situation the digging events
will have a higher maximum deviation owing to their higher variation in segment maxima. This can be seen by
comparing the two signals represented by Fig. 5b (intrusion) and Fig. 6b (adjacent nuisance) where there are
more periods of inactivity in the digging signal (lower in amplitude) when compared with the continuous
nuisance signal. This translates into a higher maximum deviation for the digging signal.
Fig. 7 New alarming system feature extraction method
Proc. of SPIE Vol. 8351 83513J-6
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
4.2 Classification Using Simple Decision Tree
By using the features described in the previous section in the right combination it is possible to suppress a large number
of nuisance alarms in buried intrusion detection systems. This can be done by implementing a decision tree. Decision
trees represent a series of IF…THEN type rules which are linked together and can be used to classify or predict events
based upon the values of a select number of features. For this work we use a simple decision tree to discriminate between
intrusion and nuisance events. Neural networks can also be used with these features to discriminate between true alarms
and nuisance events. Intrusion events will generate alarms while nuisance events will be ignored.
Practical data from two sites have been tested using the proposed algorithm. Table 1 shows the values of the features for
the intrusion and nuisance events in Fig. 5 and 6, respectively. The values in Table 1 are used to derive appropriate
threshold values, Thresh_1, Thresh_2, Thresh_3, and Thresh_4 for the decision tree as shown in Fig. 8. The algorithm
successfully classified digging events accurately (Fig. 5a and 5b) while traffic nuisances (such as those in Fig. 6a and 6b)
were rejected.
Fig 8 Practical example of the decision tree
Table 1: Feature values of intrusion and nuisance events.
Features Values
Event Type Continuity MAS MD Node decision
Intrusion (hand and
assisted digging, Fig. 5b)
0.4 34.86 3.595 Node-2
Vehicular traffic (adjacent
traffic, Fig. 6a)
1 1.86 0.76 Node-3
Road intersection and
train crossings, Fig. 6b
1 18.99 0.4712 Node-3
Proc. of SPIE Vol. 8351 83513J-7
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
5. CONCLUSIONS
A number of signal processing techniques for suppressing nuisance alarms in a fibre-optic intrusion detection system
have been proposed and presented. The use of a level crossings based algorithm for suppressing torrential rain-induced
nuisance alarms in fence-based fibre-optic perimeter intrusion detection systems has demonstrated its effectiveness
against torrential rainfall rates in excess of 100 mm//hr. It has also demonstrated the successful simultaneous detection of
intrusion events during rainfall periods. A level crossing based detection and novel classification algorithm has also been
applied to buried fibre-optic pipeline intrusion detection systems. The use of a decision tree classification algorithm has
demonstrated the effective classification of both traffic induced nuisance events and digging and excavation intrusion
events. Future work is focusing on increasing the library of signal features to achieve the classification of more intrusion
and nuisance events.
REFERENCES
[1] Tarr, S., and Leach, G., “The dependence of detection system performance on fence construction and detector
location,” Proceedings of the International Carnahan Conference on Security Technology, pp. 196-200, (1998).
[2] Jiang, L., Liu, X., Zhang, F., “Mutli-target recognition used in airport fibre fence warning system”, Proceedings of
the Ninth International Conference on Machine Learning and Cybernetics, pp. 1126- 1129 Qingdao, (2010).
[3] J. De Vries, “A low cost fence impact classification system with neural networks,” in IEEE Africon, pp. 131-136,
Vol. 2, (2004).
[4] A. Yousefi, A. A. Dibazar, and Theodore Berger, “Intelligent Fence Intrusion Detection System: Detection of
Intentional Fence Breaching and Recognition of Fence Climbing,” IEEE International Conference on Technologies
for Homeland Security, Boston, (2008).
[5] Min, H., Lee, C., Lee, J. and Park, C. H., “Abnormal signal detection in gas pipes using neural networks”, 33rd
Annual Conference of the IEEE Industrial Electronics, pp. 2503-2508, Taiwan, 2007.
[6] Katsifolis, J. and McIntosh, L., “Apparatus and method for using a counter-propagating signal method for locating
events”, Patent No US 7,499,177 (2009).
[7] Mahmoud, S. and Katsifolis, J.,”Elimination of rain-induced nuisance alarms in distributed fibre-optic perimeter
intrusion detection systems,” Proceedings of SPIE, vol. 7316, Paper 7316-3, (2009).
[8] Mahmoud, S. and Katsifolis, J., “Robust event classification for a fibre-optic perimeter intrusion detection system
using level crossing features and artificial neural networks,” Proceedings of SPIE, vol. 7677, Paper 6, (2010).
Proc. of SPIE Vol. 8351 83513J-8
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 8/21/2017 Terms of Use: https://spiedigitallibrary.spie.org/ss/TermsOfUse.aspx
