Hybrid Feature Extraction-Based Intrusion ... - IEEE Xplore

Open Access

Hybrid Feature Extraction-Based Intrusion Discrimination in Optical Fiber Perimeter Security System Volume 9, Number 1, February 2017 Xiangdong Huang Haojie Zhang Kun Liu Tiegen Liu Yuedong Wang Chunyu Ma

DOI: 10.1109/JPHOT.2016.2636747 1943-0655 © 2016 IEEE

IEEE Photonics Journal

Hybrid Feature Extraction-Based Intrusion Discrimination

Hybrid Feature Extraction-Based Intrusion Discrimination in Optical Fiber Perimeter Security System Xiangdong Huang,1 Haojie Zhang,1 Kun Liu,2 Tiegen Liu,2 Yuedong Wang,1 and Chunyu Ma2 1 School of Electronic Information Engineering, Tianjin University, Tianjin 300072, China 2 College of Precision Instrument and Optoelectronics Engineering, Tianjin University,

Tianjin 300072, China DOI:10.1109/JPHOT.2016.2636747 C 2016 IEEE. Translations and content mining are permitted for academic research only. 1943-0655 Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Manuscript received October 20, 2016; revised November 28, 2016; accepted December 4, 2016. Date of publication December 7, 2016; date of current version December 29, 2016. This work was supported in part by the National Natural Science Foundation of China under Grant 61671012 and Grant 61475114 and in part by the National Instrument Program under Grant 2013YQ030915. Corresponding author: K. Liu (e-mail: [email protected]).

Abstract: This paper proposes a hybrid feature extraction-based intrusion discrimination scheme for an optical fiber perimeter security system, which concurrently possesses high classification rate and high efficiency. The high classification rate lies in two aspects: On one hand, plentiful contents (including bandwidth segmentation in frequency domain, kurtosis in statistics, and the zero-crossing rate in time domain) are incorporated into the proposed hybrid feature vector; on the other hand, a configurable filter bank is adopted to reduce the intercoupling between features in the hybrid vector. The high efficiency also arises for two reasons: For one thing, the configurable filter bank works in a pipeline stream; for another, an efficient support vector machine is employed to classify hybrid vectors. Experiments demonstrated that the proposed scheme can accurately identify four common intrusions (fence climbing, knocking the cable, waggling, and fence cutting) with an average recognition rate higher than 94%. Moreover, the recognition efficiency is also high. Index Terms: Filter bank, hybrid feature, intrusion discrimination, optical fibers.

1. Introduction The dual Mach-Zehnder interferometry (DMZI) vibration system [1], [2] adopts the phase-modulation fiber sensing technique and thus possesses the superiority of high sensitivity and fast response, which has been applied in various perimeter security fields, such as submarine cable security [3], pipeline leakage detection [4], and airport guarding [5]. Perimeter security of DMZI system involves several critical techniques including intrusion positioning [6], [7], endpoint detection [8] and intrusion discrimination. Up to now, due to the difficulty of feature description for various kinds of invasion signals, the technique of intrusion discrimination is still immature. Specifically, the existing schemes of intrusion discrimination failed to concurrently possess high classification rate and high efficiency. Mahmoud et al. [9] proposed a robust classification scheme consisting of a level crossing based feature extractor and a supervised neural network for Mach-Zehnder interferometer system. This method can discriminate four kinds of events by means of mining some distinctive features. However, the classification rate heavily relies on a threshold which can hardly be properly specified in practical

Vol. 9, No. 1, February 2017

7800212



Fig. 1. Schematic diagram of DMZI vibration sensor. DAQ: Data Acquisition; IPC: Industrial Personal Computer; C1, C2, C3: 3dB fiber coupler; PD1, PD2: Photo-detector.

applications. Liu et al. [10] proposed a wavelet decomposition based scheme, which utilizes the signal energies distributed over different frequency bands as the feature vector to recognize three kinds of intrusions. However, this method is not efficient, because wavelet decomposition needs to half-split the frequency bands in multiple levels (i.e., it cannot work in a parallel pipeline mode). Following this, a high-accuracy scheme combining empirical mode decomposition (EMD) with the RBF neural network was proposed in [11], in which features are extracted from several intrinsic mode functions (IMFs) by means of EMD. Nevertheless, in order to adapt the desired IMFs to the given signal, EMD has to work in an iterative mode and thus consumes excessive computation. From the above intrusion discrimination schemes, one can find that, their feature vectors seem to lack abundant information. Specifically, each feature vector only involves the information of one or two aspects, thereby degrading the classification accuracy. Moreover, these algorithms of feature extraction generally experience complex processes (iterative update or multi-level decomposition), thus influencing the efficiency. To overcome the contradiction between recognition accuracy and efficiency, this paper proposes a hybrid feature vector based intrusion discrimination scheme. Through incorporating multiple aspects of contents (including bandwidth segmentation in frequency domain, kurtosis in statistics, the zero-crossing rate in time domain) into the feature vector, we can obtain a more accurate intrusion description than the aforementioned schemes, thereby enhancing the classification rate. Furthermore, our feature vectors are generated in a pipeline mode by means of a configurable closed-form filter bank, which improves the classification efficiency together with the support vector machine (SVM) classifier. Field experiments verified that, our proposed scheme can accurately and efficiently identify four common intrusions: fence climbing, knocking the cable, waggling and fence cutting. This paper is organized as follows. In Section II, we give a brief description of DMZI system. In Section III, we give the dataflow of the proposed scheme. In Section IV, we elaborate the principle of this scheme. In Section V, we compare the event discrimination experiments between the proposed scheme and the EMD based scheme. In Section VI, we have some discussions. In Section VII, we come to some conclusions.

2. DMZI Optic Fiber Perimeter Security System Fig. 1 illustrates the structure of DMZI vibration system: At coupler C1, the output of the laser with narrow linewidth is split equally through an isolator, and launched into a dual Mach-Zehnder Interferometer consisting of coupler C2 and coupler C3. Two light beams then propagate oppositely in clockwise (CW) and counter-clockwise (CCW) directions and interfere at their counterpart coupler (C3 or C2). The interference outputs are detected by PIN diodes PD1 and PD2. Then, the output signals of these PIN diodes are collected by two Data Acquisition (DAQ) cards, which are respectively specified with distinctive sampling rates for various purposes. Specifically, the D A Q 1 is used for


7800212



Fig. 2. Dataflow of the proposed intrusion discrimination scheme.

endpoint detection (i.e., determine the starting moment of the intrusions) and event discrimination [11], [12], and D A Q 2 is used for intrusion positioning. Finally, these desired functions are realized by running some signal processing algorithms implanted in the industrial personal computer (IPC).

3. General Description of the Proposed Scheme 3.1. The Data Flow As shown by Fig. 2 (z −1 refers to a delayer), the scheme of the intrusion discrimination mainly consists of 3 stages: filter-bank based pre-processing, constructing the hybrid feature vector and SVM pattern recognition, which are elaborated as follows. Initialization: Implement FFT analysis on multiple realizations of non-invasive signals to determine the environment’s cut-off frequency f e . Then, use f e and an upper intrusion frequency f u to configure Q sub-filters h 0 , . . . , h Q −1 . Stage 1 (filter-bank based pre-processing): Add the 2N − 1 input samples x(n + N − 1), . . . , x(n − N + 1) in pairs to generate an N -length sequence [x(n), x(n + 1) + x(n − 1), . . . , x(n + N − 1) + x(n − N + 1)]; then, this sequence is weighted by an N -length window [w c (0), w c (1), . . . , w c (N − 1)] to obtain an N -length vector v = [v 0 (n), v 1 (n), . . . , v N −1 (n)]; finally, calculate the Q outputs y 0 (n), . . . , y Q −1 (n) through implementing inner product between v and Q tap coefficient vectors h q = [h q (0), h q (1), . . . , h q (N − 1)], q = 0, . . . , Q − 1. Stage 2 (Constructing the hybrid feature vector): Calculate the kurtosis values K˜ 0 , . . . , K˜ Q −1 of y 0 (n), . . . , y Q −1 (n) and acquire their normalized versions K 0 , . . . , K Q −1 . Calculate the zero crossing rate ZCR of the input signal x(n). Then, construct a hybrid (Q + 1)-length feature vector as F = [K 0 , . . . , K Q −1 , Z CR ]. Stage 3 (SVM pattern recognition): Train a SVM classifier with the feature vectors of labeled intrusion signals. Then, feed the feature vectors of unknown categories into this trained SVM to recognize the types of intrusions. From Fig. 2, one can see that the construction of the feature vector F is forward and follows a pipeline mode (i.e., no complex iterations of parameter update are involved), which enhances the efficiency of the proposed scheme.


7800212



3.2. Considerations of the Hybrid Feature Vector To achieve both high classification rate and high efficiency, the following three points are taken into consideration. 1) To ensure a high recognition efficiency, the length of the feature vector should not be too long. 2) To ensure a high classification rate, this feature vector should incorporate multiple aspects of information. Furthermore, to shorten the vector length as possible, a vector entry allows to carry 2 aspects of information. 3) To further highlight features, the intercoupling among vector entries should be as low as possible. From the dataflow of Fig. 2, one can find that our feature vector F = [K 0 , . . . , K Q −1 , Z CR ] directly meets the second requirement. Specifically, apart from the kurtosis of statistics, the entries K 0 , . . . , K Q −1 also carry frequency-domain information, since each of them derives from a sub-filter of a particular frequency band. Hence, together with the time-domain counting value Z CR , the vector F actually integrates 3 aspects of features. Whether the vector F meets the other 2 requirements depends on the performance of the filter bank in Fig. 2.

4. Principle of the Proposed Intrusion Discrimination Method 4.1. Filter Bank based Pre-processing As Fig. 2 illustrates, this stage is based on a closed-form filter bank, which divides the input x(n) into Q sub-signals y 0 (n), . . . , y Q −1 (n) over different frequency channels. As long as the inter-channel interference (ICI) of this filter bank is sufficiently small (i.e., the intercoupling among K 0 , . . . , K Q −1 is low enough), these Q kurtosis values will independently reflect the statistical information of their particular frequency bands. In return, this allows to use a short feature vector. Ref. [12] pointed out that, for the filtering structure in Fig. 2, the ICI of the filter bank is closely related to the selection of the window w c (n) in Fig. 2. In particular, if w c (n) is constructed by convolving a commonly-used N -length window f (n) with its reversed version, i.e., w c (n) = f (n) ∗ f (−n) ,

− N +1 ≤ n ≤ N − 1

(1)

then each sub-filter acquires a large stop-band attenuation, which suppresses the ICI of the entire filter bank. What’s more, the filter bank’s ICI is also related to the Q tap coefficient vectors hq in Fig. 2. Ref. [12] proved that, hq is no more than the inverse Discrete Fourier Transform (IDFT) of a N -length frequency sampling vector hq . Hence, through setting an individual entry of the vector hq with the value 1 or 0, one can acquire a sub-filter with the desired transfer performance. For the DMZI perimeter security system, the desired transfer performance contains two meanings: 1) Each hq should exclude the influence of the environmental interferences; 2) Passband overlappings among H 0 , . . . , H Q −1 should be avoided. Based on these two considerations, the format of hq , q = 0, . . . Q − 1, is set as hq = [ 0, . . . , 0 0,.., 0 1, . . . , 1 0, . . . , 0 1, . . . , 1 0, . . . 0 0, . . . 0 ] e

mq

m

r

m

mq

(2)

e−1

where r = N − (2mq + 2m + 2e − 1). In (2), each entry occupies an analog frequency bandwidth f = f s /N , where f s is the sampling rate of DAQ1 in Fig. 1. As is known, the energy of the environmental interference is mainly distributed in a low-frequency region [8] (i.e., f ∈ [0, f e ]). Hence, once the cut-off frequency f e is determined by FFT (as Fig. 2 depicts), the integer e in (2) should be configured as e = [f e /f ] = [N f e /f s ]

(3)

where the operator “[·]” represents the round operation.


7800212



Fig. 3. Transfer curves of filter bank. (a) Transfer curves of the proposed filter bank. (b) Transfer curves of classical frequency sampling method based filter bank.

To determine the integer m in (2), the upper frequency f u of the common invasive actions needs to be specified a priori. Therefore, for the purpose of uniformly dividing the intrusion bandwidth B = f u − f c into Q subbands, the integer m should be configured as (f u − f e )N B . (4) = m= Q · f Q · fs With the above parameters configured, each sub-filter is a bandpass filter with the bandwidth mf . For the q-th sub-filter, its ideal passband falls in the following range f ∈ [(qm + e)f, (qm + e + m)f ]. Further, one can derive the IDFT h q (n) of H q in (2) as ⎧ 2cos[πn(2mq + 2e + m − 1)/N ] sin(mπn/N ) ⎪ ⎪ ⎪ ⎪ ⎪ N sin(πn/N ) ⎪ ⎨ n ∈ [−N + 1, 1] ∪ [1, N − 1] h q (n) = ⎪ ⎪ ⎪ ⎪ ⎪ 2m ⎪ ⎩ , n =0 N

(5)

(6)

Combining (3)–(6), one can find that every coefficient h q (n) in Fig. 2 can be easily configured in a closed-form mode, thereby enhancing the efficiency and the flexibility. Ref. [12] proved that, the q-th filter is equivalent to a (2N − 1)-length FIR filter g q whose tap coefficients are g q (n) = w c (n)h q (n),

− N +1 ≤ n ≤ N − 1.

(7)

For example, specify N = 256, the sampling rate f s = 10 kHz, the number of sub-filters Q = 4, the environmental cut-off frequency f e = 250 Hz, the upper frequency f u = 3200 Hz (i.e., e = 6 and m = 19) and select the window f as a Hamming window. Then, substituting these parameters into (1) ∼ (6) yields 4 sub-filters g 0 (n), . . . , g 3 (n), whose transfer curves |G 0 (j2πf )|, . . . , |G 3 (j2πf )| are plotted in Fig. 3(a). As a contrast, the transfer curves |H 0 (j2πf )|, . . . , |H 3 (j2πf )| generated by the classical frequency sampling filter design [13] are given in Fig. 3(b). As Fig. 3 illustrates, for the the classical frequency sampling method, large ripples appear in both passbands and stopbands. In contrast, for our proposed filter bank design, each subfilter’s passband curve is much flatter and its stopband ripples are almost negligible. Hence,


7800212



the ICI is greatly suppressed, indicating that the filter bank’s outputs y 0 (n), . . . , y Q −1 (n) are separated in independent frequency bands. Thus, the intercoupling among K 0 , . . . , K Q −1 is greatly inhibited. In addition, from Fig. 3(a), one can see that, in the region where the frequency is lower than environmental cut-off frequency 250 Hz, for any individual channel, |G q (j2πf )| nearly equals zero. This implies that the background environmental interference is thoroughly removed, thereby highlighting the features of intrusion signals. 4.2. Constructing the Hybrid Feature Vector Constructing the feature vector is crucial for the subsequent pattern recognition. Recall that our hybrid feature extraction characterizes an intrusion signal with a (Q + 1)-length vector, which contains comprehensive information (statistics, frequency domain and time-domain). Thus, this feature vector is a refined and comprehensive descriptor of an intrusion signal. 4.2.1. Kurtosis: As is known to all, kurtosis is a statistical descriptor of a probability distribution. It is suitable for characterizing vibration signals due to its sensitivity to the pulse signal. The kurtosis (or the general kurtosis) can be directly calculated using the given temporal-domain samples, and it is useful for characterizing stationary signals affected by periodical fluctuation. However, the general kurtosis cannot localize transient or hidden non stationarities (see [14], for details), which are typical features of the vibration signals arising from intrusions in the DMZI system. Hence, in [15]–[17], the general kurtosis was replaced by the spectral kurtosis (SK) [15] (i.e., the statistical values of kurtosis measured across individual frequency bands.), which proved effective in extracting the transient and impulsive features. Note that in the dataflow of Fig. 2, the proposed filter bank can separate an intrusion signal into Q independent frequency bands. Hence, the outputs K 0 , . . . , K Q −1 calculated over these channels are spectral kurtosis values, which are condensed features able to distinguish different intrusions. The kurtosis value K˜ q of the q-th channel’s output y q (n) is calculated as L 1 y q (n) − y¯ q 4 K˜ q = L σq

(8)

n=1

where L is the sample length, y¯ q is the mean value of y q (n), and σq is the standard deviation of y q (n). Further, these kurtosis values should be normalized as K q = Q

K˜ q

−1 q=0

K˜ q

,

q = 0, . . . , Q − 1.

(9)

4.2.2 Zero-Crossing Rate: The incorporation of Zero-crossing rate (ZCR) aims to integrate more complete information related to an intrusion event into the feature vector, which helps to enhance the recognition rate. ZCR refers to the rate of sign changes (from positive to negative) of samples, thus it is a general performance index in the time domain. The ZCR of an input intrusion signal x(n) is defined as L n=1 sign [x(n)] − sign [x(n − 1)] ZCR = (10) 2L where the operator ‘sign’ means sign [x(n)] =

1,

x(n) ≥ 0

−1,

x(n) < 0

.

(11)

Clearly, every entry in the hybrid vector F = [K 0 , K 1 . . . , K Q −1 , Z CR ] is between 0 and 1, which brings convenience to the subsequent SVM-based classification.


7800212



Fig. 4. Schematic diagram of four commonly intrusions.

4.3. SVM Based Pattern Recognition The support vector machine (SVM) realizes classification by mapping input feature vectors F i into a higher-dimensional space φ(F i ) (related to a kernel function K (F i , F j )) and constructing an optimal hyper-plane. The SVM classifier possesses great generalization abilities and high efficiency. As is known, the basic SVM classifier can only deal with two-category classification problem, which is divided into the training step and the testing step. In the training step, given a training set of instance-label pairs (F i , y i ), i = 1, 2, . . . , M (M denotes the number of training vectors), y i ∈ {−1, 1} (Those vectors belonging to the concerned intrusion category are labeled as ‘y i = 1’, others are labeled as ‘y i = −1’). By solving the following optimization problem: M 1 T ξi w w +C w,b,ξ 2 i =1 subject to y i (w T φ(F i ) + b) ≥ 1 − ξi

min

(12)

ξi ≥ 0. where C > 0 is the penalty parameter of the error term, one can acquire an SVM model characterized by the weighting vector w and the constant b. In other words, a linear separatable higherdimensional hyperplane defined by (w, b) with the maximal margin is determined. Assume that Q categories of intrusions need to be recognized (Q > 2). To utilize the basic SVM classifier to deal with this multiple-category problems, we adopt the one-versus-all (OVA) method [18]. Specifically, for an individual q-th category of intrusion (q = 0, . . . , Q − 1), we can obtain its basic binary SVM classifier characterized with (w q , b q ). Then, given a testing feature vector of unknown intrusion, one can obtain Q output decision values z q of these Q trained SVM models. Finally, the intrusion category index c can be determined by selecting the maximum decision value, i.e., c=

arg q=0,...,Q −1

max z q .

(13)

In addition, the classification performance is also related to some SVM parameters, such as the kernel type and its gamma parameter, SVM type etc [19].

5. Experiments and Analysis The experiment setup is shown in Fig. 1. The laser source was a 1550nm distributed feedback laser with 3.5 mW intensity. A long chain link fence was built up. To improve the detection sensitivity, instead of being placed horizontally on the fence, a sensing cable (2.25 kilometer long with single mode) was looped up and down and attached through hose clamps, as Fig. 4 illustrates. The sampling rate of DAQ1 was set as f s = 10 kHz and the recording duration of each trial was set as 3 s.


7800212



Fig. 5. Intrusions signals and their sub-signals. (a) Climbing signals; (b) knocking signals; (c) waggling signals; (d) cutting signals.

We carried out 570 trials including four types of intrusion actions: fence-climbing, knocking the cable, waggling and fence-cutting (repeated 122, 121, 161 and 166 times by a 70 kg person, respectively), which are portrayed in Fig. 4.

5.1. Filter bank based pre-processing In the parameter setting of the quantity Q of the frequency channels illustrated in Fig. 2, both the recognition efficiency and the recognition accuracy should be considered. In our experiment, Q is set as 4, thus the length of the feature vector F equals 5. The reasons are as follows: First, although increasing the value Q can provide detailed information in the frequency domain (for example, a large value Q = 35 was specified in [12]), the efficiency of feature extraction will be degraded; Secondly, as aforementioned, the feature entries K 0 , . . . , K Q −1 not only carry the information in the frequency domain but also the kurtosis information of statistics, which provides the feasibility of using a small value Q . Finally, to guarantee a high classification rate, the size of the feature vector (i.e., Q + 1) should be larger than the number of the categories (i.e., four categories). Some other parameters of the filter bank are set as follows: the length of the frequency vector N = 256 and thus the frequency resolution f = f s /N = 39.0625 Hz. Through implementing FFT analysis on the both environmental records and various intrusion signals, we obtained rough priori knowledge that f e = 220 Hz and f u = 4000 Hz. Then, the observed frequency range B = f u − f e = 3780 Hz. Further, in terms of (3) and (4), one can calculate the starting position e = 6, the bandwidth parameter m = 24. Hence, from (5), one can derive the passbands of these 4 sub-filters are f ∈ [234.3750, 1171.9] Hz, f ∈ [1171.9, 2109.4] Hz, f ∈ [2109.4, 3046.9] Hz, and f ∈ [3046.9, 3984.4] Hz, respectively. Fig. 5(a)–(d) illustrate the original signal x(n) and the filtered sub-signals y 0 (n), . . . , y 3 (n) of these four intrusions. As Fig. 5(a)–(d) depict, compared with the four original waveforms x(n) of these four events, greater differences appear on the decomposed four groups of sub-signals y 0 (n), . . . , y 3 (n), which provides the feasibility to the subsequent accurate feature description.


7800212



Fig. 6. Averaged feature vectors of four intrusions. (a) Climbing; (b) knocking; (c) waggling; (d) cutting.

5.2. Constructing the Hybrid Feature Vector In terms of (8)–(11), one can calculate the hybrid feature vector F of each intrusion trial. For each category of intrusion, the averaged feature vector (averaged among 50 training trails) are plotted in Fig. 6, from which we can summarize the following three distinctions. 1) Compared with other three intrusions, the climbing signal has the largest zero-crossing rate. Moreover, its five entries exhibit a relatively flat shape. 2) For the waggling case, the entry K 3 is much higher than those of the other three intrusions. 3) For the two intrusions of knocking the cable and cutting the fence, their difference lies in the slope of the entries K 0 , . . . , K 3 . Specifically, the average slope provided by knocking intrusions is relatively smaller. The above three aspects of characteristics can be easily captured by the subsequent SVM classifier.

5.3. SVM classification Relevant parameters of the SVM model are set as follows: specify the type of the SVM as C-SVC [19], in which a RBF function with the gamma value 7.2 is chosen as the kernel function. The cost of the C-SVC is 42, and other parameters are set as default. For each trial, we used 50 trails for training and the rest are used for testing. 5.3.1. Analysis of the classification rate: To investigate the contribution of individual entries in our proposed hybrid feature vector F , we used the entire feature vector F = [K 0 , K 1 . . . , K Q −1 , Z CR ] and its partial kurtosis vector [K 0 , K 1 . . . , K Q −1 ] to discriminate these four types of intrusions, whose classification rates are listed in the 2nd row and the 3r d row of Table I, respectively. As a comparison, the experiments based on the EMD intrusion discrimination method (see [11] for details) were also conducted, whose classification rate is also listed in Table I. From Table I, the following conclusions can be drawn. a) The average classification rate of the proposed event discrimination method with the kurtosis vector [K 0 , K 1 . . . , K Q −1 ] achieve the level of the EMD based method. Specifically, our proposed method does better in recognizing the former two intrusions while the EMD based method behaves more excellent in dealing with the latter two intrusions. b) With the entry ZCR added, our proposed method has a significant improvement in the classification rate, showing an overwhelming superiority over the EMD method. This reflects the fact, incorporating multiple features (bandwidth segmentation in frequency domain, kurtosis


7800212


Hybrid Feature Extraction-Based Intrusion Discrimination TABLE 1 PATTERN RECOGNITION RATES OF DIFFERENT METHODS

Method

climbing

knocking

waggling

cutting

EMD method

87.3%

70.9%

99.7%

85.1%

[K 0 , K 1 , K 2 , K 3 ]

91.67%

84.51%

93.69%

75.00%

100%

92.96%

99.1%

85.34%

[K 0 , K 1 , K 2 , K 3 , ZCR]

TABLE 2 TIME CONSUMED BY INDIVIDUAL STAGES

Time Consumed

Proposed Method

EMD Method

Feature Extraction (t/s)

1.3658

6.1719

Pattern Recognition(t/s)

0.084166

1.080270

in statistics, the zero-crossing rate in time domain) into the hybrid vector can considerably enhance the recognition accuracy. 5.3.2. Analysis of the efficiency: Table II lists the time consumed in the feature extraction stage and pattern recognition stage of our hybrid vector based method and the EMD method, from which the following conclusions can be drawn. a) Compared with the EMD based method, our proposed method can considerably improve the efficiency of feature extraction (including the filter-bank based pre-processing and the construction of the hybrid feature vector). Specifically, the time consumed is only about 22% of that of the EMD based method. The reasons lie in the following: As Fig. 2 depicts, for our proposed method, the filter bank outputs y 0 (n), . . . , y Q −1 (n) in a pipeline stream, in which no iterative operations are involved. Conversely, the EMD decomposition addressed in [11] requires multiple iterative operations to successively generate the IMFs, which costs a lot of time. b) The recognition time of our proposed method is a little less than the EMD method, too. This may arise from the fact that the SVM classifier employed in our scheme is suitable for dealing with small amount of samples.

6. Discussions This paper proposed a hybrid feature extraction based intrusion discrimination scheme for optic fiber perimeter security system. Due to the incorporation of the information of time-domain, frequencydomain and statistics, the proposed hybrid feature vector is able to accurately and briefly describe multiple intrusions. Besides, the filter-bank based pre-processing works in a pipeline mode, which accelerates the speed of the construction of hybrid feature vectors. Experiments showed that the proposed method can achieve an average recognition rate higher than that of EMD based scheme. Moreover, the runtime of the proposed scheme is also less than the latter. Hence, the proposed scheme possesses a vast potential in DMZI perimeter security and other security-related applications.


7800212



7. Conclusion The future work focuses on developing approaches to discriminate more intrusions. To achieve this, the following aspects should be taken into accounts: 1) Increase the length of feature vector (i.e., increasing the value Q ). Hence, some other techniques of feature extraction (such as GMM-vector) are expected to be introduced. 2) Besides the features of spectral kurtosis across individual frequency bands and ZCR, it would be beneficial to explore more optional features sets and then shrink the set using well-known methods such as Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA), which helps to discriminate more intrusions. 3) It should be emphasized that, the feature generation mode of our proposed scheme is distinct from that of the well-known classifiers ensemble approach for classification (such as LPBoosting method [20], the multiclass v-LPBoost [21]). In the latter scheme, for any individual weak classifier to be combined, its feature is directly extracted from the original data. In contrast, for the feature vector F = [K 0 , K 1 . . . , K Q −1 , Z CR ] in our scheme, as Fig. 2 depicts, only the last entry ZCR is directly extracted from the original data x(n), while any other entry K i (i = 0, . . . , Q − 1) is derived from y i (n)(i = 0, . . . , Q − 1) (i.e., the i -th filtered version of the original data x(n)). Nevertheless, in the future, it is desirable to apply the latter scheme (more features are suggested to be considered) in the intrusion discrimination. 4) Further improve the optical detection device and the signal processing algorithm. Although the detected signals at PD1 and PD2 in Fig. 1 arise from the intrusion, they are not the original versions of the intrusions since optical modulation occurs (Optical modulation will produce several new frequency components, as [22] stated). Hence, if some improved detection device (like the 3 × 3 fiber coupler) and reliable demodulation algorithms are employed [22], the recognition performance of intrusions is sure to be enhanced.

References [1] Q. Sun, D. Liu, H. Liu, Y. He, and J. Yuan, “Distributed disturbance sensor based on a novel Mach–Zehnder interferometer with a fiber-loop,” Proc. SPIE, vol. 6344, pp. 63 440K-1–63 440K-7, Jun. 2006. [2] Q. Sun, D. Liu, H. Liu, and P. Shum, “Distributed fiber-optic sensor with a ring Mach–Zehnder interferometer,” Proc. SPIE, vol. 6781, pp. 67 814D-1–67 814D-8, Nov. 2007. [3] S. Xie, M. Zhang, S. Lai, and Y. Liao, “Positioning method for dual Mach—Zehnder interferometric submarine cable security system,” Proc. SPIE, vol. 7677, pp. 76 770A-1–76 770A-4, Apr. 2010. [4] Y. Zhou, S. Jin, and Z. Qu, “Study on the distributed optical fiber sensing technology for pipeline leakage protection,” Proc. SPIE, vol. 6344, pp. 634 435–1–634 435-6, Jun. 2006. [5] L. Jiang and R. Yang, “Identification technique for the intrusion of airport enclosure based on double Mach–Zehnder interferometer,” J. Comput., vol. 7, no. 6, pp. 1453–1459, 2012. [6] S. Xie, Q. Zou, L. Wang, M. Zhang, Y. Li, and Y. Liao, “Positioning error prediction theory for dual Mach–Zehnder interferometric vibration sensor,” J. Lightw. Technol., vol. 29, no. 3, pp. 362–368, Feb. 2011. [7] C. Ma et al. “Long-range distributed fiber vibration sensor using an asymmetric dual Mach–Zehnder interferometers,” J. Lightw. Technol., vol. 34, no. 9, pp. 2235–2239, May 2016. [8] X. Huang, J. Yu, K. Liu, T. Liu, and Q. Chen, “Configurable filter-based endpoint detection in DMZI vibration system,” IEEE Photon. Technol. Lett., vol. 26, no. 19, pp. 1956–1959, Oct. 2014. [9] S. S. Mahmoud, Y. Visagathilagar, and J. Katsifolis, “Real-time distributed fiber optic sensor for security systems: Performance, event classification and nuisance mitigation,” Photon. Sensors, vol. 2, no. 3, pp. 225–236, 2012. [10] L. Liu, W. Sun, Y. Zhou, Y. Li, J. Zheng, and B. Ren, “Security event classification method for fiber-optic perimeter security system based on optimized incremental support vector machine,” in Proc. 6th Chinese Conf. Pattern Recognit., 2014, pp. 595–603. [11] K. Liu et al. “A high-efficiency multiple events discrimination method in optical fiber perimeter security system,” J. Lightw. Technol., vol. 33, no. 23, pp. 4885–4890, Dec. 2015. [12] X. Huang, Y. Wang, K. Liu, T. Liu, C. Ma, and Q. Chen, “Event discrimination of fiber disturbance based on filter bank in dmzi sensing system,” IEEE Photon. J., vol. 8, no. 3, pp. 1–14, Jun. 2016. [13] A. V. Oppenheim et al., Discrete-Time Signal Processing, vol. 2. Englewood Cliffs, NJ, USA: Prentice-Hall, 1989. [14] R. Dwyer, “Detection of non-Gaussian signals by frequency domain kurtosis estimation,” in Proc. IEEE Int. Conf. Acoust. Speech, Signal Process., 1983, vol. 8, pp. 607–610. [15] J. Antoni, “The spectral kurtosis: A useful tool for characterising non-stationary signals,” Mech. Syst. Signal Process., vol. 20, no. 2, pp. 282–307, 2006. [16] F. Immovilli, M. Cocconcelli, A. Bellini, and R. Rubini, “Detection of generalized-roughness bearing fault by spectralkurtosis energy of vibration or current signals,” IEEE Trans. Ind. Electron., vol. 56, no. 11, pp. 4710–4717, Nov. 2009.


7800212



[17] B. Eftekharnejad, M. Carrasco, B. Charnley, and D. Mba, “The application of spectral kurtosis on acoustic emission and vibrations from a defective bearing,” Mech. Syst. Signal Process., vol. 25, no. 1, pp. 266–284, 2011. [18] M. Galar, A. Fernandez, E. Barrenechea, H. Bustince, and F. Herrera, “An overview of ensemble methods for binary ´ classifiers in multi-class problems: Experimental study on one-vs-one and one-vs-all schemes,” Pattern Recognit., vol. 44, no. 8, pp. 1761–1776, 2011. [19] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector machines,” ACM Trans. Intell. Syst. Technol., vol. 2, no. 3, 2011, Art. no. 27. [20] C. Hinrichs, V. Singh, L. Mukherjee, G. Xu, M. K. Chung, and S. C. Johnson, “Spatially augmented LPboosting for AD classification with evaluations on the ADNI dataset,” Neuroimage, vol. 48, no. 1, pp. 138–149, 2009. [21] A. Demiriz, K. P. Bennett, and J. Shawetaylor, “Linear programming boosting via column generation,” Mach. Learn., vol. 46, no. 1, pp. 225–254, 2002. [22] Q. Sun, D. Liu, J. Wang, and H. Liu, “Distributed fiber-optic vibration sensor using a ring Mach–Zehnder interferometer,” Opt. Commun., vol. 281, no. 6, pp. 1538–1544, 2008.


7800212